CN109903103A

CN109903103A - A kind of method and apparatus for recommending article

Info

Publication number: CN109903103A
Application number: CN201711283557.0A
Authority: CN
Inventors: 唐睿明; 何秀强; 钮敏哲; 张伟楠; 俞勇
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2017-12-07
Filing date: 2017-12-07
Publication date: 2019-06-18
Anticipated expiration: 2037-12-07
Also published as: CN109903103B; WO2019109724A1

Abstract

The embodiment of the invention discloses a kind of method and apparatus for recommending article, belong to field of computer technology.The described method includes: obtaining the attribute data of target user and the attribute data of multiple candidate items；The attribute data of the attribute data of target user and multiple candidate items is handled, target data set is generated, target data set includes the mark of each candidate item and the second interaction node list of corresponding target in the mark and the list of the first interaction node of corresponding target, multiple candidate items of target user；Target data set is inputted into scoring model, obtains marking of the target user to multiple candidate items are stated, wherein scoring model is obtained according to the attribute data of multiple users, the attribute data of multiple articles and marking data training；Marking according to target user to multiple candidate items determines that target recommends article.Using the present invention, the efficiency that user selects article can be improved.

Description

A kind of method and apparatus for recommending article

Technical field

This application involves field of computer technology, in particular to a kind of method and apparatus for recommending article.

Background technique

With the development of computer technology and Internet technology, the terminals such as mobile phone, computer are widely used, phase The type for the application program in terminal answered is more and more, function is more and more abundant.People can be by the purchase installed in terminal Object application program is done shopping, and can also watch film etc. by the video playing application program installed in terminal.

People can choose article to be dealt with (commodity or film etc.) before operating to article first.Specifically , user can send item lists acquisition request to server by operation, triggering terminal, and server receives item lists After acquisition request, the item lists being made of each article stored in server can be sent to terminal.Terminal receives article After list, it can be shown, user can browse each article in item lists one by one, determine the object finally liked Product.

During realizing the application, the inventor finds that the existing technology has at least the following problems:

Based on above-mentioned processing mode, when user wants selection article, need to select in the item lists that server is sent, Number of articles often in item lists is relatively more, thus, the efficiency for causing user to select article is lower.

Summary of the invention

In order to solve the problems, such as that user present in the relevant technologies selects the efficiency of article lower, the embodiment of the present invention is provided A kind of method and apparatus for recommending article.The technical solution is as follows:

In a first aspect, provide it is a kind of recommend article method, this method comprises: obtain target user attribute data and The attribute data of the attribute data of multiple candidate items, target user includes the mark of target user, the category of each candidate item Property data include the mark of corresponding candidate item；The attribute data of the attribute data of target user and multiple candidate items is carried out Processing, generate target data set, target data set include target user mark and the list of the first interaction node of corresponding target, The mark of each candidate item and the second interaction node list of corresponding target, the first interaction node of target in multiple candidate items List is used to indicate the interactive information of target user and other users or article, and target the second interaction node list is waited for indicating Select article and other articles or the interactive information of user；Target data set is inputted into scoring model, obtains target user to multiple The marking of candidate item, wherein scoring model is according to the attribute data of multiple users, the attribute data of multiple articles and marking Data training obtains, and multiple users include target user, and the attribute data of each user includes corresponding use in multiple users The mark at family, multiple articles include multiple candidate items, and the attribute data of each article includes corresponding article in multiple articles Mark, marking data include marking of each user to articles one or more in multiple articles in multiple users；According to mesh Marking of the user to multiple candidate items is marked, determines that target recommends article.

Scheme shown in the embodiment of the present invention, server can have the function of recommending article.Specifically, server can be with The attribute data for obtaining multiple candidate items in the attribute data and Candidate Set of target user in turn can be to target user Attribute data and the attribute datas of multiple candidate items handled, obtain the i.e. corresponding target of user comprising target user The mark of each candidate item and the second interaction node of corresponding target column in first interaction node list, multiple candidate items Table, wherein target the first interaction node list can be used to indicate that the interactive information of target user and other users or article, i.e., It may include the mark of other users or article that target user's history interacted, target in target the first interaction node list Two interaction node lists can be used to indicate that candidate item and other articles or the interactive information of user, the i.e. interaction of target second section It may include the mark of other articles or user that candidate item history interacted in point list.

Scoring model can be previously stored in server, wherein scoring model can be server according to multiple users Attribute data, multiple articles attribute data and marking data training obtain, multiple users include target user, multiple Article includes multiple candidate items.Server can predict target user to each time in multiple candidate items by scoring model Select the marking of article, specifically, server generate comprising target user mark and the list of the first interaction node of corresponding target, It, can in multiple candidate items after the mark of each candidate item and the target data set of corresponding target the second interaction node list To be entered into scoring model, marking of the target user to multiple candidate items is obtained, in turn, server can be based on mesh It marks marking of the user to multiple candidate items and determines the target recommendation to be recommended to target user in multiple candidate items Product.In this way, target user can recommend to choose oneself desired article in article in the target that server is recommended, without taking It is selected in all items stored in business device, it is thus possible to improve the efficiency that user selects article.In addition, server is being predicted When target user is to the marking of each candidate item, interactive information (the i.e. mesh of target user and other users or article is utilized Mark the first interaction node list) and each candidate item with the interactive information of other articles or user (i.e. target second interacts section Point list), it is thus possible to the accuracy of the marking improved.

In one possible implementation, the attribute data of target user further includes one of following data or more Kind: gender, height, weight, age, occupation, income, hobby, education landscape, the attribute data of each candidate item further include with One of lower data are a variety of: brand, color, size, price, comment, taste, shelf-life, icon.

In one possible implementation, by the attribute data of the attribute data of target user and multiple candidate items into Row processing, generates target data set, comprising: according to the mark of target user, each user in pre-recorded multiple users Corresponding the first interaction node of the target list of mark in, determine corresponding the first interaction node of the target list of target user, and According to the mark of each candidate item, the corresponding target of mark of each candidate item in pre-recorded multiple candidate items In second interaction node list, corresponding the second interaction node of the target list of each candidate item is determined；According to target user's Mark, corresponding the first interaction node of the target list of target user, the mark of each candidate item and each candidate item pair Target the second interaction node list answered generates target data set.

