Invention content
In view of this, this specification one or more embodiment provides a kind of Products Show method and apparatus, to improve numberProducts Show quality in the case of according to lacking.
Specifically, this specification one or more embodiment is achieved by the following technical solution：
In a first aspect, provide a kind of Products Show method, the method be used to determine whether by Products Show to be recommended toTarget user, the method includes：
The associated multi-field information of the target user is obtained, the multi-field information includes：The target user is in instituteState the purchase data of the product scope of product to be recommended and the purchase data of other product scopes；
According to the multi-field information, the user characteristics matrix of the target user, the user characteristics matrix packet are builtIt includes：According to multiple characteristic values of the multi-field information quantization；
For a product to be recommended, the user characteristics for the multiple users for buying the product to be recommended are obtainedMatrix, and the characteristic value in the user characteristics matrix based on the multiple user, it is corresponding to obtain the product to be recommendedProduct feature matrix；
The machine learning model for respectively training the user characteristics matrix and product feature Input matrix in advance, is usedFamily preference vector and product preference vector, the user preference vector are used to indicate preference of the target user in product purchase,The product preference vector is used to indicate to buy user's feature of the product to be recommended；
According to the user preference vector and product preference vector, obtain the product to be recommended and the target user itBetween selection assessed value, the selection assessed value is for indicating that the target user buys the probability of the product to be recommended；
When the selection assessed value is more than scheduled recommendation threshold value, it is determined that by the Products Show to be recommended to describedTarget user.
Second aspect, provides a kind of Products Show device, described device be used to determine whether by Products Show to be recommended toTarget user, described device include：
Data obtaining module, for obtaining the associated multi-field information of the target user, the multi-field information includes：The target user is in the purchase data of the product scope of the product to be recommended and the purchase data of other product scopes；
User's matrix builds module, for according to the multi-field information, building the user characteristics square of the target userBattle array, the user characteristics matrix include：According to multiple characteristic values of the multi-field information quantization；
Product matrix builds module, for for a product to be recommended, obtaining and buying the product to be recommendedThe user characteristics matrix of multiple users, and the characteristic value in the user characteristics matrix based on the multiple user, obtainTo the corresponding product feature matrix of the product to be recommended；
Model processing modules, the machine for respectively training the user characteristics matrix and product feature Input matrix in advanceDevice learning model obtains user preference vector and product preference vector, and the user preference vector is for indicating that target user existsPreference in product purchase, the product preference vector are used to indicate to buy user's feature of the product to be recommended；
Output processing module, for according to the user preference vector and product preference vector, obtaining the production to be recommendedSelection assessed value between product and the target user, the selection assessed value are waited for for indicating described in target user's purchaseThe probability of recommended products；
Recommend determining module, for when the selection assessed value is more than scheduled recommendation threshold value, it is determined that waited for describedRecommended products recommends the target user.
The third aspect provides a kind of Products Show equipment, and the equipment includes memory, processor, and is stored inOn reservoir and the computer instruction that can run on a processor, the processor realize following steps when executing instruction：
The associated multi-field information of the target user is obtained, the multi-field information includes：The target user is in instituteState the purchase data of the product scope of product to be recommended and the purchase data of other product scopes；
According to the multi-field information, the user characteristics matrix of the target user, the user characteristics matrix packet are builtIt includes：According to multiple characteristic values of the multi-field information quantization；
For a product to be recommended, the user characteristics for the multiple users for buying the product to be recommended are obtainedMatrix, and the characteristic value in the user characteristics matrix based on the multiple user, it is corresponding to obtain the product to be recommendedProduct feature matrix；
The machine learning model for respectively training the user characteristics matrix and product feature Input matrix in advance, is usedFamily preference vector and product preference vector, the user preference vector are used to indicate preference of the target user in product purchase,The product preference vector is used to indicate to buy user's feature of the product to be recommended；
According to the user preference vector and product preference vector, obtain the product to be recommended and the target user itBetween selection assessed value, the selection assessed value is for indicating that the target user buys the probability of the product to be recommended；
When the selection assessed value is more than scheduled recommendation threshold value, it is determined that by the Products Show to be recommended to describedTarget user.
