Detailed Description
The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to fall within the scope of the application.
The embodiment of the application provides a content recommendation method, a content recommendation device, electronic equipment and a storage medium.
The content recommendation device may be integrated in an electronic device, and the electronic device may be a terminal, a server, or other devices. The terminal can be a mobile phone, a tablet personal computer, an intelligent Bluetooth device, a notebook computer, a personal computer (Personal Computer, PC) or other devices, and the server can be a single server or a server cluster formed by a plurality of servers.
In some embodiments, the content recommendation device may also be integrated in a plurality of electronic devices, for example, the content recommendation device may be integrated in a plurality of servers, and the content recommendation method of the present application is implemented by the plurality of servers.
In some embodiments, the terminal may also function as a server to implement some or all of the functionality of the server.
For example, referring to fig. 1a, the electronic device may be a server, the server may obtain a historical behavior sequence of the target object from an object information base, the historical behavior sequence includes behavior items arranged in time sequence, the behavior items represent interaction behaviors of the target object with respect to content, the server may determine similarity between two behavior items separated by a fixed interval in the historical behavior sequence, construct an interest-degree curve according to the similarity, the interest-degree curve reflects a change of the interest-degree with time, the interest-degree represents an interest-degree of the target object, determine recommended content adapted to the target object in a content pool based on the interest-degree curve, and finally send the recommended content to a mobile terminal of the target object.
For example, in some embodiments, the news recommendation server may generate a personalized recommendation for the mobile terminal object, and the news recommendation server may retrieve the latest 40 news click records of the object from the object information base, and arrange the latest 40 news click records in time sequence to form a historical behavior sequence [ behavior item 1, behavior item 2,..and behavior item 40], where each behavior item may include an article ID, a category tag, a reading duration, and other interactive behavior features of news, and then calculate similarities between every two adjacent behavior items in the historical behavior sequence, and connect the obtained 39 similarity values in time sequence to form an interest generalization curve.
If the interest generalization curve of the object A is an interest generalization curve of high-frequency fluctuation, the interest of the object A is wider, so that 30% of non-historical preference categories break through the information cocoon room when the news recommendation server pushes news to the object A.
If the interest generalization curve of the object B is an interest generalization curve with low-frequency fluctuation, which indicates that the interest of the object B is more concentrated, the news recommendation server can perform vertical field depth mining on the field of interest of the object B when pushing news to the object B, and add long-text content to strengthen the viscosity of the object based on the field with high behavior proportion such as collection, praise and the like in behavior items.
The following will describe in detail. The numbers of the following examples are not intended to limit the preferred order of the examples.
In this embodiment, a content recommendation method is provided, as shown in fig. 1b, and the specific flow of the content recommendation method may be as follows:
110. And acquiring a historical behavior sequence of the target object, wherein the historical behavior sequence comprises behavior items arranged in time sequence, and the behavior items represent the interactive behavior of the target object for the content.
Wherein, the historical behavior sequence refers to behavior records of interaction of the target object with the content in a period of time, and is a set arranged in time sequence.
The information carrier in which the content is interacted with may comprise articles, music, video, merchandise, advertisements, images, etc.
The behavior item is a minimum recording unit of single content interaction, can record the behavior of an object on a certain content, and can contain content characteristics and behavior characteristics. The content features describe the properties of the content itself and the behavior features describe the way the object interacts with the content.
Content features are used to characterize the inherent properties of the interacted with content, helping to understand the classification, topic or value of the content. In some embodiments the content features may include:
classification such as seal type, primary class, secondary class, etc.;
Themes such as content ID, content keywords, tags, etc.;
Sources such as seal authors, published media, etc.;
Release time, such as freshness of the content;
hotness, such as click quantity, forwarding quantity, comment quantity and the like of the content.
The behavior features are used for describing the behavior mode of the interaction of the object and the content and reflecting the interest intensity, preference and intention of the object. In some embodiments the behavioral features may include:
object identity such as object account number, object identity information, etc.;
interaction types such as clicking, praying, collecting, forwarding, commenting and the like;
The interaction time length is as browsing time length, page sliding speed, repeated access times and the like;
The interaction depth is whether to browse completely, whether to view the comment area, whether to click on the related recommendation and the like.
In some embodiments, the historical behavior sequence of the target object may be obtained in a variety of ways. For example, the behavior items of the target object may be collected from different data sources and sorted into a historical behavior sequence in chronological order. The data source may include a client of the target object, a server log, an object information database, third party data, and so on.
For example, in some embodiments, the client may record the interaction behavior of the object with the content, while the backend service may record the log data of the object request, including the interaction behavior of the object with the content. After the initial data is obtained, the data can be cleaned, such as denoising, complement, deduplication, and the like.
120. The similarity between two behavior items in the historical behavior sequence that are separated by a fixed interval is determined.
Wherein, the calculation mode of the similarity has multiple possibilities and is not a single fixed mode. For example, similarities between adjacent behavior items in a historical behavior sequence may be calculated to capture interest drift of a target object. For example, the similarity of each behavior item to the global average of all the behavior items of the historical behavior sequence can be calculated to establish a baseline of long-term interest of the subject and filter the interference of occasional behaviors, for example, the aggregate value of the similarity of the behavior items can be calculated in a sliding time window to balance the contradictory relationship of short-term behavior fluctuation and long-term stability preference.
The fixed interval may include a location interval, a time interval, and the like. The position interval refers to the fact that the action items only have fixed position intervals in the sequence, such as the pure sequence relations of 1 node, 2 nodes and the like, and the time interval refers to the fixed time interval between the action items, such as every 5 minutes. Wherein, the value of the fixed interval has a significant effect on interest generalization modeling. In some embodiments, the fixed interval may be sized based on the application scenario. For example, a small fixed interval value of 1-3 is suitable for capturing immediate interest drift, a medium fixed interval value of 4-10 is suitable for identifying interest migration patterns, and a large fixed interval value of greater than 10 is suitable for reflecting long-term interest evolution.
Similarity between behavior items is the similarity between multidimensional features contained in the behavior items, which may contain content features, behavior features, and the like. For example, the multidimensional feature of a behavior item may be expressed as [ click, reading duration, article ID, category, class 1, class 2, etc. ], and the similarity between two behavior items is obtained by calculating the similarity of the multidimensional feature.
The similarity may be calculated in various ways, for example, cosine similarity, euclidean distance, manhattan distance, jaccard coefficient, dice coefficient, edit distance, dynamic time warping distance, KL divergence, pearson correlation coefficient, semantic similarity based on deep learning, similarity calculation in combination with time decay, and combinations of the above, and the like.
For example, in some embodiments, for a historical behavior sequence [ behavior item 1, behavior item 2, ], behavior item 40], assuming a fixed interval of 1, cosine similarity between two neighboring behavior items may be calculated to obtain a plurality of similarities [ cos (behavior item 1, behavior item 2), cos (behavior item 2, behavior item 3), cos (behavior item 3, behavior item 4) ], cos (behavior item 39, behavior item 40) ], and these interests are finally connected as a curve using the cosine similarity subtracted by 1 to obtain an interest flooding curve [1-cos (behavior item 1, behavior item 2), 1-cos (behavior item 2, behavior item 3), 1-cos (behavior item 3, behavior item 4) ].
In some embodiments, two sequences may be truncated from the historical behavior sequence through a sliding window. Thus, determining the similarity between two behavior items in a historical behavior sequence that are separated by a fixed interval may include:
Extracting a first subsequence and a second subsequence from the historical behavior sequence, wherein the first subsequence comprises ith to jth behavior items in the historical behavior sequence, the second subsequence comprises ith to jth+k behavior items in the historical behavior sequence, i and j are positive integers, and k is a fixed interval;
and determining the similarity between the parity item pairs in the first subsequence and the second subsequence, wherein the parity item pairs are two behavior items corresponding to the positions in the first subsequence and the second subsequence.
For example, referring to fig. 1c, where the 1 st action item in the historical action sequence is the action item generated by the target object last time, the fixed interval k is 1, i=1, j=39, then the first sub-sequence S1 includes the 1 st to 39 th action items in the historical action sequence, i.e. S1＝[x11,x12,x13,…,x139, and the second sub-sequence S2 includes the 2 nd to 40 th action items in the historical action sequence, i.e. S2＝[x22,x23,x24,…,x240.
The parity pair includes a first parity in the first sub-sequence and a second parity in the second sub-sequence corresponding to the first sub-sequence. Then x11 of S1 and x22 of S2 are parity pairs, the first parity is x11, the second parity is x12 of x22;S1 and x23 of S2 are parity pairs, the first parity is x12, the second parity is x23, and so on. Similarity is calculated between parity terms in pairs.
