Disclosure of Invention
In order to solve the problem that objective and quantized point value ordering cannot be recommended to advertisers for preferential selection in the existing elevator advertisement point selection process, the invention aims to provide an elevator advertisement point recommendation method, device, computer equipment and computer readable storage medium, which can provide an objective and quantized point value ordering for advertisers to facilitate preferential selection of elevator advertisement points, so that time required for selection of elevator advertisement points is shortened while advertisement delivery deviation and advertisement effect are avoided, customer experience and efficiency of the whole advertisement delivery process are improved, and practical application and popularization are facilitated.
In a first aspect, the invention provides an elevator advertisement point recommendation method, which comprises the following steps:
acquiring first target crowd label information of an advertiser, wherein the first target crowd label information comprises at least one point position label appointed by the advertiser according to a target crowd positioning result of an advertisement to be thrown;
for each elevator advertisement point in a plurality of elevator advertisement points to be selected, taking the corresponding tag actual value on each point tag in the first target crowd tag information as input data, inputting the input data into a multi-tag fusion model which is based on a neural network and has been learned, and outputting to obtain a corresponding first multi-tag fusion target audience concentration value;
Calculating the product of the corresponding first multi-label fusion target audience concentration value and the corresponding touch ability estimated value aiming at each elevator advertisement point to be selected to obtain a corresponding first point value score;
and sequentially sequencing the plurality of elevator advertisement spots to be selected according to the high-to-low sequence of the first spot value scores to obtain a first spot list for recommending the elevator advertisement spots to the advertiser.
Based on the above summary of the invention, an elevator advertisement point recommendation scheme based on a neural network algorithm is provided, namely after at least one point label specified by an advertiser according to a target crowd positioning result of an advertisement to be thrown is obtained, each of a plurality of elevator advertisement points to be selected is firstly aimed at, a corresponding multi-label fusion target audience concentration value is obtained according to a corresponding label actual value and a multi-label fusion model which is based on the neural network and is learned, and the product of the multi-label fusion target audience concentration value and a corresponding touch capability estimated value is calculated, a corresponding point value score is obtained, then the plurality of elevator advertisement points to be selected are sequentially sequenced according to a high-to-low order of the point value score, a point list for recommending the elevator advertisement points to the advertiser is obtained, and then an objective quantized point value sequencing is provided for the advertiser to facilitate the preferred selection of the elevator advertisement points, so that the time required for selecting the elevator advertisement points is shortened while avoiding causing unsatisfactory advertisement throwing bias and advertisement effect, the customer experience and the whole advertisement throwing process are improved, and practical application is facilitated.
In one possible design, the first target crowd label information further includes label condition values on each of the at least one point location label, and for each of a plurality of elevator advertisement points to be selected, the label actual values corresponding to and on each of the first target crowd label information are used as input data to be input into a neural network-based and learning-completed multi-label fusion model, including:
according to the label condition values on all the point labels in the first target crowd label information, carrying out query processing on all the elevator advertisement points to be selected to obtain a plurality of elevator advertisement points to be selected, which meet the label condition intersection of all the point labels in the first target crowd label information;
and aiming at each elevator advertisement point position to be selected in the plurality of elevator advertisement point positions to be selected, taking the corresponding tag actual value on each point position tag in the first target crowd tag information as input data, and inputting the input data into a multi-tag fusion model which is based on a neural network and has been learned.
In one possible design, before inputting the multi-label fusion model that is based on the neural network and has completed learning, the method further includes the following steps S201 to S208:
S201, acquiring second target crowd label information of a cast advertiser and at least one cast elevator advertisement point position, wherein the second target crowd label information comprises at least one point position label specified by the cast advertiser according to a target crowd positioning result of the cast advertiser and label condition values on each point position label;
s202, according to the label condition values of all the point labels in the second target crowd label information, carrying out query processing on all the alternative elevator advertisement points to obtain a plurality of alternative elevator advertisement points which accord with the label condition intersection of all the point labels in the second target crowd label information;
s203, aiming at each alternative elevator advertisement point in the plurality of alternative elevator advertisement points, taking the corresponding tag actual value on each point tag in the second target crowd tag information as input data, and inputting the input data into a neural network model to obtain a corresponding second multi-tag fusion target audience concentration value;
s204, calculating products of the corresponding second multi-label fusion target audience concentration values and the corresponding touch ability estimated values aiming at the alternative elevator advertisement points to obtain corresponding second point value scores;
S205, sequentially sequencing the plurality of alternative elevator advertisement points according to the sequence from high to low of the second point value scores to obtain a second point list for recommending elevator advertisement points to the cast advertisers;
s206, calculating a loss function according to the comparison results of the plurality of alternative elevator advertisement points with the at least one cast elevator advertisement point and the first N alternative elevator advertisement points in the second point list, wherein N represents the point count of the at least one cast elevator advertisement point;
s207, correcting network parameters of the neural network model through a back propagation algorithm according to a loss function calculation result to obtain a new neural network model;
s208, for another advertiser, returning to execute steps S201-S207 based on the new neural network model until the loss function calculation result reaches a preset target value or the steps S201-S207 are executed for all the advertisers, and taking the new neural network model as a multi-label fusion model which is learned and used for outputting a multi-label fusion target audience concentration value according to input data.