The Products Show method and apparatus of this specification one or more embodiment, by the user's row for merging multiple fieldsFor data and essential information, and relevant preference profiles, side are bought using deep neural network intelligence perception user and productUser is helped to select suitable finance and money management product, it is sparse with cold start-up problem to be effectively relieved transaction data that the sector is faced,The accuracy for effectively increasing finance and money management personalization of product recommendation, more accurate recommendation service is provided for target user.
Specific implementation mode
In order to make those skilled in the art more fully understand the technical solution in this specification one or more embodiment,Below in conjunction with the attached drawing in this specification one or more embodiment, to the technology in this specification one or more embodimentScheme is clearly and completely described, it is clear that and described embodiment is only this specification a part of the embodiment, rather thanWhole embodiments.Based on this specification one or more embodiment, those of ordinary skill in the art are not making creativenessThe every other embodiment obtained under the premise of labour should all belong to the range of disclosure protection.
This specification one or more embodiment provides a kind of Products Show method when Sparse, and this method is retouchedIt states by taking the recommendation of finance and money management product as an example, but it is understood that, this method may be equally applicable for other and be opened with coldThe Products Show scene of dynamic feature.
Wherein, which has merged the user behavior data from different field, by product scope to be recommendedExcept other field behavioural information, alleviate the Sparse and cold start-up problem of product scope to be recommended.Because user existsThe buying behavior of other field, which can also react identity characteristic, environmental characteristic, life taste of user etc., to be helped to react userProduct buys the information of preference, and the recommendation for treating recommended products field also has good reference function.
In addition, the recommendation method also uses machine learning model, for example, by taking deep neural network as an example, the depth is utilizedThe output result of neural network model is spent to assist carrying out Products Show.Certainly, deep neural network model can first carry out mouldType training, and the use for the model progress Products Show completed using training.
The training of model：
First, deep neural network model of the training for finance and money management Products Show.
The actual acquired data that can be bought according to product, builds the objective matrix of model training.The actual acquired dataIn may include purchase data of the user to product, for example can be user and the actual purchase of finance and money management product is recorded, exampleSuch as, user A has purchased fund J1, and user B has purchased stock G1 and fund J1, and user C has purchased fund J2, etc..According to above-mentionedActual acquired data, objective matrix can be built, such as a kind of objective matrix of the following table 1 example, but not limited to this：
1 objective matrix of table
| Product 1 (fund J1) | Product 2 (stock G1) | Product 3 (fund J2) |
| User A | 1 | 0 | 0 |
| User B | 1 | 1 | 0 |
| User C | 0 | 0 | 1 |
In above-mentioned table 1, objective matrix may include purchase selective value of the user to product, and the purchase selective value is used forIndicate whether user buys product.Illustratively, purchase selective value may include " 1 " or " 0 ", when numerical value is 1, indicate to useFamily has purchased the product；When numerical value is 0, indicate that user does not buy the product.The objective matrix can be used as depth nerve netThe training objective of network model, when the deviation between the output result and the training objective of the model in training is smaller and smaller, andWhen deviation reaches predetermined threshold, just terminate the training of model, and the model that training terminates is directly used in follow-up finance and money management and is producedThe recommendation of product.
Then, in objective matrix each user (for example, user A, user B) and each product (for example, product 1,Product 2), the user characteristics matrix of each user can be built respectively, and builds the product feature matrix of each product respectively.AndBy the user characteristics matrix and product feature Input matrix of structure machine learning model to be trained, output model output matrix,The model output matrix includes each purchase selective value exported by the machine learning model.In the model output matrixWhen reaching predetermined threshold with the deviation of objective matrix, model training terminates.
As follows by the process of the detailed descriptive model training of Fig. 1, should describe how to build above-mentioned user in the processEigenmatrix, product feature matrix, and how by Input matrix model with the process of training pattern.