In some embodiments, the contradictory relationship of short-term behavior fluctuations to long-term stability preferences may be balanced in calculating the similarity in consideration of the context information. Thus, determining the similarity between pairs of parity items in the first sub-sequence and the second sub-sequence comprises:
determining a first set, wherein the first set comprises a first parity item and adjacent behavior items thereof in a first subsequence, namely the first set is a context set of the first parity item;
Determining a second set, wherein the second set comprises a second parity item and adjacent behavior items thereof in a second subsequence, namely the second set is a context set of the second parity item;
and obtaining the similarity between the first parity item and the second parity item according to the average value of the first set and the average value of the second set.
For example, accidental actions exist in user behaviors, and the embodiment expands single-point behaviors into local context windows by aggregating adjacent items, so that the influence of short-term noise is reduced, and interest generalization fluctuation caused by accidental operations is smoothed, and more stable information is captured. For example, in a recommendation system, a user may suddenly click on an unusual piece of content on a day, but their interests may be more stable in the long term, so it is desirable to combine the context to reduce this noise.
The mean value of the calculation set may be the average value of the terms, or may be other aggregation manners, such as taking an average vector. The mean value is used for integrating information of a plurality of points, and reducing the random fluctuation influence of a single point, thereby reflecting the trend of the position more stably.
In some embodiments, in order to improve accuracy of analysis, it is avoided that too small a window causes excessive noise when data is sparse, or that details are lost when the window is too large when data is dense, the sequence lengths of the acquired first subsequence and second subsequence may be dynamically adjusted by analyzing the liveness of the object in real time. Thus, extracting the first sub-sequence and the second sub-sequence from the historical behavioral sequence may include:
acquiring the behavior liveness of a target object;
Determining a window size of the sliding window based on the activity level of the behavior;
Placing a sliding window in the historical behavior sequence to obtain a first subsequence, wherein the first subsequence comprises the ith to jth behavior items in the historical behavior sequence, and i and j differ by a window size;
and moving a sliding window in the historical behavior sequence by taking k as a step length to obtain a second subsequence.
Where behavioral activity refers to the frequency or degree of intensity of the activity of the subject.
In some embodiments, smaller windows may be used when objects are behaving densely, and larger windows may be used when they are sparsely. For example, for high liveness objects, the window is scaled down to the hour level, capturing subtle interest drift of the object in dense behavior, and for low liveness objects, the window is extended to the week/month level to integrate discrete behavior of the object.
The manner of obtaining the behavior activity of the target object is various, for example, the behavior activity may be obtained by analysis and statistics according to the daily login times, the use duration, the interaction behavior, and the like of the target object in a period of time.
For example, core events such as login, clicking, interaction, consumption and the like of an object can be acquired, interaction types such as object ID, behavior type, timestamp, operation duration, page path and the like are recorded, abnormal values in data are cleaned, missing values are filled, indexes such as login times, behavior frequency, behavior diversity and the like are calculated according to the data, and finally a plurality of analysis modes such as weighted summation, exponential decay calculation and the like are carried out on the indexes, so that the liveness score of the object is finally obtained.
130. According to the similarity, an interest generalization curve is constructed, the interest generalization curve reflects the change of the interest generalization along with time, and the interest generalization represents the interest extensive degree of the target object.
The interest generalization curve provided by the embodiment of the application describes the change of the interest extensive degree along with time and provides decision basis for personalized service.
After the similarity is obtained in step 120, the similarity needs to be further processed to obtain the interest generalization, so that the interest generalization can accurately reflect the interest change of the object. For example, multiple processes such as normalization, linear transformation, exponential transformation, logarithmic transformation, etc. may be performed on the similarity to obtain the interestingness.
For example:
interest generalization= (1-similarity)k
Wherein when k >1, the interest generalization is more sensitive to low similarity, the change is steeper, and when 0< k <1, the interest generalization change is flatter.
For example:
The attenuation rate is controlled by adjusting k, the interest generalization of the high-similarity region is reduced more rapidly as k is larger, the interest generalization is 1 when similarity=0 after normalization, and the interest generalization is 0 when similarity=1.
For example:
the interest generalization of the low-similarity region is more remarkable when k is larger, and the interest generalization approaches 0 when the similarity approaches 1.
For example:
If the similarity is less than or equal to t, then
If similarity > t, then
Where t=0.5, the interest generalization drops faster in the low similarity region.
For example:
Wherein, the threshold value of interest generalization dip is set by m, and k controls the steepness of the curve. For example, when m=0.5, the interest generalization rapidly decreases around similarity=0.5.
For example:
Interest generalization = 1-similarity2
When the similarity is high, the interest generalization drops faster, and the low interest generalization of the high-similarity behavior is highlighted.
For example:
Wherein the interest generalization changes more gently in the low similarity region and the high similarity region drops more rapidly.
For example:
interest generalization = -similarity log (similarity) - (1-similarity) log (1-similarity)
When the similarity is close to 0 or 1, the interest generalization is lower, and the interest generalization is highest in the middle area (similarity is approximately equal to 0.5), so that uncertainty maximization is reflected.
Wherein the interestingness is considered to be 1 only when the similarity is below the threshold θ, and 0 otherwise.
In summary, in this embodiment, the interest generalization needs to decrease with increasing similarity, so as to ensure that low similarity corresponds to high interest generalization. In some embodiments, the interestingness needs to be normalized so that it falls within a reasonable range (e.g., [0,1 ]) for subsequent analysis.
For example, the interest-generalization may be set to 1-similarity, e.g., the interest-generalization between the 2 nd and 3 rd action items in the sequence may be set to 1-cos (action item 2, action item 3).
140. And determining the recommended content adapted to the target object based on the interest generalization curve.
According to the embodiment of the application, the change trend of the extensive degree of the interests along with time is known by analyzing the interest generalization curve of the target object, so that the interest preference of the target object in different time periods is deduced, and the content which is more in line with the current interest of the target object is recommended for the target object according to the interest preference.
For example, if the curve is high, indicating that the interest of the target object is broader, it may be interested in various types of content, and the target object may be more willing to accept diversified content, so various types of content may be recommended. If the curve is low, indicating that the interest of the target object is more concentrated, it may be more interesting for a particular type of content, the target object may be more inclined to a particular type of content, so that content highly correlated to its current interest may be recommended. If the curve exhibits a rising or falling trend, indicating that the interest range of the target object is expanding or contracting, for example, if the interest-generalization curve shows that the interest range of the target object is expanding, an attempt may be made to recommend some content that is related to, but slightly different from, its existing interest to explore its potential interest.
Based on the interest generalization curve, the scheme for determining the recommended content adapted by the target object is various, for example, a machine learning model, a clustering and classifying algorithm can be adopted to determine the recommended content.
For example, a behavior prediction model based on machine learning can be adopted, a field of interest of a target object in a future period can be predicted according to the interest flooding curve, and the content of the field can be used as recommended content to be pushed to the object;
For example, the target object may be classified according to the interest generalization curve, and according to the classification type, content conforming to the classification type is selected from the content pool and pushed to the object as recommended content;
For example, the target object may be classified according to an interest generalization curve, and when the conventional pushing is performed on the object, the non-explored domain content of the target object with different proportions is doped into the conventional pushing flow according to the classification type, and so on.
For example, the timeliness and diversity of the recommended content may be dynamically adjusted in combination with the fluctuation of the interest-generalization curve and the time-decay factor. For example, the interest generalization curve is divided into different time windows, the stability of the interest generalization of the object is analyzed, if the interest generalization curve of the object fluctuates severely, the short-term interest of the object is indicated to change rapidly, the content with strong timeliness such as hot news, short-term trend content and the like can be recommended, and if the curve fluctuation is gentle, the object is indicated to have stable long-term interest, and the depth content such as a thematic article and a series of courses can be recommended.
For example, the weights of collaborative filtering and content filtering may be dynamically switched according to the characteristics of the interest generalization curve. For example, when the curve peak value of the interest generalization curve is larger than a preset peak value, the weight of content filtering is enhanced, diversified contents are recommended, and when the curve valley of the interest generalization curve is smaller than a preset valley value, the weight of collaborative filtering is enhanced, and the contents preferred by similar objects are recommended. And judging the interest change speed through the slope of the curve, and rapidly switching strategies.
For example, object interest maps can be constructed using interest generalization curves, and cross-domain content can be recommended through map diffusion. For example, the historical behavior sequence of the object is mapped to the interest node, an interest map is constructed according to the behavior similarity, when the interest generalization curve is increased, the content of the adjacent interest field is expanded and recommended from the map, and when the curve is decreased, only the content of the core node of the map is focused, so that excessive expansion is not performed.
For example, when the end of the object interest generalization curve is at an ascending, the probability of exploring the new domain content is increased, and when the end of the curve is at a descending, the exploration is reduced, and the verified interest content is preferentially recommended.