Scheme shown in the embodiment of the present invention can be previously stored with the mark of each user in multiple users in server The corresponding interaction of target second of the mark of each candidate item in corresponding the first interaction node of target list, multiple candidate items Node listing, wherein server can record corresponding the first interaction node of the target column of mark of each user in the form of a table Corresponding the second interaction node list of target of the mark of table and each candidate item, can also be recorded each in the form of bigraph (bipartite graph) The corresponding interaction of target second section of mark for identifying corresponding target the first interaction node list and each candidate item of user Point list.It, can be pre-recorded multiple after server gets the mark of target user and the mark of each candidate item In user in corresponding the first interaction node of the target list of the mark of each user, determine that the corresponding target first of target user is handed over Mutual node listing, and the corresponding target second of the mark of each candidate item can be handed in pre-recorded multiple candidate items In mutual node listing, corresponding the second interaction node of the target list of each candidate item is determined.Determine that target user is corresponding After the list of the first interaction node of target, corresponding the second interaction node of the target list of each candidate item, server be can be generated Mark comprising target user and the mark and corresponding target of the list of the first interaction node of corresponding target, each candidate item The target data set of second interaction node list, wherein it may include target user's that target data, which concentrates each target data, Mark and the list of the first interaction node of corresponding target, the mark of candidate item j and the second interaction node list of corresponding target, Candidate item j is any candidate item in multiple candidate items.

In one possible implementation, scoring model includes feature learning model, feedback learning model and nerve net Network model；

Wherein, target data set is inputted into scoring model, obtains marking of the target user to multiple candidate items, comprising: The mark for the target user that target data is concentrated and the mark input feature vector learning model of candidate item j, obtain target user Corresponding feature vector and the corresponding feature vector of candidate item j, and the corresponding target of target user that target data is concentrated First interaction node list and corresponding the second interaction node of the target list of candidate item j, input feedback learning model obtain mesh Mark the corresponding implicit feedback of user and the corresponding implicit feedback of candidate item j, wherein article j is appointing in multiple candidate items One candidate item；The corresponding feature vector of target user, the corresponding feature vector of candidate item j, target user is corresponding hidden Formula feeds back implicit feedback corresponding with candidate item j, inputs neural network model, obtains target user and beat candidate item j Point.

Wherein, the corresponding feature vector of target user can be the feature (or characteristic) for characterizing the user itself to Amount.The corresponding feature vector of candidate item j can be the vector of the feature (or characteristic) for characterizing candidate item j itself.

Scheme shown in the embodiment of the present invention, scoring model may include feature learning model, feedback learning model and mind Through network model, wherein feature learning model can be for learning objective user and the corresponding feature of each candidate item to Amount, the model parameter of feature learning model may include user characteristics matrix and article characteristics matrix, wherein user characteristics matrix It is made of that (i.e. every row vector of user characteristics matrix is corresponding user respectively the feature vector of each user in multiple users Feature vector, the line number of user characteristics matrix are the quantity of multiple users), article characteristics matrix is by each in multiple articles (i.e. every row vector of article characteristics matrix is the feature vector of corresponding article, article characteristics respectively to the feature vector composition of article The line number of matrix is the quantity of multiple articles).After obtaining target data set, server can be by the target of target data concentration The mark of user and the mark input feature vector learning model of candidate item j, obtain the corresponding feature vector of target user and candidate The corresponding feature vector of article j.Specifically, server is by the mark input feature vector of the mark of target user and candidate item j After practising model, according to the mark of the mark of target user and candidate item j, through feature learning model in user characteristics matrix The corresponding feature vector of target user is extracted, the corresponding feature vector of candidate item j is extracted in article characteristics matrix, obtains mesh Mark the corresponding feature vector of user and the corresponding feature vector of candidate item j.

Feedback learning model can be for learning objective user and the corresponding implicit feedback of each candidate item, and feedback is learned The model parameter for practising model may include user feedback matrix (can be indicated with Y) and article feedback matrix (can be indicated with X), Wherein, user feedback matrix can be by feedback vector form (every row vector in user feedback matrix represents a node Corresponding feedback vector), article feedback matrix can be (the every row vector generation in article feedback matrix being made of feedback vector The corresponding feedback vector of one node of table).Determine corresponding the first interaction node of the target column of target user's (can be indicated with k) Table (can use R_kIndicate) afterwards, corresponding the second interaction node of the target list of candidate item j (R can be used_jIndicate) after, it can incite somebody to action Its input feedback learning model obtains the corresponding implicit feedback of target user k and the corresponding implicit feedback of candidate item j.Specifically , server can pass through feedback for after corresponding the first interaction node of the target list input feedback learning model of target user k Learning model extracts the corresponding multiple feedback vectors of target the first interaction node list in user feedback matrix and (wherein, feeds back The quantity of vector is the quantity for the node for including in target the first interaction node list), obtain the corresponding feedback of target user k Vector.After getting the corresponding feedback vector of target user k, multiple feedback vectors can be added, it is corresponding to obtain target user k Implicit feedback.The specific processing for obtaining the corresponding implicit feedback of candidate item j can be such that server by j pairs of candidate item It, can be by feedback learning model in article feedback matrix after target the second interaction node list input feedback learning model answered (wherein, the quantity of feedback vector is target second to the corresponding multiple feedback vectors of middle extraction target the second interaction node list The quantity for the node for including in interaction node list), obtain the corresponding feedback vector of candidate item j.Get j pairs of candidate item After the feedback vector answered, multiple feedback vectors can be added, obtain the corresponding implicit feedback of candidate item j.

It is corresponding hidden to obtain the corresponding feature vector of target user k, the corresponding feature vector of candidate item j, target user k After formula feeds back implicit feedback corresponding with candidate item j, server can be inputted neural network model, obtain target user Marking to candidate item j.

In one possible implementation, the list of the first interaction node of target includes multistage the first interaction node of target column Table, corresponding the second interaction node of the target list of each candidate item includes multistage the second interaction node of target list, multistage mesh That marks that odd-order target the first interaction node list in the first interaction node list is used to indicate target user and article interacts letter It ceases, the first interaction node list of even-order target is for indicating target user and other in multistage the first interaction node of target list The interactive information of user, the second interaction node list of odd-order target is waited for indicating in multistage the second interaction node of target list Article and the interactive information of user are selected, the second interaction node list of even-order target is used in multistage the second interaction node of target list In the interactive information for indicating candidate item and other articles；The corresponding interaction of target first of the target user that target data is concentrated Node listing and corresponding the second interaction node of the target list of candidate item j, input feedback learning model obtain target user couple The implicit feedback and the corresponding implicit feedback of candidate item j answered, comprising: the target user for concentrating target data is corresponding multistage Target the first interaction node list and corresponding multistage the second interaction node of the target list of candidate item j, input feedback learn mould Type obtains the corresponding implicit feedback of target user and the corresponding implicit feedback of candidate item j.