Based on the possible design, the multi-label fusion model can continuously perform self-learning based on the point position label appointed by the cast advertiser according to the target crowd positioning result of the cast advertisement, the label condition value on the point position label and the elevator advertisement point position selection result, so that the neural network parameters are continuously optimized, the prediction accuracy of the multi-label fusion target audience concentration value is improved, and the accuracy of the subsequent point position value sequencing is further ensured
In one possible design, calculating the loss function according to the comparison results of the plurality of candidate elevator advertisement spots with the at least one cast elevator advertisement spot and the first N candidate elevator advertisement spots in the second spot list, respectively, includes:
judging whether each alternative elevator advertisement point in the plurality of alternative elevator advertisement points belongs to the at least one cast elevator advertisement point, if so, recording a corresponding first comparison result value as 1, otherwise, recording a corresponding first comparison result value as 0;
judging whether the elevator advertisement points belong to the first N alternative elevator advertisement points in the second point list according to each alternative elevator advertisement point in the plurality of alternative elevator advertisement points, if so, recording a corresponding second comparison result value as 1, otherwise, recording a corresponding second comparison result value as 0, wherein N represents the total number of the points of the at least one cast elevator advertisement point;
The Loss function Loss is calculated using the following formula:
wherein M represents the total number of points in the plurality of alternative elevator advertisement points, M represents a natural number, ym A first comparison result value representing an mth alternative elevator ad spot of the plurality of alternative elevator ad spots, am And a second comparison result value representing the m-th alternative elevator advertisement point position.
In one possible design, for each of the candidate elevator advertisement points, calculating a product of the corresponding first multi-tag fusion target audience concentration value and the corresponding estimated touch capability value to obtain a corresponding first point value score, including:
and calculating to obtain a touch capability estimated value CPR of the advertisement point of the elevator to be selected according to the following formula:
wherein H represents the average number of floors of the elevator belonging to the elevator advertisement point position to be selected in the running floor zone, L represents the floor number in the running floor zone, R represents the occupancy rate in the running floor zone, F represents the current periodical rate of the elevator advertisement point position to be selected, and B represents a preset coefficient;
and calculating the point value score P of the advertisement point of the elevator to be selected according to the following formula:
P＝ηTA *CPR
wherein eta isTA And the first multi-label fusion target audience concentration value of the advertisement point position of the elevator to be selected is represented.
In one possible design, before calculating the estimated touchdown ability CPR for the selected elevator ad spot, the method further comprises:
for each elevator advertisement point in a plurality of elevator advertisement points, acquiring corresponding in-car video data acquired on a plurality of continuous days, wherein the elevator advertisement points are arranged in the area of the elevator advertisement point to be selected, the in-car video data are acquired by an in-car top camera installed in an elevator car to which the corresponding elevator advertisement point belongs, and the field of view of the in-car top camera covers the in-car ground area;
for each elevator advertisement point, according to the corresponding video data in the carriage, firstly adopting a human head detection model to carry out human head identification processing on video frame images in the carriage, and then carrying out touch human number statistics processing according to human head identification results to obtain corresponding single advertisement daily average touch human number;
for each elevator advertisement point position, calculating to obtain a corresponding coefficient value b according to the following formula:
wherein h represents the average number of floors of the elevator belonging to the corresponding elevator advertisement point in the operation floor zone, l represents the number of floors of the elevator belonging to the corresponding elevator advertisement point in the operation floor zone, r represents the occupancy rate of the elevator belonging to the corresponding elevator advertisement point in the operation floor zone, cpr represents the average daily contact time of a single advertisement corresponding to the elevator advertisement point, and f represents the average daily periodical rate of the corresponding elevator advertisement point in the continuous multiple days;
Fitting all the calculated coefficient values B to obtain a preset coefficient B of the advertisement point position of the elevator to be selected.
Based on the possible design, the preset coefficient in the touch capability estimation formula can be accurately measured, the accuracy of the subsequent touch capability estimation result is improved, and the accuracy of the final point position value ordering is further ensured.
In one possible design, for each elevator advertisement point, according to the corresponding video data in the car, first, a head detection model is adopted to perform head recognition processing on video frame images in the car, then, touch time statistics processing is performed according to a head recognition result, and corresponding single advertisement day average touch time is obtained, including:
extracting one frame from the video data in the carriage in unit time to obtain multi-frame video frame images in the carriage, wherein the unit time is smaller than the single playing time of the advertisement;
aiming at each frame of intra-frame video frame image in the multi-frame intra-frame video frame images, adopting a head detection model to carry out head recognition processing to obtain the number of people in the corresponding image;
aiming at each advertisement played in the elevator advertisement point, taking the maximum number of people obtained by recognition as the corresponding number of people touching in the single playing time period according to at least one frame of video frame image in the carriage corresponding to the single playing time period;
For each day of the continuous multiple days, firstly accumulating and calculating the number of times of the played advertisements in the corresponding days in all playing time slots, and then dividing the number of times of the played advertisements in the corresponding days by the total number of the carousel advertisements in the corresponding days to obtain the total number of times of the played advertisements in the corresponding single day;
and carrying out average value calculation processing on all the single-day single-advertisement total touch times to obtain single-advertisement-day average touch times corresponding to the elevator advertisement points.