Therefore, in some embodiments, a preliminary candidate set may be screened from a huge amount of content, and the scoring of the candidate content by the target object may be predicted by using a prediction model in combination with an interest generalization curve of the target object, and the high-score content is selected and pushed to the target object after the scoring is ordered, so that the determining the recommended content adapted to the target object based on the interest generalization curve may include:
Acquiring contents to be detected in a candidate set;
Adopting a prediction model to predict recommendation scores of target objects aiming at the content to be detected based on the interest generalization curve;
And selecting recommended contents from the candidate set by sequencing the recommendation scores of all the contents to be detected in the candidate set.
In the embodiment, the preliminary candidate set is screened from mass contents through coarse-granularity filtering, so that the complexity of subsequent calculation is reduced, and the basic correlation of the candidate contents is ensured. In some embodiments, the candidate set may rely on default policies, such as trending content, regional preferences, etc., to provide base recommendations for new objects lacking behavioral data.
The prediction model may be a deep learning model, such as a time series model, a multi-task model, and the like. The time sequence model can comprise a Long Short-Term Memory network (LSTM), a gate control circulation unit (GRU, gated Recurrent Unit), a Transformer and the like, and can directly model the time sequence change of the interest generalization curve and predict the future interest. The multitasking model may embed the behavior data and the content to be detected of the object, respectively, and calculate a matching score, and may include a deep structured semantic model (DSSM, deep Structured Semantic Model), sentence-BERT (SBERT), a multitasking recommendation model (User-Item Twin Towers), and the like.
The method for sorting the recommended content in the candidate set may include sorting the content of the top-K in the descending order of the prediction scores, and so on.
The following is a detailed description:
in some embodiments, the behavior item may include a content feature and a behavior feature, the content to be detected may include a domain label, and predicting a recommendation score of the target object for the content to be detected based on the interest generalization curve using a prediction model may include:
Performing feature stitching on the content features and the behavior features of each behavior item in the interest generalization curve and the historical behavior sequence to obtain personal features of the target object;
Predicting a behavior index of a target object for the content to be detected based on the personal characteristics and the domain label of the content to be detected;
And determining recommendation scores of the content to be detected based on the behavior indexes.
According to the embodiment of the application, whether the target object is interested in the domain label of the content to be detected, namely the behavior index, is predicted through the personal characteristics of the target object. The behavior index refers to a quantized evaluation value of specific interaction behavior possibly generated by the target object to be detected and obtained through model prediction. It is a core intermediate variable that connects object features with final recommendation scores, and may include interaction probability, interaction duration, interest matching, negative feedback risk, and so on.
For example, the interaction probability may include a click rate (CTR), a play rate, a purchase rate, a download rate, etc., the interaction duration may include a stay duration, a review number, an interaction comment probability, a sharing wish, etc., the negative feedback risk may include a skip rate, a bad comment rate, a reporting probability, etc., and the interest matching degree may include a degree of matching of the content with the object interest, etc.
The content features are features describing the self attribute of the content, and can comprise basic attribute features, semantic features, aging features and the like of the content. The basic attribute characteristics of the content can comprise unique Identification (ID), type, classification label, keywords, theme and the like of the content, the semantic characteristics of the content can comprise title embedding, global abstract characteristics, emotion tendencies and the like of the content, and the aging characteristics of the content can comprise release time stamps, hot event relatedness, aging attenuation coefficients and the like.
For example, the content feature may be (first class cat1=science and technology, second class cat2=artificial intelligence, keyword= [ "GPT-4", "big model" ], topic= [0.7,0.2,0.1], emotion tendency=0.8).
The behavior features are dynamic behavior features generated in the process of describing the interaction of the object and the content, and can comprise explicit behavior features, implicit behavior features, sequence features and the like. The explicit behavior features may include whether to interact, duration of interaction, type of interaction, speed of interaction, etc., the implicit behavior features may include attention heat maps, completion rate of interaction, time of interaction, etc., and the sequential features may include interest transfer matrices, periodic patterns of behavior, time intervals of continuous behavior, etc.
For example, the behavioral characteristics may be (whether click=1, reading duration=45 s, time distribution= [09:30,12:15,20:40 ]).
The domain tag description content belongs to the domain, and can comprise a domain type, a primary class, a secondary class, a tertiary class, a social propagation coefficient and the like.
In some embodiments, the recommendation score may be obtained by weighted summation of behavior indexes, wherein the interaction probability and the weight allocation of the interaction duration are dynamically adjusted according to the business objective. For example, in the scene of the interaction probability, such as information flow advertisement, electronic commerce promotion, news headline, application download page and the like, the weight of the interaction probability can be improved, and in the scene of the interaction duration, such as online education course, long video platform, music streaming media and reading platform, the weight of the interaction duration can be improved.
For example, the prediction model may be a multi-task model, wherein the input of the multi-task model is an interest generalization curve, content characteristics and behavior characteristics of each behavior item in a historical behavior sequence, and a domain label of the content to be detected, the output of one tower in the multi-task model is the interaction probability of the target object for the content to be detected, the output of the other tower is the interaction duration of the target object for the content to be detected, and finally, the interaction probability and the interaction duration are weighted and summed to obtain the recommendation score of the content to be detected.
Thus, in some embodiments, the prediction model may include a click type prediction network and a reading duration prediction network, and based on the personal characteristics and the domain label of the content to be detected, predicting the behavior index of the target object for the content to be detected may include:
a click type prediction network is adopted, and the click type of a target object aiming at the content to be detected is predicted based on personal characteristics and the domain label of the content to be detected;
a reading duration prediction network is adopted, and the reading duration of a target object aiming at the content to be detected is predicted based on personal characteristics and the domain label of the content to be detected;
And calculating the behavior index according to the click type and the reading time.
The click type prediction network is used to predict the click probability or click type (i.e. whether the object clicks) of the content, and may include a deep neural network (Deep Neural Network, DNN), a factorer neural network (DeepFM), a transducer-Encoder network, a gradient boosting decision tree (Gradient Boosting Decision Tree, GBDT), and so on.
The reading duration prediction network is used for predicting the stay duration of the object on the content, and may include a deep regression network (Deep Regression Network), a survival analysis network (Deep Survival Network), a time sequence convolution network (Temporal Convolutional Network, TCN), a Multi-task learning network (Multi-TASK LEARNING) and the like.
According to the method, personal characteristics for comprehensively describing interest preference of the object are constructed by integrating the interest generalization curve of the object and the content characteristics and behavior characteristics in the historical behavior sequence, the behavior index of the object to the content is accurately predicted based on the personal characteristics, and the recommendation score is finally calculated. The method has the effects of improving the individuation degree and accuracy of the recommendation system, realizing more efficient content matching and object satisfaction improvement, and optimizing the service conversion rate and the object retention rate.
In the embodiment of the application, the content recommendation method can acquire the object information of the target object, such as the interest generalization curve, the historical behavior sequence and the like of the target object, and determine the recommendation content adapted by the target object from the content pool. In some embodiments, the content recommendation method may include three steps of coarse sorting, scoring and reordering, where the scoring step is a step of scoring the recommendation of the predicted target object for the content to be detected according to the above embodiment of the present application. Prior to the scoring step, a coarse ordering step may be included:
acquiring object information of a target object;
And screening a plurality of candidate contents in the content pool based on the object information to obtain a candidate set.
The object information can also comprise various data such as the known interested field, the known uninteresting field, the probably interested field, the probably uninteresting field and the like of the object, and the rough sorting can quickly screen out a candidate set related to the target object from a mass content pool so as to provide high-quality input for the follow-up.
After determining the recommended content for the target object adaptation, a reordering step may be further included:
and reordering all the recommended contents, and sending the reordered recommended contents to the target object.
The reordering can carry out refined ordering on the recommended content, and the individuation and the service value of the recommended result are optimized. For example, in video recommendation, if the initially generated recommendation lists are all the same type of movie, the reordering stage may introduce diversity strategies to insert comedy, love, etc. other types of movies into the list, for example, in news recommendation, the reordering stage may preferentially show the latest released news, moving the outdated news down.
In some embodiments, the reordering stage may dynamically adjust the ordering priority of the recommended content according to the interest generalization degree of the target object, so as to more accurately meet the requirement of the target object and promote the recommendation effect. Thus, reordering all recommended content may include:
determining the interest generalization type of the target object based on the interest generalization curve;
If the target object is of the first generalization type, the priority of sequencing the recommended content conforming to the first generalization type is improved;
if the target object is of the second generalization type, the priority of sequencing the recommended content conforming to the second generalization type is improved;
and reordering all the recommended contents according to the priority of the sequencing of the recommended contents.