Scheme shown in the embodiment of the present invention, server, can be with when predicting target user to the marking of candidate item j Utilize corresponding multistage the first interaction node of the target list of target user, the corresponding multistage interaction of target second section of candidate item j Point list, wherein multistage the first interaction node of target list can be the list of the first interaction node of single order target, second order mesh respectively Mark the first interaction node list ..., target the first interaction node list of A rank, multistage user feedback matrix may include single order use Family feedback matrix, second order user feedback matrix ..., A rank user feedback matrix, A is default value (such as A be 3), and A is default The maximum step number that can be reached in user-article bigraph (bipartite graph) of target user, multistage the second interaction node of target list can divide Be not the list of the second interaction node of single order target, the list of the second interaction node of second order target ..., B rank the second interaction node of target List, multistage article feedback matrix may include single order article feedback matrix, second order article feedback matrix ..., B rank article feedback Matrix, B are default value, and B is the maximum step number that preset candidate item j can be reached in user-article bigraph (bipartite graph), wherein A It may be the same or different with B.Single order user feedback matrix, can use Y¹, every row vector in single order user feedback matrix Can be vector of the corresponding article as node in the first interaction node list of single order target when indicates, second order user feedback Matrix can use Y²It indicates, every row vector in second order user feedback matrix can be corresponding user as second order target the Vector when node in one interaction node list indicates, and so on.Single order article feedback matrix, can use X¹, single order object When every row vector in product feedback matrix can be corresponding user as node in the second interaction node list of single order target Vector indicate that second order article feedback matrix can use X²It indicates, every row vector in second order article feedback matrix can be pair The vector when article answered is as node in second order target the second interaction node list indicates, and so on.

For such situation, the corresponding multistage interaction of target first of the target user that server can concentrate target data Node listing and corresponding multistage the second interaction node of the target list of candidate item j, are input to feedback learning model, obtain target The corresponding implicit feedback of user and the corresponding implicit feedback of candidate item j.Specifically, being arranged for every the first interaction node of rank target Table(a=1,2 ..., A), server can be by feedback learning models, in rank user feedback matrix Y^aIn, extract target First interaction node listCorresponding feedback vector.Server it is corresponding can to select target user in the manner described above All feedback vectors selected can be added, obtain in turn by the corresponding feedback vector of each the first interaction node of rank target list To the corresponding implicit feedback of target user.Every rank target corresponding for candidate item j the second interaction node list(b=1, 2 ..., B), server can be by feedback learning model, in article feedback matrix X^bIn, choose the second interaction node of target column TableCorresponding feedback vector obtains corresponding the second interaction node of the target list of candidate item jCorresponding feedback vector. Server it is corresponding anti-can to select corresponding each the second interaction node of the rank target list of candidate item j in the manner described above All feedback vectors selected can be added in turn, obtain the corresponding implicit feedback of candidate item j by feedback vector.In this way, When predicting target user to the marking of each candidate item, the corresponding each rank history mutual information of target user and every is utilized Each rank history mutual information of one candidate item, it is thus possible to improve the marking of the target user that predicts to candidate item Accuracy.

In one possible implementation, the model parameter of feedback learning model includes: each user in multiple users Feedback vector weight, the weight of the feedback vector of each article in multiple articles；The target user that target data is concentrated Corresponding target the first interaction node list and corresponding the second interaction node of the target list of candidate item j, input feedback study Model obtains the corresponding implicit feedback of target user and the corresponding implicit feedback of candidate item j, comprising: concentrate target data Target user mark and the list of the first interaction node of corresponding target, candidate item j mark and corresponding target second Interaction node list, input feedback learning model, obtains the corresponding implicit feedback of target user and candidate item j is corresponding implicit Feedback.

Scheme shown in the embodiment of the present invention, the model parameter for feedback learning model further include each in multiple users In the weight of the feedback vector of user and multiple articles the case where the weight of the feedback vector of each article, server can be by mesh Mark the mark of the target user in data set and the mark and correspondence of the list of the first interaction node of corresponding target, candidate item j Target the second interaction node list, input feedback learning model obtains the corresponding implicit feedback of target user and candidate item j Corresponding implicit feedback.Specifically, server can be according to above-mentioned determining target user and the corresponding feedback vector of candidate item j Method, the corresponding feedback vector of target user and the corresponding feedback vector of candidate item j are determined by learning model.Then, Server can pass through the target user's in feedback learning model according to the mark of target user and the mark of candidate item j The weight of feedback vector (can use Φ_ktIndicate) the corresponding feedback vector of target user k is weighted and is handled, obtain target The corresponding implicit feedback of user, and Ω (can be used by the weight of the feedback vector of the candidate item j in feedback learning model_vj Indicate) the corresponding feedback vector of candidate item j is weighted and is handled, obtain the corresponding implicit feedback of candidate item j.This Sample introduces the weight of the feedback vector of target user and each when predicting target user to the marking of each candidate item The weight of the feedback vector of candidate item, it is thus possible to improve the accurate of the marking of the target user that predicts to candidate item Property.

In one possible implementation, the marking according to target user to multiple candidate items determines that target is recommended Article, comprising: the marking according to target user to multiple candidate items determines that corresponding marking meets the mesh of default recommendation condition Mark recommends article.

Scheme shown in the embodiment of the present invention can be previously stored with default recommendation condition in server, and server obtains After target user is to the marking of multiple candidate items, it can choose corresponding marking in multiple candidate items and meet default push away The target for recommending condition recommends article.

In one possible implementation, the marking according to target user to multiple candidate items determines corresponding beat The target for meeting default recommendation condition is divided to recommend article, comprising: the marking according to target user to multiple candidate items, determining pair The maximum preset number target of the marking answered recommends article；Alternatively, the marking according to target user to multiple candidate items, really The target that fixed corresponding marking is greater than preset fraction threshold value recommends article.

Scheme shown in the embodiment of the present invention can after server determines target user to the marking of multiple candidate items With the sequence descending according to corresponding marking, multiple candidate items are ranked up, in turn, by the forward present count that sorts Mesh candidate item is determined as target and recommends article.Alternatively, preset fraction threshold value can be previously stored in server.Service After device determines target user to the marking of multiple candidate items, it is big can to choose corresponding marking in multiple candidate items In the candidate item of preset fraction threshold value, in turn, determining candidate item can be determined as to target and recommend article.

In one possible implementation, training obtains scoring model by the following method: obtaining the category of multiple users Property data, the attribute data of multiple articles and the marking data；To the attribute of the attribute data of multiple users, multiple articles Data and marking data are handled, and obtain training dataset, and training dataset includes the mark of each user and corresponding First interaction node list, the mark of each article and corresponding second interaction node list, each user are in multiple articles The marking of one or more articles, the first interaction node list are used to indicate the interactive information of user and other users or article, Second interaction node list is used to indicate the interactive information of article and other articles or user；According to training dataset, to marking Model is trained.