The invention provides an elevator advertisement point position recommending device, which comprises a label information acquisition module, a target audience concentration estimating module, a point position value scoring calculating module and a point position ordering module which are connected in sequence in a communication mode;
the tag information acquisition module is used for acquiring first target crowd tag information of an advertiser, wherein the first target crowd tag information comprises at least one point location tag designated by the advertiser according to a target crowd positioning result of an advertisement to be placed;
the target audience concentration estimation module is used for inputting the corresponding tag actual values on the tag of each point in the first target crowd tag information into a multi-tag fusion model which is based on a neural network and has been learned, and outputting the corresponding first multi-tag fusion target audience concentration value for each elevator advertisement point in a plurality of elevator advertisement points to be selected;
The point value scoring calculation module is used for calculating the product of the corresponding first multi-label fusion target audience concentration value and the corresponding touch capability estimated value aiming at each elevator advertisement point to be selected to obtain a corresponding first point value score;
and the point position ordering module is used for sequentially ordering the plurality of elevator advertisement points to be selected according to the high-to-low sequence of the first point position value scores to obtain a first point position list for recommending the elevator advertisement points to the advertiser.
In a third aspect, the present invention provides a computer device, comprising a memory and a processor communicatively connected, wherein the memory is configured to store a computer program, and the processor is configured to read the computer program and execute the elevator advertisement spot recommendation method according to the first aspect or any possible design of the first aspect.
In a fourth aspect, the present invention provides a computer readable storage medium having instructions stored thereon which, when run on a computer, perform the elevator ad spot recommendation method as described in the first aspect or any of the possible designs of the first aspect.
In a fifth aspect, the invention provides a computer program product comprising instructions which, when run on a computer, cause the computer to perform the elevator ad spot recommendation method according to any of the above first or any of the possible designs of the first aspect.
Detailed Description
The invention will be further elucidated with reference to the drawings and to specific embodiments. The present invention is not limited to these examples, although they are described in order to assist understanding of the present invention. Specific structural and functional details disclosed herein are merely representative for purposes of describing example embodiments of the present invention. This invention may, however, be embodied in many alternate forms and should not be construed as limited to the embodiments set forth herein.
It should be understood that although the terms first and second, etc. may be used herein to describe various objects, these objects should not be limited by these terms. These terms are only used to distinguish one object from another. For example, a first object may be referred to as a second object, and similarly a second object may be referred to as a first object, without departing from the scope of example embodiments of the invention.
It should be understood that for the term "and/or" that may appear herein, it is merely one association relationship that describes an associated object, meaning that there may be three relationships, e.g., a and/or B, may represent: three cases of A alone, B alone or both A and B exist; for the term "/and" that may appear herein, which is descriptive of another associative object relationship, it means that there may be two relationships, e.g., a/and B, it may be expressed that: the two cases of A and B exist independently or simultaneously; in addition, for the character "/" that may appear herein, it is generally indicated that the context associated object is an "or" relationship.
As shown in fig. 1, the elevator advertisement spot recommendation method provided in the first aspect of the present embodiment may be performed by, but not limited to, a computer device with a certain computing resource, for example, a personal computer (Personal Computer, PC, refer to a multipurpose computer with a size, price and performance suitable for personal use, and desktop, notebook to small notebook, tablet, super-notebook etc. all belong to personal computers), a smart phone, a personal digital assistant (Personal digital assistant, PAD) or an electronic device such as a wearable device, so as to provide an objective and quantized spot value ordering for an advertiser to facilitate the preferred selection of an elevator advertisement spot, so that the time required for selecting an elevator advertisement spot is shortened while avoiding causing an advertisement placement deviation and an advertisement effect to be not ideal, the customer experience and the whole advertisement placement process efficiency are improved, and the actual application and popularization are facilitated. As shown in fig. 1, the elevator advertisement spot recommendation method may include, but is not limited to, the following steps S1 to S4.
S1, acquiring first target crowd label information of an advertiser, wherein the first target crowd label information comprises, but is not limited to, at least one point position label designated by the advertiser according to a target crowd positioning result to be advertised.
In the step S1, the advertiser is a potential client with respect to the advertisement server; the point location tag is a point location index designated by the advertiser according to the Target crowd positioning result of the advertisement to be advertised and used for estimating a concentration value of a multi-tag fusion Target audience (abbreviated as TA), and specifically may, but not limited to, include indexes such as a building average price, a female duty ratio, and/or a duty ratio above the family. The multi-tag fusion target audience concentration value is used for representing the ratio of the number of target people to the total number of people determined at the elevator advertisement point based on fusion results of a plurality of point location tags, so that after the estimated value of the touch ability of the elevator advertisement point is multiplied, the point location value score of the elevator advertisement point relative to the advertisement. In addition, the first target crowd label information can be input by the advertiser through a man-machine interaction mode, but is not limited to the first target crowd label information.