The interest generalization type of the target object can be determined according to the shape of the interest generalization curve. For example, whether the interests of the object are gradually dispersed or concentrated may be determined according to the overall slope of the curve, whether the interests of the object are stable may be determined according to the variance of the curve, and whether there is a periodic or abrupt interest change may be determined according to the local shape similarity of the curve, for example.
The interest generalization types may include a first generalization type, a second generalization type, a third generalization type, a fourth generalization type, and the like. For example, the interest generalization types can be classified into a first generalization type, a second generalization type, and a third generalization type by high, medium, and low. The object with high generalization type has wide interest and strong exploratory property, and the recommended content needs to be compatible with diversity, exploratory property and interest matching degree. The low-generalization type object interest set, the recommended content needs to be more accurately matched with the core interest, and unnecessary exploration is reduced.
For example, if the target object is of a high-generalization type, the content of the target object unexplored region may be ranked in the front portion and the content of the region of known interest may be ranked in the rear portion, if the target object is of a medium-generalization type, the content of the target object unexplored region may be ranked in balance with the content of the region of known interest, and if the target object is of a low-generalization type, the content of the target object unexplored region may be ranked in the rear portion and the content of the region of known interest may be ranked in the front portion.
In some embodiments, unexplored content may be dynamically recommended by combining historical behavior sequences of objects, interest generalization curves, and time periods. For example, when the object is in a time period with wide interest and strong exploratory property, the content of the unexplored field, such as articles of 'travel' or 'art' class, which the object never clicks, can be recommended to the object more, and when the object is in a time period with concentrated interest and weak exploratory property, the content of the unexplored field can be recommended to the object less.
Thus, after determining the recommended content adapted to the target object, it may further include:
Determining a domain label of the target object which never generates interactive behaviors according to the historical behavior sequence of the target object, wherein the domain label represents the content domain to which the content belongs;
Determining the interest generalization types of the target object in different time periods based on the interest generalization curves;
If the target object is of the first generalization type in the target time period, selecting unexplored content in the content pool, wherein the unexplored content is content with a domain label of which the target object never generates interaction behavior;
And sending unexplored content simultaneously when the recommended content is sent to the target object in the target time period.
The duty cycle of sending unexplored content can be dynamically adjusted according to the interest generalization curve, so in some embodiments, when sending recommended content to a target object in a target time period, sending unexplored content simultaneously can include:
determining a mixing proportion weight according to the interest generalization curve, wherein the mixing proportion weight can be used for adjusting the proportion of recommended content and unexplored content in transmission;
And sending unexplored content at the same time when sending the recommended content to the target object according to the mixed proportion weight.
For example, through the interest generalization curve, the interest generalization types of the object in different time periods such as early, middle, late and the like are identified in real time, and the recommendation strategy is dynamically adjusted. For example, in the morning commute period, the target object is of a high-generalization type, the target object may be in an open state of information receiving, and can recommend diversified contents such as international news and unexplored scientific and technological cold knowledge, in the noon break period, the target object is of a medium-generalization type, and can recommend moderately diversified contents such as easy entertainment, food, short reading and the like, and in the evening break period, the target object is of a low-generalization type, and can recommend accurate contents such as movie dramas, deep long texts and the like which the target normally sees.
For example, through the interest generalization curve, the interest generalization types of the object in different time periods of the working day and the non-working day are identified in real time, and the recommendation strategy is dynamically adjusted. For example, on monday to friday, the target object is in a working state, time-slicing is prone to efficiently acquire core interest contents related to work and study, so that the content in the history high-frequency interest field of the target object is recommended mainly by accurate matching, the ratio of the content in the unexplored field is limited, for example, on friday to friday, the target object is time-free and has more open interest, the content in the entertainment, leisure and masses fields is prone to be explored, the cross-field content is recommended for the target object, and the ratio of the unexplored field content is greatly improved.
From the above, the embodiment of the application can acquire the historical behavior sequence of the target object, wherein the historical behavior sequence comprises behavior items arranged in time sequence, the behavior items represent the interactive behavior of the target object for the content, the similarity between two behavior items which are separated by a fixed interval in the historical behavior sequence is determined, an interest generalization curve is constructed according to the similarity, the interest generalization curve reflects the change of the interest generalization along with time, the interest generalization represents the interest generalization degree of the target object, and the recommended content adapted to the target object is determined based on the interest generalization curve.
According to the embodiment of the application, the historical behavior sequence of the object is deeply mined, the behavior mode of the object is comprehensively captured, the interest generalization curve similar to the heartbeat curve is drawn, and the change of the object interest generalization degree along with time is intuitively reflected, so that the objects with extensive or concentrated interests are accurately distinguished. Based on the method, the system recommends diversified contents for objects with wide interests, avoids information cocoons, improves exploratory performance, recommends accurate matching contents for objects with concentrated interests, and improves satisfaction. The proposal widens the range of recommended content, satisfies interest matching and avoids content convergence, thereby improving the quality of content recommendation.
The method described in the above embodiments will be described in further detail below.
In this embodiment, a method according to an embodiment of the present application will be described in detail by taking news push as an example.
As shown in fig. 2a, an embodiment of the present application provides a content recommendation system, which includes an object information module, a content pool module, a coarse ordering module, a scoring module and a rearrangement module. The object information module integrates the static characteristics and the dynamic behavior characteristics of the object, and the content pool module integrates massive recommended content data.
The embodiment of the application provides a depth information mining method based on a target object behavior sequence, realizes the imaging representation of the object interest dynamic change by constructing an interest generalization curve, adopts a time sequence similarity calculation method, intuitively reflects the evolution rule of the object interest breadth in an electrocardiosignal-like mode, and effectively improves the intention recognition precision and the service prediction effect of a recommendation system. When recommending the target object, the predicted recommended content comprehensively considers the interest generalization and the potential interest prediction, and is not the content very similar to the historical click, so that the recommended content range is widened.
The specific flow is as follows:
And (one) a coarse ordering stage.
The coarse ordering stage may coarsely sift through the candidate set from the content pool based on object information of the target object. The object information may include multi-dimensional data such as social attributes, consumption habits, interest preferences, behavior preferences, and the like of the target object.
Wherein the coarse ordering stage may comprise the steps of:
acquiring object information of a target object;
Screening a plurality of candidate contents in the content pool based on the object information to obtain a candidate set;
the object information of the target object may be from an object information module, and may include social attribute features, historical behavior data, interest patterns, and the like of the target object.
Wherein the candidate content needs to include content that the target object may be interested in but has not yet contacted, in addition to content that exactly matches the interest of the target object. Therefore, after the classification tag of the content is directly and accurately matched with the interest tag in the object information, it is also necessary to add content that may be of interest to the target object but has not been touched.
Wherein, the content that may be of interest to the target object but has not been contacted (i.e. unexplored content) can be searched in the content library by various methods:
Excluding content that the target object has contacted from the content library;
Content in the content library that may be of interest to the target object but has not been contacted is screened.
Wherein, the database can be used for recording all object information of the user, and the contacted content can be automatically removed before recommendation.
Wherein, the associated untouched content can be matched in the content library as unexplored content according to the object information.
In some embodiments, similar objects for the target object may be found by user collaborative filtering, such that content of interest to the similar object is recommended as unexplored content to the target object in the content library.
In some embodiments, content that may be of interest to the target object but has not been contacted may be predicted in the content library by a neural network model.
And (II) scoring stage.
The scoring stage can acquire a historical behavior sequence of the target object, the historical behavior sequence can comprise behavior items arranged in time sequence, the behavior items represent interaction behaviors of the target object for content, an interest generalization curve is constructed based on similarity among the behavior items in the historical behavior sequence, the interest generalization curve reflects change of the interest generalization degree of the target object along with time, and recommendation content adapted to the target object is determined based on the interest generalization curve.
Specifically, the scoring stage may be divided into the following flows:
(1) Preprocessing of historical behavior sequences.
The last n click sequences S' = [ S1,s2,s3…sn ] of object a are acquired. Where s is a behavior item, including the content features and behavior features of object a. The content characteristics comprise a unique ID (identity) of the content read by the target object, a content type, a content primary category cat1 and a content secondary category cat2, and the behavior characteristics comprise access time, interaction duration, behavior type click and the like of the content.
Then, a series of preprocessing is performed on the sequence S' to obtain a historical behavior sequence S.
In some embodiments, the sequence S' may be subjected to data cleansing, such as deduplication, missing value processing, time ordering, and the like.
The de-duplication processing can reserve the last operation in a plurality of identical operations in a short time, and avoid the interference caused by repeated clicking of continuously clicking the same content in a short time.
Wherein the missing value process may fill in the missing field with a default value, such as marking the empty item as unknown, culling incomplete records, and so forth.