Scheme shown in the embodiment of the present invention, for training scoring model, server can predefine training dataset.Tool Body, the attribute data of the available multiple users of server, the attribute data of multiple articles and marking data, wherein multiple The attribute data of each user may include the mark of corresponding user in user, and the attribute data of each article can in multiple articles To include the mark of corresponding article, marking data may include in multiple users each user to an article in multiple articles or The marking of multiple articles.After getting the attribute data of multiple users, the attribute data of multiple articles and marking data, server It can be handled, obtain training dataset, wherein training dataset may include the mark of each user in multiple users Know and corresponding first interaction node list, multiple articles in each article mark and corresponding second interaction node list, Marking of each user to articles one or more in multiple articles.After obtaining training dataset, server can be beaten above-mentioned Sub-model is trained, it can is adjusted to the model parameter in scoring model, the scoring model after being trained.

In one possible implementation, the attribute data of each user further includes in following number breath in multiple users It is one or more: gender, height, weight, age, occupation, income, hobby, education landscape, the category of each article in multiple articles Property data further include one of following data or a variety of: brand, color, size, price, comment, taste, shelf-life, icon； Data of giving a mark further include one of following data or a variety of: operating time, currently used equipment, discount situation.

In one possible implementation, obtain the attribute data of multiple users, the attribute data of multiple articles and Give a mark data, comprising: obtain multiple marking record, in multiple marking records each marking record include user u attribute data, For the attribute data and user u of article i to the marking data of article i, user u is to beat article i in excessive multiple users Any user, article i are any article in multiple articles；To the attribute data of the attribute data of multiple users, multiple articles And marking data are handled, and training dataset is obtained, comprising: are handled multiple marking record, obtained training data Collection, it includes the mark of user u and the mark of corresponding first interaction node list, article i that training data, which concentrates each training data, Knowledge and corresponding second interaction node list, marking of the user u to article i.

Scheme shown in the embodiment of the present invention, the available multiple marking of server record, each in multiple marking records Marking record includes the marking data of the attribute data and user u of the attribute data of user u, article i to article i, article i For any article in multiple articles, user u is any user beaten article i in excessive multiple users, the attribute of user u Data include the mark of user u, and the attribute data of article i includes the mark of article i, and user u can be with to the marking data of article i Marking including user u to article i, wherein marking record is alternatively referred to as intersection record (for example, user bought certain article, then 1) marking data in corresponding marking record can be.For example, multiple marking records are respectively (u₀, i₀, 1), (u₀, i₁, 1), (u₀, i₂, 1).After getting multiple marking records, w is recorded for each marking in multiple marking record, can be remembered according to marking The marking record of w and time of origin before marking records w is recorded, the corresponding training data g of marking record w is obtained.For example, first It gets marking and is recorded as w₀(u₀, i₀, 1), since marking records w₀It gets for the first time, therefore, user u₀Corresponding first List is empty for interaction node, article i₀List is empty for corresponding second interaction node, obtained marking record w₀Corresponding trained number According to g₀For the mark u of user u₀, article i mark i₀, list is empty, article i is corresponding for corresponding first interaction node of user u List is empty for second interaction node, marking is 1；Next marking got is recorded as w₁(u₀, i₁, 1), it can be seen that, user u₀ To article i₀It beats excessively, article i₁It is not beaten excessively by other users, therefore, user u₀Corresponding first interaction node list is i₀, Article i₁List is empty for corresponding second interaction node, obtained marking record w₁Corresponding training data g₁For the mark of user u u₀, article i mark i₁, the corresponding first interaction node list of user u is i₀, the corresponding second interaction node list of article i is Empty, marking is 1；Then the marking record w got₂(u₁, i₁, 1), it can be seen that, user u₁Other articles are not beaten excessively, Article i₁By user u₀It beats excessively, therefore, user u₁List is empty for corresponding first interaction node, article i₁Corresponding second hands over Mutual node listing is u₀, obtain marking record w₂Corresponding training data g₂For the mark u of user u₁, article i mark i₁, use List is empty for corresponding first interaction node of family u, and the corresponding second interaction node list of article i is u₀, marking be 1.

In one possible implementation, scoring model includes feature learning model, feedback learning model and nerve net Network model；Wherein, according to training dataset, scoring model is trained, comprising: by the mark of user u, the mark of article i Input feature vector learning model obtains the corresponding feature vector of user u and the corresponding feature vector of article i, and user u is corresponding First interaction node list, the corresponding second interaction node list input feedback learning model of article i, it is corresponding to obtain user u Implicit feedback and the corresponding implicit feedback of article i；By the corresponding feature vector of user u and the corresponding feature vector of article i, user The corresponding implicit feedback of u and the corresponding implicit feedback of article i input the neural network model, obtain prediction score；According to pre- The marking of score and user u to article i is surveyed, feature learning model, feedback learning model and neural network model are adjusted It is whole, the scoring model after being trained.

Scheme shown in the embodiment of the present invention, after obtaining training dataset, training data is concentrated each training by server The mark of user u in data and the mark input feature vector learning model of article i, obtain the corresponding feature vector of user u and object The corresponding feature vector of product i, and can by each training data the corresponding first interaction node list of user u and article i Corresponding second interaction node list input feedback learning model, obtains the corresponding implicit feedback of user u and article i is corresponding hidden Formula feedback, wherein obtain the concrete mode of the corresponding feature vector of user u and the corresponding feature vector of article i and obtain target The mode of the corresponding feature vector of user feature vector corresponding with candidate item j is similar, obtains the corresponding implicit feedback of user u The concrete mode of implicit feedback corresponding with article i implicit feedback corresponding with target user is obtained and candidate item j are corresponding The mode of implicit feedback is similar, is no longer repeated herein.Obtain the corresponding feature vector of user u, the corresponding feature of article i to After amount, the corresponding implicit feedback of user u and the corresponding implicit feedback of article i, server can be inputted neural network model, Obtain prediction score.After user u is obtained to the prediction score of article i, it can be concentrated according to prediction score and training data every Marking of the user u to article i in one training data, to feature learning model, feedback learning model and neural network model Model parameter is adjusted, the scoring model after being trained, wherein can level off to user u to article i based on prediction score Marking training philosophy, the model parameter of feature learning model, feedback learning model and neural network model is adjusted, Scoring model after being trained.

In one possible implementation, the corresponding first interaction node list of user u includes multistage first interaction section Point list, the corresponding second interaction node list of article i include multistage second interaction node list, the model of feedback learning model Parameter includes: multistage user feedback matrix and multistage article feedback matrix, wherein the corresponding first interaction node list of user u Order it is identical as user feedback order of matrix number, the order of the corresponding second interaction node list of article i and article feed back square The order of battle array is identical, and odd-order the first interaction node list is used to indicate user and article in multistage first interaction node list Interactive information, even-order the first interaction node list is used to indicate user and other users in multistage first interaction node list Interactive information, odd-order the second interaction node list is for indicating article and the interaction of user in multistage second interaction node list Information, even-order the second interaction node list is used to indicate the interaction of article Yu other articles in multistage second interaction node list Information；The corresponding first interaction node list of user u, the corresponding second interaction node list input feedback of article i are learnt into mould Type obtains the corresponding implicit feedback of user u and the corresponding implicit feedback of article i, comprising: hands over user u corresponding multistage first Mutual node listing, the corresponding multistage second interaction node list input feedback learning model of article i, it is corresponding hidden to obtain user u Formula feeds back implicit feedback corresponding with article i.