S2, aiming at each elevator advertisement point position to be selected in a plurality of elevator advertisement point positions to be selected, taking the corresponding actual value of the label on each point position label in the label information of the first target crowd as input data, inputting the actual value into a multi-label fusion model which is based on a neural network and is studied, and outputting the actual value to obtain the corresponding first multi-label fusion target audience concentration value.
In the step S2, the plurality of candidate elevator advertisement points may be all current optional elevator advertisement points, or may be part of the current optional elevator advertisement points. The actual value of the tag is a statistical result which is obtained through offline investigation and is on a point label aiming at the elevator advertisement point to be selected, for example, the actual value of the tag can be 3 ten thousand RMB/square meter on the point label of the average price of a building aiming at a certain elevator advertisement point to be selected; on a female duty cycle point label, the actual value of the label can be 70%; on a spot label above the family, the actual value of the label may be 60%, etc. Considering that the multi-tag fusion target audience concentration is closely related to the product positioning of the advertiser (i.e. the target crowd positioning result of the advertisement product), namely, factors such as building average price, gender proportion, age distribution, professional distribution, academic distribution, school age child proportion, income distribution, automobile family proportion and the like, although a plurality of point tags can be set based on these factors, the conventional multi-tag fusion target audience concentration estimation is usually performed by performing linear calculation processing on each point tag respectively, and then taking intersection or union of all processing results to obtain a final estimation result. The neural network is a complex network system formed by a large number of and simple processing units (called neurons) which are widely connected with each other, reflects a plurality of basic characteristics of human brain functions, is a highly complex nonlinear power learning system, and can obtain the multi-label fusion model for outputting multi-label fusion target audience concentration values according to input data through a conventional learning mode, so that after corresponding data is input to advertisement points of an elevator to be selected, corresponding information such as the multi-label fusion target audience concentration values used as estimation results can be output. In particular, the neural network may be, but is not limited to, a two-class neural network, a fully-connected neural network, a back propagation BP (Back Propagation) network, a Hopfield network, an adaptive resonance theory ART (Adaptive Resonance Theory) network, a Kohonen network, or the like. In addition, in order to make the input data meet the model input requirement, the actual value of the label is usually required to be input into the multi-label fusion model after conventional normalization processing.
In the step S2, in order to reduce the number of the elevator advertisement spots to be selected, the operand may be reduced, and the plurality of elevator advertisement spots to be selected may be obtained by first screening according to the condition specified by the advertiser for all the current selectable elevator advertisement spots, that is, when the first target crowd label information further includes the label condition value on each of the at least one point label, the corresponding actual label value on each of the first target crowd label information is used as input data for each of the plurality of elevator advertisement spots to be selected, and the obtained label actual value is input into the neural network-based and learning-completed multi-label fusion model, including, but not limited to, the following steps S21 to S22.
S21, according to the label condition values of all the point labels in the first target crowd label information, carrying out query processing on all the elevator advertisement points to be selected, and obtaining a plurality of elevator advertisement points to be selected, which accord with the label condition intersection of all the point labels in the first target crowd label information.
In the step S21, the tag condition value is a point screening condition specified by the advertiser according to the target crowd positioning result of the advertisement to be advertised, for example, on a point tag of a building average price, the tag condition value may be greater than 2 ten thousand; on a female duty cycle point label, the label demand value may be greater than 60%; on the point label with the ratio above the family, the label requirement value can be more than 50%, and the like, 1000 elevator advertisement points to be selected can be obtained through a conventional query processing mode (for example, the intersection of all label conditions is taken), and then the multi-label fusion target audience concentration value is estimated for each elevator advertisement point in the 1000 elevator advertisement points to be selected.
S22, aiming at each elevator advertisement point position to be selected in the plurality of elevator advertisement point positions to be selected, taking the corresponding tag actual value on each point position tag in the first target crowd tag information as input data, and inputting the input data into a multi-tag fusion model which is based on a neural network and has been learned.
S3, calculating products of the corresponding first multi-label fusion target audience concentration values and the corresponding touch ability estimated values aiming at the elevator advertisement points to be selected, and obtaining corresponding first point value scores.
In the step S3, the estimated touch capability value is used to represent the number of touch people corresponding to the advertisement point in a unit time and aiming at a single advertisement under a multi-advertisement carousel mechanism, and may be obtained in advance through offline investigation statistics, for example, aiming at the elevator advertisement point, the elevator advertisement machine is considered to play the advertisement in a form of sound+picture, so for the on-playing advertisement, when a passenger enters the elevator, the advertisement sound affects the passenger, namely, the passenger is considered to be touched; when a person takes the ladder for multiple times in different advertisement playing time slots of a certain advertisement, the person can be regarded as multiple touch, so that the method is not limited to the method that the single advertisement day of the advertisement point of the elevator to be selected is touched by the person for multiple times as a corresponding touch capability estimation result.
In the step S3, specifically, for each of the candidate elevator advertisement points, a product of the corresponding first multi-tag fusion target audience concentration value and the corresponding estimated touch capability value is calculated to obtain a corresponding first point value score, which includes, but is not limited to, the following steps S31 to S32.