In step 100, the associated multi-field information of target user is obtained, the multi-field information includes：Target userIn the purchase data of the purchase data and other product scopes of the product scope of product to be recommended.
In this step, target user is the user of product to be recommended, such as, it is desirable to user's A recommended products, but still notKnow to the user A and which product recommended, needs recommendation method through this embodiment to determine the production to recommend to user AProduct, then the user A is properly termed as target user.
It should be noted that in model training, goal user can be the user in objective matrix, these useFamily has occurred that actual buying behavior in fact.And model after subsequent model training is in use, target userCan be the user that not yet certain products are bought of pending Products Show.
By taking finance and money management product as an example, product scope, that is, finance and money management product of product to be recommended, target user waits at thisThe purchase data of the product scope of recommended products, such as may include：User buys the transaction amount of some finance and money management product.And the purchase data of other product scopes can be with the purchase of right and wrong finance and money management product, for example, it may be purchase clothes, purchase electricityRice cooker etc..The purchase data of other product scopes can buy the purchasing price of the other field product, for example, purchaseThe clothes bought is 200 yuan, and the electric cooker of purchase is 350 yuan.Regardless of being the product scope or other products of product to be recommendedThe purchase data in field are all the data generated by target user to buy.
In addition, multi-field information is also not limited to the purchase data in above-mentioned different product field, can also include otherInformation.Several, including but not limited to following information of following example：
For example, the customer attribute information of the target user.The customer attribute information can be the gender of user, the age,Educational background etc..
For example, purchase data of the association user of the target user in the product scope of product to be recommended.Wherein, targetThe association user of user can have friend relation, relationship of transferring accounts etc. with target user.Can be mesh by taking friend relation as an exampleThe data for the purchase finance and money management product that the good friend of mark user occurred, for example, the good friend user a of user A bought some goldMelt finance product, and transaction amount is 20,000.
For example, the lend-borrow action data of target user.The lend-borrow action that the lend-borrow action data can be occurred with target user,The product of some category is borrowed or lent money, and the amount of money borrowed or lent money is how many.
In a step 102, according to the multi-field information, the user characteristics matrix of the target user, the use are builtFamily eigenmatrix includes：According to multiple characteristic values of the multi-field information quantization.
In this step, collected data in step 100 can be based on and quantified, characteristic value is converted into.
A kind of form of user characteristics matrix of following 2 example of table：
2 user characteristics matrix of table
As upper table 2 can carry out the coarseness processing of product first before the quantization for carrying out characteristic value.At coarsenessReason is that the data that will more be refined in data set are converted into generality, the higher data of synthesis degree.If for a product categoryPurchase data, the product quantity bought under the product category reaches coarseness treatment conditions, then by the product categoryUnder multiple products carry out coarseness processing.As an example it is assumed that the purchase data of other product scopes include purchase clothes,Multiple thinner categories such as electric cooker, also, target user just has purchased universal love and thinks DFB-B in electric cooker this categoryThe processing of the low capacities electric cookers such as 0.8L, oaks AR-Y0801, Lip river shellfish LBF-091BM is the mini electric cooker of 0-1L non-computers, beautifulA variety of electric cookers such as the domestic full-automatic intelligent electric cooker such as MB-WHS30C96, rice man pressure IH, Panasonic SR-AE101-K.ThatIf when user characteristics matrix is built, these products are divided it is very thin, for example, other product buying behaviors in table 2In, including many products such as product 1, product 2, product 3, such as above-mentioned universal love think DFB-B 0.8L, oaks AR-Y0801,Multiple products such as Lip river shellfish LBF-091BM, then prodigious calculating pressure will be caused.Therefore, coarseness processing can be by granularity waterEqual the granular level that thinner characteristic dimension is aggregated into a relative coarseness.