Scheme shown in the embodiment of the present invention, server, can also be corresponding more using user u in training scoring model The list of the first interaction node of rank, the corresponding multistage second interaction node list of article i.For such situation, server can be incited somebody to action The corresponding multistage first interaction node list of user u and the corresponding multistage second interaction node list input feedback study of article i Model obtains the corresponding implicit feedback of user u and the corresponding implicit feedback of article i.

In one possible implementation, the model parameter of feedback learning model includes: each user in multiple users Feedback vector weight, the weight of the feedback vector of each article in multiple articles；By corresponding first interaction node of user u List, the corresponding second interaction node list input feedback learning model of article i, obtain the corresponding implicit feedback of user u and object The corresponding implicit feedback of product i, comprising: by the mark of user u and corresponding first interaction node list, the mark of article i and right The the second interaction node list input feedback learning model answered, obtains the corresponding implicit feedback of user u and article i is corresponding implicit Feedback.

Training data can be concentrated the user u in each training data by scheme shown in the embodiment of the present invention, server Mark and corresponding first interaction node list, article i mark and corresponding second interaction node list, input feedback Model is practised, the corresponding implicit feedback of user u and the corresponding implicit feedback of article i are obtained.

Second aspect provides a kind of training method of scoring model, this method comprises: obtaining the attribute number of multiple users According to the attribute data and marking data of, multiple articles；To the attribute data of multiple users, the attribute data of multiple articles and Marking data are handled, and training dataset is obtained, and training dataset includes the mark and corresponding first interaction of each user Node listing, the mark of each article and corresponding second interaction node list, each user are to one in multiple articles or more The marking of a article, the first interaction node list are used to indicate the interactive information of user and other users or article, the second interaction Node listing is used to indicate the interactive information of article and other articles or user；According to training dataset, scoring model is carried out Training.

The third aspect provides a kind of device for recommending article, which includes at least one module, at least one mould Block is for realizing the method for recommending article provided by above-mentioned first aspect.

Fourth aspect provides a kind of equipment, which includes processor, memory and transmitter, and processor is configured To execute the instruction stored in memory；Processor executes instruction so that the equipment realizes recommendation provided by above-mentioned first aspect The method of article.

5th aspect provides computer readable storage medium, including instruction, when the computer readable storage medium exists When being run on computer, so that the computer executes method described in above-mentioned first aspect.

6th aspect, provides a kind of computer program product comprising instruction, when the computer program product is being counted When being run on calculation machine, so that the computer executes method described in above-mentioned first aspect.

7th aspect, provides a kind of training device of scoring model, which includes at least one module, this at least one A module for realizing scoring model provided by above-mentioned second aspect training method.

Eighth aspect provides a kind of equipment, which includes processor, memory and transmitter, and processor is configured To execute the instruction stored in memory；Processor executes instruction so that the equipment realizes marking provided by above-mentioned second aspect The training method of model.

9th aspect provides computer readable storage medium, including instruction, when the computer readable storage medium exists When being run on computer, so that the computer executes method described in above-mentioned second aspect.

Tenth aspect, provides a kind of computer program product comprising instruction, when the computer program product is being counted When being run on calculation machine, so that the computer executes method described in above-mentioned second aspect.

Technical solution provided in an embodiment of the present invention has the benefit that

In the embodiment of the present invention, the attribute data of target user and the attribute data of multiple candidate items are obtained, target is used The attribute data at family includes the mark of target user, and the attribute data of each candidate item includes the mark of corresponding candidate item； The attribute data of the attribute data of target user and multiple candidate items is handled, target data set, target data are generated Collection includes each candidate item in the mark and the list of the first interaction node of corresponding target, multiple candidate items of target user Mark and the second interaction node list of corresponding target, target the first interaction node list is for indicating target user and other use The interactive information of family or article, the second interaction node list of target are used to indicate the interaction of candidate item and other articles or user Information；Target data set is inputted into scoring model, obtains marking of the target user to multiple candidate items, wherein scoring model It is obtained according to the attribute data of multiple users, the attribute data of multiple articles and marking data training, multiple users include Target user, the attribute data of each user includes the mark of corresponding user in multiple users, and multiple articles include multiple times Article is selected, the attribute data of each article includes the mark of corresponding article in multiple articles, and marking data include multiple users In marking of each user to articles one or more in multiple articles；Marking according to target user to multiple candidate items, Determine that target recommends article.In this way, target user can recommend in article in the target that server is recommended, choose what oneself was wanted Article, without being selected in all items that store in the server, it is thus possible to improve the efficiency that user selects article.

Detailed description of the invention

Fig. 1 is a kind of system framework schematic diagram provided in an embodiment of the present invention；

Fig. 2 is a kind of server architecture schematic diagram provided in an embodiment of the present invention；

Fig. 3 is a kind of bigraph (bipartite graph) schematic diagram provided in an embodiment of the present invention；

Fig. 4 is a kind of method flow diagram for recommending article provided in an embodiment of the present invention；

Fig. 5 is a kind of training method flow chart of scoring model provided in an embodiment of the present invention；

Fig. 6 is a kind of apparatus structure schematic diagram for recommending article provided in an embodiment of the present invention；

Fig. 7 is a kind of apparatus structure schematic diagram for recommending article provided in an embodiment of the present invention；

Fig. 8 is a kind of training device structural schematic diagram of scoring model provided in an embodiment of the present invention.

Specific embodiment

The embodiment of the invention provides a kind of method for recommending article, the executing subject of this method is equipment, which can To be server.Wherein, which can be the background server for recommending article function, which can be one individually Server, be also possible to the server group being made of multiple servers, the embodiment of the present invention with server be one individually It is described in detail for server, other situations are similar therewith, no longer repeated.To improve the effect that user selects article Rate can send the object of corresponding target user when target user wants selection article by operation triggering terminal to server Product recommendation request, correspondingly, after server receives article recommendation request, can in multiple candidate items in Candidate Set, It determines that the target that target user may like recommends article, in turn, target can be sent to terminal and recommend article, terminal receives Target can show it after recommending article, so that target user can recommend to select oneself to want in article in target Article, wherein system framework figure is as shown in Figure 1.