S31, calculating to obtain a touch ability estimated value CPR of the advertisement point of the elevator to be selected according to the following formula:
wherein H represents the average number of floors of the elevator belonging to the advertisement point position of the elevator to be selected in the running floor zone, L represents the floor number in the running floor zone, R represents the occupancy rate in the running floor zone, F represents the current periodical rate of the advertisement point position of the elevator to be selected, and B represents a preset coefficient.
In the step S31, the single advertisement hit number considering the elevator advertisement spot is related to the following factors: (A) The elevator is in linear correlation with the average number of floors of the elevator belonging to the elevator advertisement point in the running floor interval, namely, the more the average number of floors is, the more the single advertisement is contacted with the people; (B) The number of floors of the elevator belonging to the elevator advertisement point is linearly related to the number of floors in the operation floor section, namely, the more the number of floors is, the more the number of households is, and the more the number of people is reached by a single advertisement; (C) The time length of the elevator is related to the time length of the elevator, and the time length of the elevator is linearly related to the number of floors, namely, the more the number of floors is, the longer the time length of the elevator is, the more the single advertisement is touched to the number of people; (D) The elevator service life is linearly related to the service life rate of the elevator belonging to the elevator advertisement point in the operation floor zone, namely, the higher the service life rate is, the more passengers take the elevator, and the more single advertisement calls; (E) The method is inversely proportional to the periodical rate of the elevator advertisement points, namely the periodical period of advertisement carousel is shortened when the elevator advertisement points are not full of the elevator advertisement points, the number of times of carousel is increased, so that single advertisement contact is increased to a human number, for example, if the single playing time of the advertisement is 15 seconds, 12 advertisements in carousel are full of the elevator advertisement points (namely the periodical rate is 100%), 180 seconds are needed for playing one round, if the periodical rate is 50% (namely only 6 advertisements are periodical), only 90 seconds are needed for playing one round, the number of times of single advertisement play is doubled, and therefore the single advertisement contact is increased to a human number. Therefore, after the average number of households, the number of floors and the occupancy rate of the elevator belonging to the elevator advertisement point to be selected in the running floor interval are obtained through investigation, the current periodical rate of the elevator advertisement point to be selected is obtained through the calculation formula, and the estimated touch ability value CPR of the elevator advertisement point to be selected is obtained. In addition, the preset coefficient B is a fixed coefficient related to travel habits, the number of floors, and the like, and may be preset according to experience, or may be preset according to an investigation result.
S4, sequentially sequencing the plurality of elevator advertisement points to be selected according to the high-to-low sequence of the first point value scores to obtain a first point list for recommending the elevator advertisement points to the advertiser.
In the step S4, because the multi-label fusion target audience concentration value and the touch capability estimated value of each elevator advertisement point to be selected are more or less different from the corresponding values of other elevator advertisement points to be selected, the point value scores of the elevator advertisement points to be selected are also different, and the plurality of elevator advertisement points to be selected are sequentially ordered according to the order from high to low of the first point value score, so that an objective and quantized point value ordering is obtained, thereby facilitating the advertisement master-slave to select satisfactory elevator advertisement points in the first point list, shortening the time required for selecting the elevator advertisement points, and improving the customer experience and the efficiency of the whole advertisement putting process. For example, in the first point location list, if 1000 elevator advertisement points to be selected are sequentially ordered according to the order from high to low of the first point location value score, the advertiser can sequentially determine whether each elevator advertisement point to be selected is satisfied according to the order until the number of the elevator advertisement points to be selected reaches the required point number (for example, 100), so that a plurality of elevator advertisement points to be satisfied are selected without finishing looking at the 1000 elevator advertisement points to be selected, the purpose of shortening the time required for selecting the elevator advertisement points is achieved, and the customer experience is greatly improved.
According to the elevator advertisement point recommending method described in the steps S1-S4, an elevator advertisement point recommending scheme based on a neural network algorithm is provided, namely after at least one point label appointed by an advertiser according to a target crowd positioning result of an advertisement to be thrown is obtained, each elevator advertisement point in a plurality of elevator advertisement points to be selected is firstly predicted according to a corresponding label actual value and a multi-label fusion model which is based on the neural network and is learned, a corresponding multi-label fusion target audience concentration value is obtained, the product of the multi-label fusion target audience concentration value and a corresponding accessibility estimated value is calculated, a corresponding point value score is obtained, then the plurality of elevator advertisement points to be selected are sequentially sequenced according to a high-to-low order of the point value score, a point list for recommending the elevator advertisement points to the advertiser is obtained, and further objective quantitative point value sequencing is provided for the advertiser to facilitate the preferential selection of the elevator advertisement points, so that the time required for elevator advertisement point selection and the actual application experience are shortened while the advertisement point selection and the advertisement point position advertisement point placement efficiency are shortened while the advertisement effect is avoided from being unsatisfactory.
The present embodiment further provides a possible design of how to learn to obtain the multi-label fusion model based on the technical solution of the first aspect, that is, as shown in fig. 2, before inputting the multi-label fusion model that is based on the neural network and has been learned, the method further includes, but is not limited to, the following steps S201 to S208.