The time sequencing can be arranged according to the field ascending sequence of the access time, so that the behavior sequence is ensured to be strictly arranged according to the time occurrence sequence.
In some embodiments, the sequence S' may be feature engineered, such as by separately feature engineering the content features and the behavior features.
Wherein the content feature processing includes directly encoding the content ID of the high frequency and merging the content ID of the low frequency into a < UNK > category, and the content feature may also be mapped to a low-dimensional dense vector, or through an embedding layer.
Content feature processing also includes encoding classification features (type/cat 1/cat 2), such as encoding using One-Hot encoding (One-Hot) or word embedding (Embedding).
In some embodiments, if a hierarchical relationship exists, such as cat1→cat2, then it may be directly stitched into one dimension.
The behavior feature processing includes converting time features (time) into time stamps, calculating interval time_gap of adjacent behaviors, extracting periodic features, normalizing interaction duration (duration), classifying short/medium/long duration according to business logic, and directly encoding behavior types (click) into category labels (such as click=0, collection=1, share=2).
In some embodiments, sequence S' may be sequence constructed.
For example, the last n behaviors are reserved, the partial truncation is exceeded, and the < PAD > is filled when the partial truncation is insufficient;
For example, the multidimensional features of each behavior item are spliced into a unified vector, such as [ content ID_emb, type_emb, cata1_emb, cata2_emb, duration_norm, time_gap, click_type ];
For example, add behavioral position coding (e.g., position embedding of a transducer);
In some embodiments, the sequence S' may be normalized and stored.
For example, the duration, time _gap, etc., continuous values are normalized.
(2) And (5) constructing an interest generalization curve.
1. Acquiring the behavior liveness of a target object;
2. determining a window size=40 of the sliding window based on the activity level;
3. placing the sliding window in a historical behavior sequence to obtain a first subsequence S1＝[s1,s2,s3…s39;
4. Moving a sliding window in the historical behavior sequence by taking 1 as a step length to obtain a second subsequence S2＝[s2,s3,s4…s40 ];
5. determining similarity between pairs of parity items in the first subsequence S1 and the second subsequence S2 [cos(s1,s2),cos(s2,s3),cos(s3,s4),…,cos(s39,s40)];
6. Generating an interest generalization curve [1-cos(s1,s2),1-cos(s2,s3),1-cos(s3,s4),…,1-cos(s39,s40)], the interest generalization curve includes an interest generalization, interest generalization = 1-similarity.
The method for calculating the behavior activity of the target object is various, for example, in some embodiments, the behavior activity of the target object may be obtained by performing weighted summation on the login frequency, the online time length and the interaction times:
behavior liveness=k1 (login times/period) +k2 (total online time/period) +k3 (w1 praise+w2 comment+w3 share +..
Wherein, k 1-k 3 and w 1-w 3 are weights, which can be set according to actual demands.
Wherein the window size of the sliding window may be determined based on a relationship between liveness and liveness threshold. For example, liveness is 5 or more, the window size is set to the first size, 5> liveness is 10 or more, the window size is set to the second size, 10> liveness is 15 or more, the window size is set to the third size, 15> liveness is 20 or more, the window size is set to the fourth size, and so on.
In some embodiments, the behavior liveness of the target object may be updated in real time to obtain the latest behavior liveness, so that the window size may also be adjusted in real time.
For example, object A detects that liveness is continuously higher than the threshold value by clicking 20 times per minute for 3 minutes, the window is narrowed to 40, and subsequently, object A detects that liveness is continuously lower than the threshold value by clicking 3 times for the past 5 minutes, the window is enlarged to 200, and more behavior data is accumulated.
(3) And the prediction model determines recommended content adapted to the target object according to the interest generalization curve.
1. Acquiring content to be detected in the candidate set, wherein the content to be detected can comprise a domain label D;
2. Performing feature stitching on the content features and the behavior features of each behavior item in the historical behavior sequence S= [ S1,s2,s3…sn ] of the interest generalization curve [1-cos(s1,s2),1-cos(s2,s3),1-cos(s3,s4),…,1-cos(s39,s40)]、 to obtain a personal feature U of the target object;
3. A click type prediction network in a prediction model is adopted, and the click type of a target object aiming at the content to be detected is predicted based on the personal characteristic U and the domain label D of the content to be detected;
4. a reading duration prediction network in a prediction model is adopted, and the reading duration of a target object aiming at the content to be detected is predicted based on personal characteristics and a domain label of the content to be detected;
5. calculating a behavior index according to the click type and the reading time length;
6. determining recommendation scores of the content to be detected based on the behavior indexes;
7. and selecting recommended contents from the candidate set by sequencing the recommendation scores of all the contents to be detected in the candidate set.
The domain label may include a domain to which the content belongs, and may include a plurality of feature data that can indicate the domain to which the content belongs, such as a primary class, a secondary class, a type, and a theme.
The characteristic splicing modes are various and can comprise transverse splicing, longitudinal splicing, characteristic crossing, embedded splicing, sequence splicing and the like.
In some embodiments, the prediction model can construct a click type prediction training data set based on personal feature vectors, to-be-detected content field labels and corresponding click type labels extracted from historical behavior data, design a deep neural network structure formed by a full-connection layer and a softmax classifier, input the personal features and the field labels into a network after feature cross fusion, introduce a cross entropy loss function to measure the difference between predicted click type distribution and real labels, iteratively update network parameters by adopting an Adam optimization algorithm, and dynamically adjust learning rate through an exponential decay strategy to improve model convergence effect.
In some embodiments, training sample sets comprising personal characteristics, field labels and standardized reading time length correlation coefficients can be constructed based on user historical reading time length log data, a regression neural network architecture with nonlinear activation functions is built, user attributes and content field characteristics are fused through a feature splicing layer, deviation degree between predicted time length and real time length is measured by means of a mean square error loss function, network weight parameters are optimized by means of a gradient descent algorithm with L2 regularization, and overfitting is prevented by means of early shutdown.
In some embodiments, the click type prediction network and the reading duration prediction network may use independent training pipelines, complete model iterative optimization through small-batch data parallel computation under a distributed training framework, and finally generate a dual-task prediction model capable of synchronously outputting click type probability distribution and reading duration predicted values.
Wherein the click type classification distinguishes different interaction behaviors of the user, each type representing a different user intent. For example, active clicking may set the weight higher, passive clicking, such as false touching, may set the weight lower or negative, deep interactions, such as clicking comments, sharing buttons, etc., may set the weight higher.
In some embodiments, the behavior index may be set as:
Behavior index = Σ (click type weight×click number) +α×log (reading duration+1)
Wherein alpha is an adjustment coefficient of reading duration and is used for balancing the influence of clicking and duration.
In some embodiments, the behavioral indicators may be converted to normalized recommendation scores of 0-1 or 1-5 points.
And (III) a reordering stage.
The reordering stage can reorder the recommended content to realize diversity control and strategy intervention of the recommended content, and finally, the reordered recommended content is sent to the target object.
For example, to avoid overcommitted similarity of recommended content, diversity may be increased by computing similarity between content, introducing diversity constraints in the ranking. For example, maximum Marginal Relevance (MMR) may be introduced to comprehensively consider relevance and diversity of content in ranking. For example, cluster reordering may be introduced to cluster recommended content and then select content from different categories for reordering. For example, random sampling may be introduced, randomly scrambling portions of the content to increase diversity based on correlation.
For example, the recommendation order may be adjusted according to business needs or policies, and the ordering may be adjusted according to predefined rules. For example, different content or features may be weighted and reordered in combination with the original score. For example, the time-decay process may be performed on the content with higher timeliness, and new content may be recommended preferentially.
For example, the recommendation order may be optimized based on historical behavior of the user or real-time feedback. For example, a machine learning model may be used to predict user click probabilities, reordered by predicted values. For example, the order may be adjusted according to the matching degree between the target image and the content of the target object. For example, the recommendation order may be dynamically adjusted based on the user's real-time behavior.
For example, to balance various metrics such as click through rate, diversity, user satisfaction, etc., multi-objective optimization algorithms are used to reorder and dynamically adjust the ordering strategy through reinforcement learning to optimize long term revenue.
For example, the recommendation order may be adjusted according to the current application scenario of the user, such as time, place, device, etc.