For example, universal love thinks the processing of the low capacities electric cookers such as DFB-B 0.8L, oaks AR-Y0801, Lip river shellfish LBF-091BMFor the mini electric cooker of 0-1L non-computers, beautiful MB-WHS30C96, rice man pressure IH, Panasonic SR-AE101-K etc. are domestic full-automaticIntelligent electric cooker processing is 3L-4L intelligent microcomputer electric cookers.And whether coarse grain is carried out to the purchase data of a product categoryDegree processing, can be arranged coarseness treatment conditions.For example, the condition can be the product quantity bought under the product categoryReach certain amount threshold, for example, the product quantity under the same product category has reached 6 or more.And in table 2Customer attribute information, social networks, the purchase of finance and money management product and the characteristic dimensions such as lend-borrow action, due to its characteristic dimensionLess, information content is high, can not have to carry out coarseness processing.
The following characteristic value quantization for illustrating how to carry out each dimension respectively, wherein it should be noted that following quantizationMethod is only example, is not limited thereto in actual implementation, can be executed according to other quantitative criterias：
1) the finance and money management product buying behavior of social networks user is established for target user and with target user：
For example, can be multiple areas by transaction amount classifying rationally according to the transaction amount of the finance and money management product of purchaseBetween, such as the " < in table 2;Multiple sections such as P1 ", " P1-P2 ", " P2-P3 ".If user buys the amount of money of the finance and money management productIn the section, then 1 is labeled as；Otherwise it is 0.
Wherein, the social networks column in table 2 with target user there is the user of incidence relation to buy finance and money management productPurchase data will first can establish the institute of social networks since the user of incidence relation may be multiple users with target userThere is the transaction amount of user to be averaged, is marked according to the amount of money of average value.For example, if average value is in section " P1-P2 ", then can be in the characteristic value label 1 in the corresponding section.
2) for the buying behavior of other products：
As described above, the purchase data of other products have carried out coarseness processing, are in same coarser particle size levelThere can be multiple products, and there can be relatively large difference in the price of these products.At this time can with price thisIndex buys frequency totality 0-1 standards by each attribute section of all product classifying rationallies under the category to table 2, and by userFor value after change as its attribute value, reaction target user buys the frequent degree of product in the price range under the category.
For example：Assuming that target user is in the buying behavior of other products, 3L-4L intelligent microcomputers electric cooker thisUnder one category, has purchased universal love and think at the low capacities electric cookers such as DFB-B 0.8L, oaks AR-Y0801, Lip river shellfish LBF-091BMReason is the mini electric cooker of 0-1L non-computers, and the households such as beautiful MB-WHS30C96, rice man pressure IH, Panasonic SR-AE101-K are complete certainlyIntelligent electric cooker is moved, i.e., has purchased multiple product under same category.It can so be looked into according to the respective purchasing price of these productsIt sees in " <The product quantity bought in the sections P1 ", and using the quantity as the characteristic value in the corresponding section.For example, described"<3 products are had purchased in the sections P1 ", then characteristic value is 3；1 product is had purchased in described section " P1-P2 ", then it is rightShould the characteristic value in section can be 1.
3) for the lend-borrow action of target user：
For example, the quantization of the lend-borrow action is similar with the quantization of finance and money management product, equally it is rationally to draw credit amountIt is divided into multiple sections, if the amount of money that user borrows or lends money the category product is in the section, is labeled as 1；Otherwise it is 0.
4) for the essential information of user：
For example, for numeric type variable, such as age can be divided according to method identical with transaction amount.ExampleProperty, 18 years old~25 years old correspondence, one quantized value, 26 years old~35 years old correspondence, one quantized value.
For example, for classification type variable, such as gender, educational background, then marked after can encoding Variable Factors.Illustratively,Undergraduate course educational background can correspond to a quantized value, and postgraduate's educational background can correspond to a quantized value.
At step 104, for multiple products, the user characteristics square for the multiple users for buying the product is obtainedBattle array, and the characteristic value in the user characteristics matrix based on the multiple user, obtain the product feature matrix of the product.