S201, obtaining second target crowd label information of the cast advertiser and at least one cast elevator advertisement point position, wherein the second target crowd label information comprises, but is not limited to, at least one point position label specified by the cast advertiser according to the target crowd positioning result of the cast advertisement and label condition values on each point position label.
In the step S201, the posted advertiser is a transacted client corresponding to the advertisement server; the description of the point location tag and the tag condition value in the tag information of the second target crowd can be referred to the corresponding description content of the first aspect, and will not be repeated here; the at least one cast elevator advertisement point is all elevator advertisement points selected by the cast advertisers and used for casting the cast advertisements.
S202, according to the label condition values of all the point labels in the second target crowd label information, carrying out query processing on all the alternative elevator advertisement points to obtain a plurality of alternative elevator advertisement points which accord with the label condition intersection of all the point labels in the second target crowd label information.
In the step S202, the specific query processing manner may refer to the step S21 of the first aspect, which is not described herein.
S203, aiming at each alternative elevator advertisement point position in the plurality of alternative elevator advertisement points, taking the corresponding tag actual value on each point position tag in the second target crowd tag information as input data, and inputting the input data into a neural network model to obtain a corresponding second multi-tag fusion target audience concentration value.
In the step S203, although the neural network model is the multi-label fusion model that has not yet completed learning, the specific process description may also refer to the step S2 of the first aspect, which is not described herein again.
S204, calculating products of the corresponding second multi-label fusion target audience concentration values and the corresponding touch ability estimated values according to the alternative elevator advertisement points to obtain corresponding second point value scores.
In the step S204, the specific calculation process can refer to the step S3 of the first aspect, which is not described herein.
S205, sequentially sequencing the plurality of alternative elevator advertisement spots according to the order from high to low of the second spot value scores to obtain a second spot list for recommending elevator advertisement spots to the cast advertisers.
In the step S205, the specific sorting manner can be referred to as step S4 of the first aspect, which is not described herein.
S206, calculating a loss function according to comparison results of the plurality of alternative elevator advertisement points and the at least one cast elevator advertisement point and the first N alternative elevator advertisement points in the second point list, wherein N represents the point count of the at least one cast elevator advertisement point.
In the step S206, specifically, according to the comparison results of the multiple candidate elevator advertisement points with the at least one cast elevator advertisement point and the first N candidate elevator advertisement points in the second point list, a loss function is calculated, including but not limited to the following steps S2061 to S2063.
S2061, judging whether the elevator advertisement points belong to the at least one elevator advertisement point according to each alternative elevator advertisement point in the plurality of alternative elevator advertisement points, if so, recording a corresponding first comparison result value as 1, otherwise, recording a corresponding first comparison result value as 0.
S2062, judging whether the elevator advertisement points belong to the first N alternative elevator advertisement points in the second point list according to each alternative elevator advertisement point in the plurality of alternative elevator advertisement points, if so, recording a corresponding second comparison result value as 1, otherwise, recording a corresponding second comparison result value as 0, wherein N represents the total number of the points of the at least one cast elevator advertisement point.
In the steps S2061 to S2062, if a certain alternative elevator advertisement spot belongs to the first N alternative elevator advertisement spots and the at least one cast elevator advertisement spot at the same time, the alternative elevator advertisement spot is indicated to be recommended preferentially and finally selected by the cast advertiser, and the prediction is indicated to be correct; if the alternative elevator advertisement point does not belong to the first N alternative elevator advertisement point positions and the at least one cast elevator advertisement point position at the same time, the alternative elevator advertisement point position is not recommended preferentially and finally is also removed by the cast advertiser, and the prediction can be indicated to be correct. In addition, a prediction error is indicated.
S2063, calculating a Loss function Loss by adopting the following formula:
wherein M represents the total number of points in the plurality of alternative elevator advertisement points, M represents a natural number, ym A first comparison result value representing an mth alternative elevator ad spot of the plurality of alternative elevator ad spots, am And a second comparison result value representing the m-th alternative elevator advertisement point position.
In the step S2063, the Loss function Loss is conventionally modified by referring to the two-class cross entropy Loss function, and in this case, the neural network model needs to adopt a model based on the two-class neural network structure as shown in fig. 3, and the point location tag needs to be subjected to one-hot encoding so as to input the corresponding tag actual value.
S207, correcting network parameters of the neural network model through a back propagation algorithm according to a loss function calculation result to obtain a new neural network model.
In the step S207, the back propagation algorithm is a learning algorithm suitable for a multi-layer neural network, and is mainly implemented by repeating loop iteration of two links (i.e., an excitation propagation link and a weight update link) until the response of the neural network to the input reaches a predetermined target range, so that the network parameters of the neural network model can be modified by a conventional modification manner to obtain the new neural network model.
S208, for another advertiser, returning to execute steps S201-S207 based on the new neural network model until the loss function calculation result reaches a preset target value or the steps S201-S207 are executed for all the advertisers, and taking the new neural network model as a multi-label fusion model which is learned and used for outputting a multi-label fusion target audience concentration value according to input data.