According to the embodiment of the application, through analyzing the form of the interest curve, the preference characteristics of the target user can be deeply and insignificantly obtained, so that a more accurate content recommendation strategy is formulated. When the curve is observed to exhibit a higher average peak, this indicates that the user exhibits a tendency to diversify interest, with a broader coverage of content consumption behavior. For such users, it is desirable to employ diversified content recommendation strategies to meet their wide interest needs by providing cross-domain quality content. Conversely, when the curve shows a lower average trough, this tends to mean that the user has a more focused area of interest, which shows a stronger preference for a particular type of content. In this case, the recommendation system should emphasize mining the core points of interest of the user, providing the orthographic content that matches highly with its preferences. In addition, the trend change of the curve also contains important information, namely when the interest range shows an expanded situation, new content related to the existing interest of the user and having exploratory property can be properly introduced to guide the user to find potential interest, and when the interest range shows a contracted trend, the user needs to pay attention to the core interest field of the user, so that more accurate vertical content is provided to improve the viscosity of the user. The recommendation strategy based on the dynamic change of the interest curve can effectively improve the accuracy of content recommendation and the satisfaction degree of users.
Thus, in some embodiments, the steps performed by the reorder stage may include:
1. determining the interest generalization type of the target object based on the interest generalization curve;
2. If the target object is of a high-generalization type, the priority of sequencing the recommended content which accords with the high-generalization type is improved;
3. If the target object is of a low-generalization type, the priority of sequencing the recommended content which accords with the low-generalization type is improved;
4. Reordering all the recommended contents according to the priority of the sequencing of the recommended contents to obtain reordered recommended contents;
5. Determining a domain label of the target object which never generates interactive behaviors according to the historical behavior sequence of the target object, wherein the domain label represents the content domain to which the content belongs;
6. determining the interest generalization types of the target object in different time periods based on the interest generalization curves;
7. if the target object is of a high-generalization type in the target time period, selecting unexplored content in a content pool, wherein the unexplored content is content with a domain label which does not generate interaction behaviors of the target object;
8. Determining a mixing proportion weight according to the interest generalization curve, wherein the mixing proportion weight can be used for adjusting the proportion of recommended content and unexplored content in transmission;
9. and sending unexplored content at the same time when sending the recommended content to the target object according to the mixed proportion weight.
In some embodiments, when the interest generalization curve shows that the click rate of the target object on a plurality of fields such as politics, science and technology, entertainment is maintained above 0.8 and the variance of the curve is less than 0.1, the user is shown to have stable and wide news reading interest. At the moment, a cross-domain mixed recommendation strategy can be adopted, the multi-component contents such as international politics, AI technical breakthroughs, star dynamics and the like are combined in a single push card, local folk reports and sports event rapid messages are inserted in the information flow, and 3-5 different domains are covered in single push.
In some embodiments, when the interestingness curve fluctuates by more than 3 standard deviations within 15 minutes in the early season, the news hot tracking system can be accessed in real time, the instantaneous content of sudden financial data, major social events and the like can be preferentially pushed, and the recommendation weight of the deep report is reduced. If the curve presents a stable situation (fluctuation range < 0.2) lasting for 0.5 hours at night of working days, the curve is automatically switched to a thematic planning recommended mode, and continuous deep reports such as 'global climate peak series observation', 'artificial intelligence ethical ten-day talk' and the like are pushed.
In some embodiments, by analyzing the space feature vector of the interest generalization curve, when the attention duration of a user to the 'regional conflict' type label is detected to be 120% longer than that of the upper month, a dynamic region recommendation matrix can be constructed, wherein the content of direct hits of the war field reporter in the middle east region is preferentially pushed to be 40%, the pushing frequency of local municipal news is synchronously reduced, and the associated energy market analysis type content is recommended through interest map diffusion.
In some embodiments, in the news hot spot exploration mechanism, when the slope of the tail end of the curve exceeds a threshold (delta >0.7/10min, for example), a hot spot prediction engine is triggered, potential interest points are extracted based on the real-time reading track of a user (3 quantum calculation reports are continuously clicked, for example), a news semantic map is immediately invoked to generate a cross-dimension content package, and related news combination pushing such as "the latest breakthrough of quantum calculation" related enterprise stock price transaction "and" scientific research team special visit "is included, so that the exploration depth is improved by 2.3 times compared with a baseline.
In some embodiments, in a reordering strategy of news pushing, when the information entropy value of the interest generalization curve is detected to be higher than 80 quantiles of the history, a diversity enhancing algorithm is adopted, wherein the three types of contents including breaking news, background interpretation and expert comments are mixed and ordered according to the ratio of 4:3:3, so that the single feed stream is ensured to contain at least 2 untouched news tags. And meanwhile, a time attenuation factor is introduced, the ordering weight of the hot news before 30 minutes is reduced by 50%, and live image-text live broadcast content updated in real time is preferentially displayed.
In some embodiments, the region coverage can be dynamically adjusted by the initial slope of the generalization curve for the first week push of the new user, namely, when the slope of the curve rising first day is greater than 45 degrees, the 5:5 push ratio of global hot spot and local news is adopted, if the slope is less than 30 degrees, local news is focused to 70 percent of the ratio, and a small amount of national news with high interaction rate is embedded as an interest probe.
In some embodiments, in news interest attenuation early warning, when the curve duration index of a specific field (such as a sports event) continuously drops for 5 days for more than 20%, a content fresh-keeping mechanism is automatically triggered, namely, conventional event report pushing is suspended, related derivative content such as 'sports star crossing dynamics' and 'sports science and technology new release' is recommended instead, and the original content supply is gradually restored after the user interaction rate returns to a threshold value.
The recommendation strategy design based on the interest generalization curve has diversified realization paths, and personalized content pushing can be performed through various technical means and algorithm models. Specifically, the following various schemes may be employed:
For example, in some embodiments, a machine learning model may be used to construct a content scoring system, and various types of content may be quantitatively evaluated according to a popularity curve of interest, so that content with a higher score may be preferentially recommended to the target user. Meanwhile, by means of a behavior prediction model, future interest fields of the user can be predicted based on the interest generalization curve, and related subject matters can be pushed accordingly.
For example, in some embodiments, the user population may be subdivided according to the interest generalization curve by a clustering and classifying algorithm, and content matching the category features thereof may be screened from the content library for accurate pushing. In addition, in the conventional content pushing process, a certain proportion of new domain content can be properly blended according to the user classification result, so that the user interest boundary is expanded.
For example, in some embodiments, the volatility characteristics of the interestingness curve and the time decay effects may be combined, and the timeliness and diversity of the recommended content may be dynamically optimized. Specifically, the user interest stability is analyzed by dividing a time window, namely, when the curve is severely fluctuated, the content with strong timeliness such as hot news is pushed in a focused mode, and when the curve is stable, the depth content such as thematic articles is recommended.
For example, in some embodiments, the weight ratio of collaborative filtering to content filtering may be flexibly adjusted according to morphological features of the interest generalization curve. And when the curve peak value is lower than a preset value, strengthening the collaborative filtering weight, and pushing the preference content of the similar user in a weight mode. Meanwhile, the quick switching of the recommended strategy is realized through the change speed of the slope of the curve.
For example, in some embodiments, a user interest profile may be constructed based on the interest generalization curve to enable cross-domain content recommendation. The method comprises the steps of mapping a user history behavior sequence into interest nodes, and constructing a relevance map, wherein when an interest generalization curve is in an ascending trend, the content of adjacent interest fields is recommended, and when the curve is descending, core interest nodes are focused, so that excessive expansion is avoided.
In some embodiments, the probability of exploring new domain content is increased when the end of the object's interest generalization curve is at an ascent, and the exploration is decreased when the end of the curve is at a descent, preferentially recommending validated interest content.
In some embodiments, if the target object is of a high-generalization type, the content of the target object unexplored domain may be ranked in the front portion and the content of the domain of known interest may be ranked in the rear portion, if the target object is of a medium-generalization type, the content of the target object unexplored domain may be ranked in the balance with the content of the domain of known interest, and if the target object is of a low-generalization type, the content of the target object unexplored domain may be ranked in the rear portion and the content of the domain of known interest may be ranked in the front portion.
In some embodiments, unexplored content may be dynamically recommended by combining historical behavior sequences of objects, interest generalization curves, and time periods. For example, when the object is in a time period with wide interest and strong exploratory property, the content of the unexplored field, such as articles of 'travel' or 'art' class, which the object never clicks, can be recommended to the object more, and when the object is in a time period with concentrated interest and weak exploratory property, the content of the unexplored field can be recommended to the object less.
In some embodiments, if the target object is of the first generalization type in the target time period, selecting unexplored content in the content pool, wherein the unexplored content is content with a domain label which never generates interactive behaviors by the target object, and transmitting unexplored content simultaneously when transmitting recommended content to the target object in the target time period.
The duty cycle of sending unexplored content can be dynamically adjusted according to the interest flooding curve, so in some embodiments, a mixing proportion weight can be determined according to the interest flooding curve, the mixing proportion weight can be used for adjusting the proportion of the recommended content and the unexplored content when being sent, and the unexplored content is simultaneously sent when the recommended content is sent to the target object according to the mixing proportion weight.