Server may include processor 210, transmitter 220, receiver 230 and memory 240, receiver 230 and hair Emitter 220, memory 240 can be connect with processor 210 respectively, as shown in Figure 2.Receiver 230 can be used for receiving message Or data, transmitter 220 and receiver 230 can be network interface card, transmitter 220 can be used for sending message or data, it can Target, which is sent, to the terminal of target user recommends article.Processor 210 can be the control centre of server, utilize various interfaces With the various pieces of the entire server of connection, such as receiver 230, transmitter 220 and memory 240.In the present invention, Processor 210 can be CPU (Central Processing Unit, central processing unit), be determined for target recommendation The relevant treatment of product, optionally, processor 210 may include one or more processing units；Processor 210 can integrate at Manage device and modem processor, wherein the main processing operation system of application processor, modem processor mainly handle nothing Line communication.Processor 210 can also be digital signal processor, specific integrated circuit, field programmable gate array or other Programmable logic device etc..Memory 240 can be used for storing software program and module, and processor 210 is stored in by reading The software code and module of memory, thereby executing the various function application and data processing of server.

For the ease of the understanding to the embodiment of the present invention, introduce first below the present embodiments relate to basic conception.

1, user-article bigraph (bipartite graph)

I.e. the interaction of user and article can be indicated by bigraph (bipartite graph), wherein the company side between user and article indicates should User's history interacted the article, and bigraph (bipartite graph) is described in detail below:

The available each user of server historical behavior data (for example, for the case where article is film, each user's Historical behavior data can be the film that each user once downloaded, watched, collecting), in turn, for each user, clothes Business device can establish the company side for the article that the user interacted with history according to the historical behavior data of the user, thus, it obtains User-article bigraph (bipartite graph).

For example, user-article bigraph (bipartite graph) is as shown in figure 3, user1 to the user5 in Fig. 3 respectively indicates 5 users, Item1 to item8 respectively indicates 8 articles.From a node in Fig. 3, be referred to as a step by a line, for For family, all nodes that a step can reach are article, all nodes that two steps can reach be with active user at least one All users of identical interactive article, wherein from user, all nodes that a step can reach are properly termed as single order first Interaction node list, all nodes that two steps can reach are properly termed as second order the first interaction node list, and so on, from article It sets out, all nodes that a step can reach are properly termed as single order the second interaction node list, and all nodes that two steps can reach can With referred to as the second interaction node of second order list, and so on.Server can obtain each user according to user-article bigraph (bipartite graph) Corresponding each rank the first interaction node list and corresponding each the second interaction node of the rank list of each article, for example, user1 is corresponding Single order the first interaction node list include item1, item2, item3, item8, second order the first interaction node list of user1 Including user2, user3, user5.In another example corresponding the second interaction node of the single order list of item1 includes user1 and user3, Corresponding the second interaction node of the second order list of item1 includes item2, item3, item4, item8.

Below in conjunction with specific embodiment, process flow shown in Fig. 4 is described in detail, content can be as Under:

Step 401, the attribute data of target user and the attribute data of multiple candidate items, the attribute of target user are obtained Data include the mark of target user, and the attribute data of each candidate item includes the mark of corresponding candidate item.

In an implementation, the article that corresponding each user can be previously stored in server recommends trigger event, wherein each The corresponding article of user recommends trigger event can be identical, recommends for example, article recommends trigger event to can be preset article Period, the corresponding article of each user recommend trigger event can also be different, for example, the corresponding article of each user recommends triggering Event can be the article recommendation request that the terminal of user is sent respectively.When server detects that the article of corresponding target user pushes away When recommending trigger event generation (for example, when detecting the article recommendation request of correspondence target user of terminal transmission), server It can determine that the favorite target of target user institute recommends article, and be recommended target user.Specifically, server can obtain Take the attribute data of each candidate item in the attribute data and multiple candidate items of target user, wherein the category of target user Property data may include target user mark, the attribute data of each candidate item may include corresponding in multiple candidate items The mark of candidate item.

Optionally, the attribute data of target user can also include one of following data or a variety of: gender, height, Weight, age, occupation, income, hobby, education landscape, the attribute data of each candidate item can also include in following data It is one or more: brand, color, size, price, comment, taste, shelf-life, icon.

Step 402, the attribute data of the attribute data of target user and multiple candidate items is handled, generates target Data set, target data set include the mark and the list of the first interaction node of corresponding target, multiple candidate items of target user In each candidate item mark and the second interaction node list of corresponding target, target the first interaction node list is for indicating The interactive information of target user and other users or article, target the second interaction node list is for indicating candidate item and other The interactive information of article or user.

It in an implementation, can after server gets the attribute data of target user and the attribute data of multiple candidate items To handle it, obtain target data set, wherein target data set may include target user mark and corresponding mesh Mark the first interaction node list, the mark of each candidate item and the second interaction node of corresponding target column in multiple candidate items Table, target the first interaction node list are used to indicate that the interactive information of target user and other users or article, target second to be handed over Mutual node listing is used to indicate the interactive information of candidate item and other articles or user.

It optionally, can be according to the mark of target user after obtaining the mark of target user and the mark of each candidate item Know the mark with each candidate item, determines that target the first interaction node list and the corresponding target second of each candidate item are handed over Mutual node listing, correspondingly, the treatment process of step 402 can be such that the mark according to target user, pre-recorded more In a user in corresponding the first interaction node of the target list of the mark of each user, the corresponding target first of target user is determined Interaction node list, and according to the mark of each candidate item, each candidate item in pre-recorded multiple candidate items Corresponding the second interaction node of the target list of mark in, determine each candidate item corresponding the second interaction node of target column Table；According to the mark of target user, corresponding the first interaction node of the target list of target user, each candidate item mark, And corresponding the second interaction node of the target list of each candidate item, generate target data set.

In an implementation, the corresponding target first of mark of each user in multiple users can be previously stored in server Corresponding the second interaction node of the target list of the mark of each candidate item in interaction node list, multiple candidate items, wherein Server can record in the form of a table each user corresponding target the first interaction node list of mark and each candidate Corresponding the second interaction node of the target list of the mark of product, the mark that each user can also be recorded in the form of bigraph (bipartite graph) are corresponding Target the first interaction node list and each candidate item corresponding the second interaction node of the target list of mark.Server obtains Get target user mark and each candidate item mark after, can in pre-recorded multiple users each user It identifies in corresponding the first interaction node of target list, determines corresponding the first interaction node of the target list of target user, and can With in corresponding the second interaction node of the target list of the mark of each candidate item in pre-recorded multiple candidate items, really Determine corresponding the second interaction node of the target list of each candidate item.Determine corresponding the first interaction node of target of target user After list, corresponding the second interaction node of the target list of each candidate item, the mark comprising target user is can be generated in server Knowledge and the list of the first interaction node of corresponding target, the mark of each candidate item and the second interaction node list of corresponding target Target data set, wherein it may include the mark and corresponding target of target user that target data, which concentrates each target data, First interaction node list, the mark of candidate item j and the second interaction node list of corresponding target, candidate item j is multiple Any candidate item in candidate item.