In the step S208, for each of the already-cast advertisers, the corresponding second target crowd label information is used as a training sample, and the corresponding at least one already-cast advertisement point is used as the label value of the training sample, and the steps S201 to S207 are executed repeatedly, so that the neural network model can continuously self-learn based on different training samples and the corresponding label values, and a final multi-label fusion model which can be used for outputting a multi-label fusion target audience concentration value according to input data is obtained. In addition, the second target crowd label information corresponding to different thrown advertisers may be the same or different, and the corresponding at least one thrown elevator advertisement point may be the same or different.
Based on the possible designs one described in the foregoing steps S201 to S208, the multi-tag fusion model can continuously perform self-learning based on the point location tag specified by the advertiser according to the target crowd positioning result of the advertised, the tag condition value on the point location tag, and the elevator advertisement point location selection result, so as to continuously optimize the neural network parameters, improve the prediction accuracy of the multi-tag fusion target audience concentration value, and further ensure the correctness of the subsequent point location value sequencing.
The embodiment further provides a second possible design of how to obtain the preset coefficient based on the first possible design of the first aspect, that is, before calculating the estimated value CPR of the touch ability of the advertisement point of the elevator to be selected, the method further includes, but is not limited to, the following steps S301 to S304.
S301, acquiring corresponding in-car video data acquired on a plurality of continuous days aiming at each elevator advertisement point in a plurality of elevator advertisement points, wherein the elevator advertisement points are arranged in the area where the elevator advertisement point to be selected is located, the in-car video data are acquired by an in-car top camera installed in an elevator car to which the elevator advertisement point corresponds, and the in-car ground area is covered by the field of view of the in-car top camera.
In the step S301, the location area may be, but is not limited to, a street, a town, a district, a city, etc. Since the field of view of the overhead camera in the car covers the floor area in the car, the captured video frame images necessarily contain all of the elevator occupants in the elevator car. Further, the consecutive days may be, but are not limited to, 7 consecutive days within a week.
S302, according to the corresponding video data in the carriage, a head detection model is adopted to conduct head recognition processing on video frame images in the carriage, then the number of times of touch is counted according to head recognition results, and corresponding single advertisement day number of touch is obtained.
In the step S302, the human head detection model may be, but is not limited to, a model that is trained based on an existing target detection algorithm and has human head recognition capability. The target detection algorithm is an existing artificial intelligent recognition algorithm for recognizing and marking the object in the picture, specifically, but not limited to, a Faster R-CNN (Faster Regions with Convolutional Neural Networks features, he Kaiming, etc. a target detection algorithm is proposed in 2015, which obtains a plurality of first target detection algorithms in the ILSVRV and COCO contests in 2015, an SSD (Single Shot MultiBox Detector, a single-lens multi-box detector, a target detection algorithm proposed by Wei Liu on ECCV 2016, one of the main detection frameworks currently popular) or a YOLO (You only look once, recently developed to the V4 version, the basic principle of which is that an input image is divided into 7x7 grids, 2 frames are predicted for each grid, then a redundant window is removed according to a target window with relatively low threshold removal possibility, and finally a detection result is obtained by using a frame merging mode), and the like. Therefore, through a conventional sample training mode, the human head detection model for identifying whether a human head exists according to the input image can be obtained, so that after the video frame image is input, corresponding identification results, confidence prediction values and other information can be output. For example, the target detection algorithm preferably employs the YOLO V4 target detection algorithm. In addition, one person head inevitably corresponds to one elevator passenger, so that the advertisement touch times can be counted according to the person head identification result.
In the step S302, specifically, for each elevator advertisement point, according to the corresponding video data in the car, a head detection model is first used to perform head recognition processing on the video frame image in the car, and then touch-up number of times of statistics processing is performed according to the head recognition result, so as to obtain corresponding single advertisement day number of touch-up numbers of times, including but not limited to the following steps S3021 to S3025.
S3021, performing frame extraction processing on video data in a carriage every unit time to obtain multi-frame video frame images in the carriage, wherein the unit time is smaller than single playing time of advertisements.
In the step S3021, for example, a frame process may be performed every second to obtain the multi-frame intra-carriage video frame image.
S3022, aiming at each frame of intra-carriage video frame image in the multi-frame intra-carriage video frame images, adopting a human head detection model to carry out human head identification processing, and obtaining the number of people in the corresponding image.
S3023, regarding each advertisement played in the elevator advertisement point, taking the maximum number of people obtained through recognition as the corresponding number of touch people in the single playing time period according to at least one frame of video frame image in the carriage corresponding to the single playing time period.
In the step S3023, for example, if there are 15 intra-frame video frame images in a single playing period of a certain advertisement, the number of people identified as the largest in the 7 th intra-frame video frame image is: 7 persons can record that the touch time of the single playing period is 7 times.
S3024, for each day of the continuous multiple days, firstly accumulating and calculating the number of times of the played advertisements in the corresponding days in all playing time slots, and then dividing the number of times of the played advertisements in the corresponding days by the total number of the carousel advertisements in the corresponding days to obtain the total number of times of the played advertisements in the corresponding single day.
S3025, carrying out average value calculation on all the single-day single-advertisement total touch times to obtain single-advertisement-day average touch times corresponding to the elevator advertisement points.