For example, through the interest generalization curve, the interest generalization types of the object in different time periods such as early, middle, late and the like are identified in real time, and the recommendation strategy is dynamically adjusted. For example, in the morning commute period, the target object is of a high-generalization type, the target object may be in an open state of information receiving, and can recommend diversified contents such as international news and unexplored scientific and technological cold knowledge, in the noon break period, the target object is of a medium-generalization type, and can recommend moderately diversified contents such as easy entertainment, food, short reading and the like, and in the evening break period, the target object is of a low-generalization type, and can recommend accurate contents such as movie dramas, deep long texts and the like which the target normally sees.
For example, through the interest generalization curve, the interest generalization types of the object in different time periods of the working day and the non-working day are identified in real time, and the recommendation strategy is dynamically adjusted. For example, on monday to friday, the target object is in a working state, time-slicing is prone to efficiently acquire core interest contents related to work and study, so that the content in the history high-frequency interest field of the target object is recommended mainly by accurate matching, the ratio of the content in the unexplored field is limited, for example, on friday to friday, the target object is time-free and has more open interest, the content in the entertainment, leisure and masses fields is prone to be explored, the cross-field content is recommended for the target object, and the ratio of the unexplored field content is greatly improved.
The schemes jointly form a multi-level and multi-dimensional recommendation strategy system based on the interest generalization curve, and the accuracy of content recommendation and the satisfaction of users can be effectively improved.
Referring to fig. 2b, the ordinate is the value of the interest generalization, the abscissa is time, and the farther the curve is from the 0 axis, the more general interest the target object is represented. The closer the curve is to the 0 axis, the more focused the interest representing the target object. The interest generalization curve of the object A has a similarity mean value of 0.695 and an interest generalization mean value of 0.305, and the interest generalization mean value is larger than that of the object B, so that when recommending news to the object A, the matching degree of the interest is considered, and meanwhile, the diversity and the exploring degree of the interest are improved. The interest generalization curve of the object B has a similarity mean value of 0.798 and an interest generalization mean value of 0.202, and the interest generalization mean value is smaller than that of the object A, so that when news is recommended to the object B, the matching degree of the interest is more weighted, and the diversity and the exploratory degree of the interest are reduced.
According to the embodiment of the application, the hidden deep associated information in the target object behavior sequence is deeply mined, and the interest prediction capability of the recommendation model is remarkably enhanced by constructing the interest generalization curve. Experimental data show that the offline AUC is improved by 1.2% and the online CTR is improved by 0.8% in the news recommendation system.
From the above, the embodiment of the application provides an interest generalization curve, and quantifies the change rule of the interest breadth of the represented object along with time by constructing a similarity fluctuation curve of a time sequence, the embodiment of the application provides a sliding window mechanism, and the size of a sequence analysis window is adaptively adjusted according to the activity of the object, the dual-network prediction model architecture designed by the embodiment of the application integrates a click type prediction network and a reading duration prediction network to carry out comprehensive scoring, and the embodiment of the application establishes an interest generalization-driven diversity recommendation mechanism, and dynamically adjusts the exploration content proportion by mixing proportion weights.
Compared with the existing content recommendation method, the method widens recommendation diversity through interest generalization analysis, overcomes the information cocoon house effect caused by the traditional method, extracts deep features which are not utilized by the traditional method through a time sequence association mode of a deep mining behavior sequence, and improves the interpretability of a recommendation system by constructing a visual analysis tool with strong interpretability. Therefore, the embodiment of the application can improve the quality of content recommendation.
It will be appreciated that, in the specific embodiment of the present application, related data such as historical behavior sequences of target objects, object information, etc. are related, when the following embodiments of the present application are applied to specific products or technologies, permission or agreement needs to be obtained, and collection, use and processing of related data need to comply with related laws and regulations and standards of related countries and regions.
In order to better implement the method, the embodiment of the application also provides a content recommendation device, which can be integrated in an electronic device, wherein the electronic device can be a terminal, a server and other devices. The terminal can be a mobile phone, a tablet personal computer, an intelligent Bluetooth device, a notebook computer, a personal computer and other devices, and the server can be a single server or a server cluster consisting of a plurality of servers.
For example, in the present embodiment, a method according to an embodiment of the present application will be described in detail by taking a specific integration of a content recommendation device in a server as an example.
For example, as shown in fig. 3a, the content recommendation apparatus may include a history unit 310, a similarity unit 320, a construction unit 330, and a recommendation unit 340, as follows:
and (one) a history unit 310.
The history unit 310 may be configured to obtain a history behavior sequence of the target object, where the history behavior sequence may include behavior items arranged in time sequence, and the behavior items characterize interaction behavior of the target object with respect to the content.
And (two) a similarity unit 320.
The similarity unit 320 may be used to determine the similarity between two behavior items in a historical behavior sequence that are separated by a fixed interval.
As shown in fig. 3b, in some embodiments, the similarity unit 320 may include a sequence sub-unit 3201 and a similarity sub-unit 3202, wherein:
The sequence sub-unit 3201 may be configured to extract a first sub-sequence and a second sub-sequence from the historical behavior sequence, where the first sub-sequence includes the ith to jth behavior items in the historical behavior sequence, the second sub-sequence includes the ith to jth to kth behavior items in the historical behavior sequence, i and j are positive integers, and k is a fixed interval.
Among other things, in some embodiments, sequence subunit 3201 may be configured to:
acquiring the behavior liveness of a target object;
Determining a window size of the sliding window based on the activity level of the behavior;
Placing a sliding window in the historical behavior sequence to obtain a first subsequence, wherein the first subsequence comprises the ith to jth behavior items in the historical behavior sequence, and i and j differ by a window size;
and moving a sliding window in the historical behavior sequence by taking k as a step length to obtain a second subsequence.
The similarity subunit 3202 may be configured to determine a similarity between a first subsequence and a pair of parity terms in the second subsequence, where the pair of parity terms is two behavior terms corresponding to positions in the first subsequence and the second subsequence.
Wherein, in some embodiments, similarity subunit 3202 is to:
The parity item pair includes a first parity item in the first sub-sequence and a second parity item in the second sub-sequence, and determining a similarity between the first sub-sequence and the parity item pair in the second sub-sequence includes:
Determining a first set, wherein the first set comprises a first parity item and adjacent behavior items thereof in a first subsequence;
determining a second set comprising a second parity item and its neighboring behavior items in a second sub-sequence;
and obtaining the similarity between the first parity item and the second parity item according to the average value of the first set and the average value of the second set.
(III) construction unit 330.
The construction unit 330 may be configured to construct an interest generalization curve according to the similarity, where the interest generalization curve reflects a change of interest generalization with time, and the interest generalization characterizes an interest generalization degree of the target object.
And (fourth) a recommendation unit 340.
The recommendation unit 340 may be configured to determine recommended content adapted to the target object based on the interest generalization curve.
As shown in fig. 3c, in some embodiments, the recommendation unit 340 may include a to-be-detected subunit 331, a scoring subunit 332, and a ranking subunit 333, wherein:
(1) The sub-unit 331 is to be detected.
The sub-unit 331 to be detected may be used for acquiring contents to be detected in the candidate set.
(2) Scoring subunit 332.
The scoring subunit 332 may be configured to predict a recommendation score for the target object for the content to be detected based on the interest-in-generalization curve using a prediction model.
In some embodiments, the behavior items may include content features and behavior features, the content to be detected may include a domain label, and the scoring subunit 332 may include a stitching subunit 3321, a behavior index subunit 3322, and a scoring subunit 3323, where:
The splicing submodule 3321 can be used for carrying out characteristic splicing on the content characteristics and the behavior characteristics of each behavior item in the interest generalization curve and the historical behavior sequence to obtain the personal characteristics of the target object;
The behavior index submodule 3322 may be used for predicting a behavior index of a target object for the content to be detected based on the personal characteristics and the domain label of the content to be detected;
the scoring submodule 3323 may be used to determine a recommendation score for the content to be detected based on the behavioral indicators.
In some embodiments, the predictive model may include a click type prediction network and a reading duration prediction network, and the behavioral indicator sub-module 3322 may be configured to:
a click type prediction network is adopted, and the click type of a target object aiming at the content to be detected is predicted based on personal characteristics and the domain label of the content to be detected;
a reading duration prediction network is adopted, and the reading duration of a target object aiming at the content to be detected is predicted based on personal characteristics and the domain label of the content to be detected;
And calculating the behavior index according to the click type and the reading time.
(3) Ordering subunit 333.
The ranking subunit 333 may be configured to select recommended content in the candidate set by ranking the recommendation scores of all the content to be detected in the candidate set.