Product in this step is finance and money management product.This step can build product feature matrix, a product featureMatrix can correspond to a product, which can be each product in objective matrix.Wherein, the structure of product feature matrixIt can be based on user characteristics matrix.
For example, by taking a product as an example, that buys the finance and money management product has multiple users, each user constructsUser characteristics matrix shown in table 2.The corresponding multiple user characteristics matrixes of multiple users can be so based on, by characteristic valueIt is weighted average.
For example, by taking the age in essential information as an example, buying each user of the product, there are one the spies at corresponding ageThe characteristic value of multiple users can be weighted averagely, obtain an age corresponding comprehensive characteristics value by value indicative.
For another example, by taking the category 1 in other product buying behaviors in table 2 as an example, each user in multiple usersThere are one the characteristic values of the corresponding category 1, can be weighted the characteristic value of multiple users averagely, it is right to obtain a category 1The comprehensive characteristics value answered.
It can further be seen that each characteristic value in table 2 corresponds to different characteristic value positions, for example, the x1 in table 2 is corresponded toCharacteristic value position Shi [The corresponding section " P1-P2 " of row, arranges corresponding " 1 " ] of category;, and the corresponding characteristic value positions characteristic value x2 are[The corresponding section " P2-P3 " of row, arranges corresponding " 1 " ] of category;.It, can be special by the user of multiple users when building product feature matrixThe characteristic value that same characteristic value position is corresponded in sign matrix is weighted averagely, obtains corresponding to the feature in product feature matrixIt is worth the characteristic value of position.
That is, the characteristic value of multiple users, can be weighted averagely, finally obtaining can be anti-by each row in table 2The product feature matrix of user's global feature of the product should be gone out to buy.
Wherein, the setting of weight when characteristic value weighted average can be determined according to practical business situation.If for example, recognizingIt is more important when reacting user's global feature for the characteristic value of some user, just its weight is arranged some higher.
In step 106, attribute interactive operation is carried out to user characteristics matrix and product feature matrix respectively.
The attribute interactive operation of user characteristics matrix and product feature matrix can be carried out in this step.Attribute interactive operationIt is to establish correlativity between the attribute that will be not directly relevant in matrix, it is first that the eigenmatrix of structure is random as unit of attribute columnSequence generates multiple new eigenmatrixes, then multiple new eigenmatrix splicings are generated the eigenmatrix after attribute interaction.It needsIt is noted that the attribute interactive operation can be that an optional operation can be more effective after executing attribute interactive operationIt was found that the potential association between different characteristic, thus also can be more accurate when later use machine learning model perceives user preferenceReally.
The principle of the attribute interactive operation of eigenmatrix may refer to shown in Fig. 2：
As described in Figure 2, each feature such as feature 1 therein, feature 2, feature 3 corresponds to different characteristic series.With userFor eigenmatrix, feature 1 can be " product 1 in the buying behavior of finance and money management product " in table 1, and feature 15 can be" the debt-credit category 1 in lend-borrow action " in table 1, i.e., different features corresponds to different lines.According to Fig.2, be equivalent to byRandom movement has been carried out between different lines in table 1, is that unit progress is randomly ordered with row, is then spliced.
In step 108, respectively by the user characteristics matrix and product feature Input matrix machine learning model after interaction,Obtain user preference vector and product preference vector.
In this step, deep neural network includes two parallel neural networks, one of them is user behavior preferenceIntellisense device, the other is buying the Intellisense device of user's general characteristic preference of the product, as shown in Figure 3.By attributeInteraction and input of the spliced eigenmatrix as parallel neural network, for example, the user characteristics matrix after attribute interaction is defeatedEnter a neural network, product feature Input matrix another neural network after attribute interaction.
After the convolutional layer of neural network, pond layer and full linked operation, it is inclined that neural network can respectively obtain userGood vector sum product preference vector.Wherein, the user preference vector can be used to indicate that preference of the user in product purchase,It is equivalent to and indicates a user likes which type of product bought.And the product preference vector can be used to indicate that purchase productWhat user is user's feature of the corresponding product of eigenmatrix have the characteristics that is, indicating for a productIt is more likely to buy the product.