S303, calculating corresponding coefficient values b according to the following formulas aiming at the elevator advertisement points:
in the formula, h represents the average number of floors of the elevator belonging to the corresponding elevator advertisement point in the operation floor zone, l represents the number of floors of the elevator belonging to the corresponding elevator advertisement point in the operation floor zone, r represents the occupancy rate of the elevator belonging to the corresponding elevator advertisement point in the operation floor zone, cpr represents the average daily contact time of a single advertisement corresponding to the elevator advertisement point, and f represents the average daily periodical rate of the corresponding elevator advertisement point in the continuous multiple days.
S304, fitting all the calculated coefficient values B to obtain a preset coefficient B of the advertisement point position of the elevator to be selected.
In the step S304, the fitting process is in a conventional manner. By means of the fitting process, the preset coefficient B obtained is typically a linear function related to the number of floors in a given city. For example, the present inventors determine the preset coefficient b=12-L/8 of a city through the foregoing steps S301 to S304, thereby obtaining the estimated touch capability value of the advertisement point of the elevator to be selected
Therefore, based on the possible designs one described in the foregoing steps S301 to S304, the preset coefficient in the touch capability estimation formula can be accurately determined, the accuracy of the subsequent touch capability estimation result is improved, and the accuracy of the final point value ranking is further ensured.
As shown in fig. 4, in a second aspect of the present embodiment, a virtual device for implementing the elevator advertisement point recommendation method according to the first aspect or any one of the first aspects may be provided, where the virtual device includes a tag information obtaining module, a target audience concentration estimating module, a point value scoring calculating module, and a point ordering module that are sequentially connected in a communication manner;
the tag information acquisition module is used for acquiring first target crowd tag information of an advertiser, wherein the first target crowd tag information comprises at least one point location tag designated by the advertiser according to a target crowd positioning result of an advertisement to be placed;
The target audience concentration estimation module is used for inputting the corresponding tag actual values on the tag of each point in the first target crowd tag information into a multi-tag fusion model which is based on a neural network and has been learned, and outputting the corresponding first multi-tag fusion target audience concentration value for each elevator advertisement point in a plurality of elevator advertisement points to be selected;
the point value scoring calculation module is used for calculating the product of the corresponding first multi-label fusion target audience concentration value and the corresponding touch capability estimated value aiming at each elevator advertisement point to be selected to obtain a corresponding first point value score;
and the point position ordering module is used for sequentially ordering the plurality of elevator advertisement points to be selected according to the high-to-low sequence of the first point position value scores to obtain a first point position list for recommending the elevator advertisement points to the advertiser.
The working process, working details and technical effects of the foregoing device provided in the second aspect of the present embodiment may refer to the first aspect or any one of the first aspects, which may be designed to be the method for recommending an advertisement point of an elevator, and will not be described herein.
As shown in fig. 5, a third aspect of the present embodiment provides a computer device for executing the elevator advertisement spot recommendation method according to the first aspect or any one of the first aspects, which includes a memory and a processor that are communicatively connected, where the memory is configured to store a computer program, and the processor is configured to read the computer program and execute the elevator advertisement spot recommendation method according to the first aspect or any one of the first aspects. By way of specific example, the Memory may include, but is not limited to, random-Access Memory (RAM), read-Only Memory (ROM), flash Memory (Flash Memory), first-in first-out Memory (First Input First Output, FIFO), and/or first-in last-out Memory (First Input Last Output, FILO), etc.; the processor may be, but is not limited to, a microprocessor of the type STM32F105 family. In addition, the computer device may include, but is not limited to, a power module, a display screen, and other necessary components.
The working process, working details and technical effects of the foregoing computer device provided in the third aspect of the present embodiment may refer to the first aspect or any one of the first aspects, which may be designed to be the elevator advertisement point recommendation method, and will not be described herein.
A fourth aspect of the present embodiment provides a computer readable storage medium storing instructions comprising the first aspect or any one of the first aspects of the possible designs of the elevator ad spot recommendation method, i.e. the computer readable storage medium has instructions stored thereon that, when run on a computer, perform the elevator ad spot recommendation method as described in the first aspect or any one of the first aspects of the possible designs. The computer readable storage medium refers to a carrier for storing data, and may include, but is not limited to, a floppy disk, an optical disk, a hard disk, a flash Memory, and/or a Memory Stick (Memory Stick), where the computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable devices.
The working process, working details and technical effects of the foregoing computer readable storage medium provided in the fourth aspect of the present embodiment may refer to the first aspect or any one of the first aspects, and may be referred to as designing the elevator advertisement point recommendation method, which is not described herein.
A fifth aspect of the present embodiments provides a computer program product comprising instructions which, when run on a computer, cause the computer to perform the elevator ad spot recommendation method according to the first aspect or any one of the possible designs of the first aspect. Wherein the computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus.
Finally, it should be noted that the invention is not limited to the alternative embodiments described above, but can be used by anyone in various other forms of products in the light of the present invention. The above detailed description should not be construed as limiting the scope of the invention, which is defined in the claims and the description may be used to interpret the claims.