As shown in fig. 3d, in some embodiments, the recommendation unit 340 may further include an object information subunit 334, a coarse ordering subunit 335, and a reordering subunit 336, as follows:
(4) The object information subunit 334, the object information subunit 334 may be configured to obtain object information of the target object.
(5) The coarse ordering subunit 335 may be configured to screen a plurality of candidate contents from the content pool based on the object information to obtain a candidate set.
The recommendation unit 340 may further include:
(6) The reordering subunit 336 may be configured to reorder all recommended content and send the reordered recommended content to the target object.
In some embodiments, reorder subunit 336 may be configured to:
determining the interest generalization type of the target object based on the interest generalization curve;
If the target object is of the first generalization type, the priority of sequencing the recommended content conforming to the first generalization type is improved;
if the target object is of the second generalization type, the priority of sequencing the recommended content conforming to the second generalization type is improved;
and reordering all the recommended contents according to the priority of the sequencing of the recommended contents.
In some embodiments, the recommendation unit 340 may further include:
The non-interactive subunit can be used for determining a domain label of the target object which never generates interactive behaviors according to the historical behavior sequence of the target object, wherein the domain label represents the content domain of the content;
a type subunit, configured to determine, based on the interest generalization curve, an interest generalization type of the target object in different time periods;
The selecting subunit may be configured to select unexplored content in the content pool if the target object is of the first generalization type in the target time period, where the unexplored content is a content having a domain label in which the target object has never generated an interaction behavior;
the simultaneous transmission subunit may be configured to simultaneously transmit unexplored content when transmitting the recommended content to the target object in the target time period.
In some embodiments, the simultaneous transmit subunit may be configured to:
determining a mixing proportion weight according to the interest generalization curve, wherein the mixing proportion weight can be used for adjusting the proportion of recommended content and unexplored content in transmission;
And sending unexplored content at the same time when sending the recommended content to the target object according to the mixed proportion weight.
In the implementation, each unit may be implemented as an independent entity, or may be implemented as the same entity or several entities in any combination, and the implementation of each unit may be referred to the foregoing method embodiment, which is not described herein again.
In the present embodiment, the term "module" or "unit" refers to a computer program or a part of a computer program having a predetermined function and working together with other relevant parts to achieve a predetermined object, and may be implemented in whole or in part by using software, hardware (such as a processing circuit or a memory), or a combination thereof. Also, a processor (or multiple processors or memories) may be used to implement one or more modules or units. Furthermore, each module or unit may be part of an overall module or unit that incorporates the functionality of the module or unit.
It can be seen from the above that the content recommendation device of the embodiment obtains, by the history unit, a history behavior sequence of the target object, where the history behavior sequence includes behavior items arranged in time sequence, the behavior items represent interaction behavior of the target object with respect to the content, determines, by the similarity unit, similarity between two behavior items separated by a fixed interval in the history behavior sequence, constructs, by the construction unit, an interest-generalization curve, the interest-generalization curve reflecting a change of the interest generalization with time, the interest-generalization represents an interest-generalization degree of the target object, and determines, by the recommendation unit, recommended content adapted to the target object based on the interest-generalization curve.
Therefore, the embodiment of the application can improve the quality of content recommendation.
The embodiment of the application also provides electronic equipment which can be a terminal, a server and other equipment. The terminal can be a mobile phone, a tablet computer, an intelligent Bluetooth device, a notebook computer, a personal computer and the like, and the server can be a single server or a server cluster formed by a plurality of servers and the like.
In some embodiments, the content recommendation device may also be integrated in a plurality of electronic devices, for example, the content recommendation device may be integrated in a plurality of servers, and the content recommendation method of the present application is implemented by the plurality of servers.
In this embodiment, a detailed description will be given taking an example that the electronic device of this embodiment is a server, for example, as shown in fig. 4, which shows a schematic structural diagram of the electronic device according to the embodiment of the present application, specifically:
The electronic device may include one or more processing cores 'processors 410, one or more computer-readable storage media's memory 420, a power supply 430, an input module 440, and a communication module 450, among other components. Those skilled in the art will appreciate that the electronic device structure shown in fig. 4 is not limiting of the electronic device and may include more or fewer components than shown, or may combine certain components, or may be arranged in different components. Wherein:
The processor 410 is a control center of the electronic device, connects various parts of the entire electronic device using various interfaces and lines, and performs various functions of the electronic device and processes data by running or executing software programs and/or modules stored in the memory 420, and calling data stored in the memory 420, thereby performing overall detection of the electronic device. In some embodiments, processor 410 may include one or more processing cores, and in some embodiments, processor 410 may integrate an application processor that primarily handles operating systems, object interfaces, applications, etc., with a modem processor that primarily handles wireless communications. It will be appreciated that the modem processor described above may not be integrated into the processor 410.
The memory 420 may be used to store software programs and modules, and the processor 410 may perform various functional applications and data processing by executing the software programs and modules stored in the memory 420. The memory 420 may mainly include a storage program area that may store an operating system, application programs required for at least one function (such as a sound playing function, an image playing function, etc.), etc., and a storage data area that may store data created according to the use of the electronic device, etc. In addition, memory 420 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, memory 420 may also include a memory controller to provide processor 410 with access to memory 420.
The electronic device also includes a power supply 430 that provides power to the various components, and in some embodiments, the power supply 430 may be logically connected to the processor 410 via a power management system, such that charge, discharge, and power consumption management functions are performed by the power management system. Power supply 430 may also include one or more of any of a direct current or alternating current power supply, a recharging system, a power failure detection circuit, a power converter or inverter, a power status indicator, and the like.
The electronic device may also include an input module 440, which input module 440 may be used to receive input numeric or character information and to generate keyboard, mouse, joystick, optical or trackball signal inputs related to object settings and function control.
The electronic device may also include a communication module 450, and in some embodiments the communication module 450 may include a wireless module through which the electronic device may wirelessly transmit over a short distance, thereby providing wireless broadband internet access to the object. For example, the communication module 450 may be used to assist objects in e-mail, browsing web pages, accessing streaming media, and the like.
Although not shown, the electronic device may further include a display unit or the like, which is not described herein. In particular, in this embodiment, the processor 410 in the electronic device loads executable files corresponding to the processes of one or more application programs into the memory 420 according to the following instructions, and the processor 410 executes the application programs stored in the memory 420, so as to implement various functions as follows:
Acquiring a historical behavior sequence of a target object, wherein the historical behavior sequence comprises behavior items arranged in time sequence, and the behavior items represent interactive behaviors of the target object for contents;
determining the similarity between two behavior items which are separated by a fixed interval in the historical behavior sequence;
According to the similarity, an interest generalization curve is constructed, the interest generalization curve reflects the change of the interest generalization along with time, and the interest generalization represents the interest extensive degree of the target object;
and determining the recommended content adapted to the target object based on the interest generalization curve.
The specific implementation of each operation above may be referred to the previous embodiments, and will not be described herein.
From the above, the embodiment of the application can improve the quality of content recommendation.
Those of ordinary skill in the art will appreciate that all or a portion of the steps of the various methods of the above embodiments may be performed by instructions, or by instructions controlling associated hardware, which may be stored in a computer-readable storage medium and loaded and executed by a processor.
To this end, an embodiment of the present application provides a computer readable storage medium having stored therein a plurality of instructions capable of being loaded by a processor to perform the steps of any of the content recommendation methods provided by the embodiments of the present application. For example, the instructions may perform the steps of:
Acquiring a historical behavior sequence of a target object, wherein the historical behavior sequence comprises behavior items arranged in time sequence, and the behavior items represent interactive behaviors of the target object for contents;
determining the similarity between two behavior items which are separated by a fixed interval in the historical behavior sequence;
According to the similarity, an interest generalization curve is constructed, the interest generalization curve reflects the change of the interest generalization along with time, and the interest generalization represents the interest extensive degree of the target object;
and determining the recommended content adapted to the target object based on the interest generalization curve.
The storage medium may include a Read Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, an optical disk, or the like.
According to one aspect of the present application, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions to cause the electronic device to perform the methods provided in various alternative implementations of the content push aspect or the message stream push aspect of news, video, articles, advertisements, etc., provided in the above-described embodiments.
The instructions stored in the storage medium may perform steps in any content recommendation method provided by the embodiments of the present application, so that the beneficial effects that any content recommendation method provided by the embodiments of the present application can be achieved are detailed in the previous embodiments, and are not repeated here.
The foregoing describes a content recommendation method, apparatus, electronic device and computer readable storage medium, and the principles and embodiments of the present application are described herein by applying specific examples, which are provided to facilitate understanding of the method and core ideas of the present application, and meanwhile, according to the ideas of the present application, those skilled in the art should not understand the limitation of the present application in terms of the specific embodiments and application scope.