In step 110, the user preference vector and product preference vector exported according to model obtains model output squareBattle array, the model output matrix includes each purchase selective value exported by machine learning model.
For example, by the corresponding user characteristics Input matrix neural network model of a user, user preference vector is obtained；It willThe corresponding product feature Input matrix neural network model of one product, obtains product preference vector.It can be inclined according to the userGood vector sum product preference vector obtains a purchase selective value.For example, can be inclined by above-mentioned user preference vector and productGood vector seeks inner product, obtains purchase selective value, and the selection value indicates that above-mentioned user buys the probability of the product.
One user characteristics matrix can be built for each user in objective matrix, it can for each productCorresponding product feature matrix is built respectively.According to above-mentioned method, product of one of user couple can be obtainedBuy selective value.These purchase selective values may be constructed model output matrix, i.e., each purchase that the model output matrix includesBuy the numerical value that selective value is neural network model output.
And the user that objective matrix includes is to the purchase selective value of product, is obtained according to actual acquired data, is to useThe buying behavior that family actually occurs, objective matrix are the mutual selection matrixs of the user really occurred and product.It can be by targetTraining objective of the matrix as neural network model, with continuing to optimize for model, the output result and reality of neural network modelThe generation numerical value on border will be closer.
In step 112, when the deviation of the model output matrix and objective matrix reaches predetermined threshold, model trainingTerminate.
For example, when can reach predetermined threshold with the deviation of setting model output matrix and objective matrix, terminate the instruction of modelPractice.It can be that deviation is less than or equal to scheduled threshold value that the deviation, which reaches predetermined threshold, i.e., deviation foot between the twoIt is enough small.Wherein, the measurement of the deviation of model output matrix and objective matrix can be there are many method, for example, weigh can be with for deviationUse root-mean-square error RMSE (Root Mean Square Error) or mean absolute error MAE (Mean AbsoluteDeviation).It is mutual between prediction user and product according to trained neural network model after model trainingWhen select probability, it will being approached with actual conditions for prediction has the prodigious prediction probability of success.
Use to the model that training terminates：
Assuming that terminating two neural metwork trainings arranged side by side, trained as follows with an example to illustrate how to useGood model judges to recommend which kind of product that will have higher success rate to user.
For example, it is assumed that include to user's Y recommendation finance and money management products, product to be recommended currently：Products C 1, productMultiple products such as C2, products C 3, then to recommend which finance and money management product that can have higher success rate to user Y, it can be according to thisThe recommendation method of example executes.
The user characteristics matrix of user Y can be first built, and builds multiple productions such as products C 1, products C 2, products C 3 respectivelyThe product feature matrix of product.Then, the user characteristics matrix of user Y and the product feature matrix of products C 1 are inputted parallel respectivelyNeural network, obtain user preference vector and product preference vector.And user Y is obtained to products C 1 based on the two vectorsAssessed value, the selection assessed value is selected to be used to indicate the probability that target user buys assessment product.The selection assessed value with it is upperThe calculation for stating the purchase selective value mentioned is identical, and it is to distinguish only to use two titles, and purchase selective value is in mouldThe numerical value calculated when type training, select assessed value be the numerical value calculated in the complete use of model training, for as whether toThe foundation of user's recommended products.
Multiple products such as above-mentioned products C 1 to be recommended, products C 2, products C 3 are properly termed as assessment product, that is, assess theseWhether product will recommend user Y.It can each divide between the product feature matrix of product and the user characteristics matrix of user YA selection assessed value is not obtained.A recommendation threshold value can be set, is more than scheduled recommendation threshold value in the selection assessed valueWhen, it is determined that give the assessment Products Show to the target user.As an example it is assumed that the selection of products C 1 and user Y is commentedValuation is 0.6, and the selection assessed value of products C 2 and user Y are 0.8, and the selection assessed value of products C 3 and user Y are 0.2, and falseIf it is 0.55 to recommend threshold value, then can really directional user Y recommended products C1 and products C 2, not recommended products C3.