Step 403, target data set is inputted into scoring model, obtains target user and the multiple candidate item is beaten Point, wherein scoring model is obtained according to the attribute data of multiple users, the attribute data of multiple articles and marking data training , the attribute data of each user includes the mark of corresponding user in multiple users, the attribute of each article in multiple articles Data include the mark of corresponding article, marking data include in multiple users each user to one or more in multiple articles The marking of article.

In an implementation, scoring model can be previously stored in server, wherein scoring model can be server according to What the attribute data of multiple users, the attribute data of multiple articles and marking data training obtained, multiple users include target User, multiple articles include multiple candidate items.Server can predict target user to multiple candidates by scoring model The marking of each candidate item in product.Specifically, target data set can be inputted and be given a mark by server after generating target data set Model obtains marking of the target user to each candidate item in multiple candidate items, wherein for target data set including more Each target data can be inputted scoring model by the case where a target data, server, obtain target user to corresponding candidate The marking of article.

Optionally, scoring model may include feature learning model, feedback learning model and neural network model, accordingly , the treatment process of step 403 can be such that the mark for the target user for concentrating target data and the mark of candidate item j Input feature vector learning model, obtains the corresponding feature vector of target user and the corresponding feature vector of candidate item j, and by target Corresponding target the first interaction node list of target user and corresponding the second interaction node of target of candidate item j in data set List, input feedback learning model obtain the corresponding implicit feedback of target user and the corresponding implicit feedback of candidate item j, In, article j is any candidate item in multiple candidate items；The corresponding feature vector of target user, candidate item j is corresponding Feature vector, the corresponding implicit feedback of target user and the corresponding implicit feedback of candidate item j, input neural network model, Obtain marking of the target user to candidate item j.

In an implementation, scoring model may include feature learning model, feedback learning model and neural network model, In, feature learning model can be used for learning objective user and the corresponding feature vector of each candidate item, feature learning model Model parameter may include user characteristics matrix and article characteristics matrix, wherein user characteristics matrix is by multiple users (i.e. every row vector of user characteristics matrix is the feature vector of corresponding user, user respectively to the feature vector composition of each user The line number of eigenmatrix is the quantity of multiple users), article characteristics matrix be from the feature of each article in multiple articles to (i.e. every row vector of article characteristics matrix is the feature vector of corresponding article respectively to amount composition, and the line number of article characteristics matrix is It is the quantity of multiple articles).After obtaining target data set, server can by target data concentrate target user mark and The mark input feature vector learning model of candidate item j, obtains the corresponding feature vector of target user and the corresponding spy of candidate item j Levy vector.Specifically, server is by after the mark input feature vector learning model of the mark of target user and candidate item j, according to The mark of target user and the mark of candidate item j extract target user in user characteristics matrix by feature learning model Corresponding feature vector extracts the corresponding feature vector of candidate item j in article characteristics matrix, it is corresponding to obtain target user Feature vector and the corresponding feature vector of candidate item j.

Feedback learning model can be used for learning objective user and the corresponding implicit feedback of each candidate item, feedback learning The model parameter of model may include user feedback matrix (can be indicated with Y) and article feedback matrix (can be indicated with X), In, user feedback matrix can be by feedback vector form (every row vector in user feedback matrix represents a node pair The feedback vector answered), article feedback matrix can be by feedback vector form (every row vector in article feedback matrix represents The corresponding feedback vector of one node).Determine corresponding the first interaction node of the target list of target user's (can be indicated with k) (R can be used_kIndicate) afterwards, corresponding the second interaction node of the target list of candidate item j (R can be used_jIndicate) after, it can be by it Input feedback learning model obtains the corresponding implicit feedback of target user k and the corresponding implicit feedback of candidate item j.Specifically, Server can be learned after corresponding the first interaction node of the target list input feedback learning model of target user k by feedback Practise model extracted in user feedback matrix the corresponding multiple feedback vectors of target the first interaction node list (wherein, feed back to The quantity of amount is the quantity for the node for including in target the first interaction node list), obtain target user k it is corresponding feed back to Amount.After getting the corresponding feedback vector of target user k, multiple feedback vectors can be added, it is corresponding to obtain target user k Implicit feedback, wherein server can obtain the corresponding implicit feedback P of target user k according to formula (1)_k,

Wherein, every row vector can use Y in user feedback matrix_tIt indicates, wherein mark of the t for node, t=1,2 ..., M, M are total line number of user feedback matrix.It should be noted that when user feedback matrix is odd-order user feedback matrix, M For the total quantity of multiple articles, when user feedback matrix is even-order user feedback matrix, M is the total quantity of multiple users.

The specific processing for obtaining the corresponding implicit feedback of candidate item j can be such that server is corresponding by candidate item j After target the second interaction node list input feedback learning model, it can be mentioned in article feedback matrix by feedback learning model Taking the corresponding multiple feedback vectors of target the second interaction node list, (wherein, the quantity of feedback vector is the interaction of target second The quantity for the node for including in node listing), obtain the corresponding feedback vector of candidate item j.It is corresponding to get candidate item j After feedback vector, multiple feedback vectors can be added, obtain the corresponding implicit feedback of candidate item j, wherein server can be with According to formula (2), the corresponding implicit feedback Q of candidate item j is obtained_j,

Wherein, every row vector can use X in article feedback matrix_vIt indicates, wherein mark of the v for node, v=1,2 ..., N, N are total line number of article feedback matrix.It should be noted that when article feedback matrix is odd-order article feedback matrix, N For the total quantity of multiple users, when article feedback matrix is even-order article feedback matrix, N is the total quantity of multiple articles.

Specifically, the neural network model trained in advance can be previously stored in server, wherein neural network mould Type may include multilayer neural network, and the input of each layer of neural network in multilayer neural network can be one layer of nerve net The output of network, wherein the formula of h layers of neural network can as shown in formula (3),

r^h+1=σ (W^hr^h+b^h) (3)

Wherein, σ () is known as activation primitive, for example can be sigmoid function, relu function, tanh function etc., r^hIt is H layers of input, b^hFor h layers of shift term, W^hConnect the weight on side for h layers of neuron and h+1 layers of neuron, wherein W^h And b^hAnd training obtains, the input r of first layer neural network¹Can as shown in formula (4),

r¹=<p+P,q+Q>(4)

Wherein, the numerical value of<x, y>expression vector x dimension corresponding with vector y is multiplied, r¹For vector, p indicates the corresponding feature of user Vector, P indicate that the corresponding implicit feedback of user, q indicate that the corresponding feature vector of article, Q indicate the corresponding implicit feedback of article. Therefore, neural network model can be as shown in formula (5), wherein H is total number of plies of neural network model.