The personalized recommendation method of the finance and money management product of this example, by merge multiple fields user behavior data withEssential information, and relevant preference profiles are bought with product using deep neural network intelligence perception user, help user to chooseSuitable finance and money management product is selected, it is sparse with cold start-up problem to be effectively relieved transaction data that the sector is faced, effectively improvesThe accuracy that finance and money management personalization of product is recommended, for target user provides more accurate recommendation service, becomes promotion and sellsThe effective measure of good interaction between platform and user.
In order to realize that the above method, at least one embodiment of this specification additionally provide a kind of Products Show device.Such as Fig. 4Shown, which can be used to determine whether that by Products Show to be recommended, to target user, which may include：Acquisition of informationModule 41, user's matrix structure module 42, product matrix structure module 43, model processing modules 44,45 and of output processing moduleRecommend determining module 46.
Data obtaining module 41, for obtaining the associated multi-field information of the target user, the multi-field packetIt includes：The target user is in the purchase data of the product scope of the product to be recommended and the purchase data of other product scopes；
User's matrix builds module 42, for according to the multi-field information, building the user characteristics of the target userMatrix, the user characteristics matrix include：According to multiple characteristic values of the multi-field information quantization；
Product matrix builds module 43, for for a product to be recommended, obtaining and buying the product to be recommendedMultiple users the user characteristics matrix, and the characteristic value in the user characteristics matrix based on the multiple user,Obtain the corresponding product feature matrix of the product to be recommended；
Model processing modules 44, for respectively by the training in advance of the user characteristics matrix and product feature Input matrixMachine learning model obtains user preference vector and product preference vector, and the user preference vector is for indicating target userPreference in product purchase, the product preference vector are used to indicate to buy user's feature of the product to be recommended；
Output processing module 45, for according to the user preference vector and product preference vector, obtaining described to be recommendedSelection assessed value between product and the target user, the selection assessed value is for indicating described in target user's purchaseThe probability of product to be recommended；
Recommend determining module 46, for when the selection assessed value is more than scheduled recommendation threshold value, it is determined that will be describedProducts Show to be recommended gives the target user.
In one example, user's matrix builds module 42, is additionally operable to：If for the purchase data of a product category,The product quantity bought under the product category reaches coarseness treatment conditions, then by multiple products under the product categoryCarry out coarseness processing.
In one example, product matrix builds module 43, is specifically used for the user characteristics matrix to the multiple userThe characteristic value of the middle same characteristic value position of correspondence is weighted averagely, obtains corresponding to the feature in the product feature matrixIt is worth the characteristic value of position.
In one example, model processing modules 44 are additionally operable to respectively by the user characteristics matrix and product featureBefore the machine learning model that Input matrix is trained in advance, the user characteristics matrix and product feature matrix are belonged to respectivelySexual intercourse interoperates；By the user characteristics matrix and product feature matrix after interaction, the machine learning model is inputted.
The device or module that above-described embodiment illustrates can specifically realize by computer chip or entity, or by havingThe product of certain function is realized.A kind of typically to realize that equipment is computer, the concrete form of computer can be personal meterCalculation machine, laptop computer, cellular phone, camera phone, smart phone, personal digital assistant, media player, navigation are setIt is arbitrary several in standby, E-mail receiver/send equipment, game console, tablet computer, wearable device or these equipmentThe combination of equipment.
For convenience of description, it is divided into various modules when description apparatus above with function to describe respectively.Certainly, implementing thisThe function of each module is realized can in the same or multiple software and or hardware when specification one or more embodiment.
Each step in above-mentioned flow as shown in the figure, execution sequence are not limited to the sequence in flow chart.In addition, eachThe description of a step can be implemented as software, hardware or its form combined, for example, those skilled in the art can be by itIt is embodied as the form of software code, can is the computer executable instructions that can realize the corresponding logic function of the step.When it is realized in the form of software, the executable instruction can store in memory, and by the processor in equipmentIt executes.
For example, corresponding to the above method, this specification one or more embodiment provides a kind of Products Show equipment simultaneously,The equipment may include processor, memory and storage on a memory and the computer instruction that can run on a processor,The processor is by executing described instruction, for realizing following steps：
The associated multi-field information of the target user is obtained, the multi-field information includes：The target user is in instituteState the purchase data of the product scope of product to be recommended and the purchase data of other product scopes；
According to the multi-field information, the user characteristics matrix of the target user, the user characteristics matrix packet are builtIt includes：According to multiple characteristic values of the multi-field information quantization；
For a product to be recommended, the user characteristics for the multiple users for buying the product to be recommended are obtainedMatrix, and the characteristic value in the user characteristics matrix based on the multiple user, it is corresponding to obtain the product to be recommendedProduct feature matrix；
The machine learning model for respectively training the user characteristics matrix and product feature Input matrix in advance, is usedFamily preference vector and product preference vector, the user preference vector are used to indicate preference of the target user in product purchase,The product preference vector is used to indicate to buy user's feature of the product to be recommended；
According to the user preference vector and product preference vector, obtain the product to be recommended and the target user itBetween selection assessed value, the selection assessed value is for indicating that the target user buys the probability of the product to be recommended；
When the selection assessed value is more than scheduled recommendation threshold value, it is determined that by the Products Show to be recommended to describedTarget user.
It should also be noted that, the terms "include", "comprise" or its any other variant are intended to nonexcludabilityIncluding so that process, method, commodity or equipment including a series of elements include not only those elements, but also wrapInclude other elements that are not explicitly listed, or further include for this process, method, commodity or equipment intrinsic wantElement.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that wanted including describedThere is also other identical elements in the process of element, method, commodity or equipment.
It will be understood by those skilled in the art that this specification one or more embodiment can be provided as method, system or calculatingMachine program product.Therefore, this specification one or more embodiment can be used complete hardware embodiment, complete software embodiment orThe form of embodiment combining software and hardware aspects.Moreover, this specification one or more embodiment can be used at one orIt is multiple wherein include computer usable program code computer-usable storage medium (include but not limited to magnetic disk storage,CD-ROM, optical memory etc.) on the form of computer program product implemented.
This specification one or more embodiment can computer executable instructions it is general onDescribed in hereafter, such as program module.Usually, program module includes executing particular task or realization particular abstract data typeRoutine, program, object, component, data structure etc..Can also put into practice in a distributed computing environment this specification one orMultiple embodiments, in these distributed computing environments, by being executed by the connected remote processing devices of communication networkTask.In a distributed computing environment, the local and remote computer that program module can be located at including storage device is depositedIn storage media.
Each embodiment in this specification is described in a progressive manner, identical similar portion between each embodimentPoint just to refer each other, and each embodiment focuses on the differences from other embodiments.At dataFor managing apparatus embodiments, since it is substantially similar to the method embodiment, so description is fairly simple, related place is referring to sideThe part of method embodiment illustrates.
It is above-mentioned that this specification specific embodiment is described.Other embodiments are in the scope of the appended claimsIt is interior.In some cases, the action recorded in detail in the claims or step can be come according to different from the sequence in embodimentIt executes and desired result still may be implemented.In addition, the process described in the accompanying drawings not necessarily require show it is specific suitableSequence or consecutive order could realize desired result.In some embodiments, multitasking and parallel processing be also canWith or it may be advantageous.
The foregoing is merely the preferred embodiments of this specification one or more embodiment, not limiting this theoryBright book one or more embodiment, all within the spirit and principle of this specification one or more embodiment, that is done is anyModification, equivalent replacement, improvement etc. should be included within the scope of the protection of this specification one or more embodiment.