Y=σ (W^H(σ(W^H-1(σ(...σ(W¹r¹+b¹)+...+b^H-1))+b^H) (5)

Server determines the corresponding feature vector p of target user k_k, the corresponding implicit feedback P of target user k_k, candidate The corresponding feature vector q of product j_j, the corresponding implicit feedback Q of candidate item j_jIt afterwards, can be by r¹ _kjAs the defeated of neural network model Enter, take in formula (5), obtains marking of the target user k to candidate item j.Wherein, r¹ _kjAs shown in formula (6).

In addition, can also include b in formula (6)_k、b_jWith b's and (b_k、b_jIt can be described as statistical basis point with the sum of b Number), wherein b is the mean value that training data concentrates all marking for including, b_kThe target user k for including is concentrated for training data The difference of mean value and b to all marking of each article, b_jMarking of all users for including to article j is concentrated for training data Mean value and b difference.

Optionally, the list of the first interaction node of target may include multistage the first interaction node of target list, each candidate Corresponding the second interaction node of the target list of article may include multistage the second interaction node of target list, and multistage target first is handed over The the first interaction node list of odd-order target is used to indicate the interactive information of target user and article, multistage mesh in mutual node listing Mark the friendship that even-order target the first interaction node list in the first interaction node list is used to indicate target user and other users Mutual information, in multistage the second interaction node of target list the second interaction node list of odd-order target for indicate candidate item with The interactive information of user, the second interaction node list of even-order target is waited for indicating in multistage the second interaction node of target list Select the interactive information of article Yu other articles；The model parameter of feedback learning model may include multistage user feedback matrix and more Rank article feedback matrix, wherein the order of target the first interaction node list is identical as user feedback order of matrix number, Mei Yihou Select the order of corresponding the second interaction node of the target list of article identical as the order of article feedback matrix.For such situation, Correspondingly, the concrete processing procedure for obtaining the corresponding implicit feedback of target user and the corresponding implicit feedback of candidate item j can be with As follows: corresponding multistage target the first interaction node list of target user and candidate item j that target data is concentrated are corresponding Multistage the second interaction node of target list, input feedback learning model obtain the corresponding implicit feedback of target user and candidate The corresponding implicit feedback of product j.

In an implementation, server can also utilize target user couple when predicting marking of the target user to candidate item j The list of the first interaction node of multistage target, corresponding multistage the second interaction node of the target list of candidate item j answered, wherein more The list of the first interaction node of rank target can be the list of the first interaction node of single order target, the first interaction node of second order target respectively List ..., target the first interaction node list of A rank, multistage user feedback matrix may include single order user feedback matrix, second order User feedback matrix ..., A rank user feedback matrix, A is default value (such as A be 3), A be preset target user with The maximum step number that can be reached in family-article bigraph (bipartite graph), multistage the second interaction node of target list can be single order target respectively Two interaction node lists, the list of the second interaction node of second order target ..., target the second interaction node list of B rank, multistage article is anti- Feedback matrix may include single order article feedback matrix, second order article feedback matrix ..., B rank article feedback matrix, B is present count Value, B is the maximum step number that preset candidate item j can be reached in user-article bigraph (bipartite graph), wherein A can be identical with B, It can be different.Single order user feedback matrix, can use Y¹, every row vector in single order user feedback matrix can be corresponding object Vector when product are as node in the first interaction node list of single order target indicates that second order user feedback matrix can use Y² It indicates, every row vector in second order user feedback matrix can be corresponding user as second order target the first interaction node list In node when vector indicate, and so on.Single order article feedback matrix, can use X¹, in single order article feedback matrix Every row vector can be vector of the corresponding user as node in the second interaction node list of single order target when and indicate, second order Article feedback matrix can use X²It indicates, every row vector in second order article feedback matrix can be corresponding article as two Vector when node in the second interaction node list of rank target indicates, and so on.

For such situation, the corresponding multistage interaction of target first of the target user that server can concentrate target data Node listing and corresponding multistage the second interaction node of the target list of candidate item j, are input to feedback learning model, obtain target The corresponding implicit feedback of user and the corresponding implicit feedback of candidate item j.Specifically, being arranged for every the first interaction node of rank target Table(a=1,2 ..., A), server can be by feedback learning models, in rank user feedback matrix Y^aIn, extract target First interaction node listCorresponding feedback vector.Server it is corresponding can to select target user in the manner described above All feedback vectors selected can be added, obtain in turn by the corresponding feedback vector of each the first interaction node of rank target list To the corresponding implicit feedback of target user.

Every rank target corresponding for candidate item j the second interaction node list(b=1,2 ..., B), server can be with By feedback learning model, in article feedback matrix X^bIn, choose the list of the second interaction node of targetCorresponding feedback vector, Obtain corresponding the second interaction node of the target list of candidate item jCorresponding feedback vector.Server can be according to above-mentioned side Formula, selecting the corresponding feedback vector of corresponding each the second interaction node of the rank target list of candidate item j can will select in turn All feedback vectors taken out are added, and obtain the corresponding implicit feedback of candidate item j.

Optionally, the model parameter of feedback learning model can also include: the feedback vector of each user in multiple users Weight, the weight of the feedback vector of each article in multiple articles, wherein weight can be that training obtains in advance by server , in such cases, obtains the corresponding implicit feedback of target user and the specific of the corresponding implicit feedback of candidate item j processes Journey can be such that the mark and the list of the first interaction node of corresponding target, candidate for the target user for concentrating target data It is corresponding implicit to obtain target user for the mark of product j and the second interaction node list of corresponding target, input feedback learning model Feed back implicit feedback corresponding with candidate item j.

In an implementation, the model parameter for feedback learning model further includes the feedback vector of each user in multiple users Weight and multiple articles in each article feedback vector weight the case where, server can by target data concentrate mesh Mark mark and the list of the first interaction node of corresponding target, the interaction of the mark of candidate item j and corresponding target second of user Node listing, input feedback learning model, obtains the corresponding implicit feedback of target user and candidate item j is corresponding implicit anti- Feedback.Specifically, server can pass through according to the method for above-mentioned determining target user and the corresponding feedback vector of candidate item j Feedback learning model determines the corresponding feedback vector of target user and the corresponding feedback vector of candidate item j.Then, server can To pass through the feedback vector of the target user in feedback learning model according to the mark of the mark of target user and candidate item j Weight (Φ can be used_ktIndicate) the corresponding feedback vector of target user k is weighted and is handled, it is corresponding to obtain target user Implicit feedback, and Ω (can be used by the weight of the feedback vector of the candidate item j in feedback learning model_vjIndicate) to time It selects the corresponding feedback vector of article j to be weighted and handle, obtains the corresponding implicit feedback of candidate item j.

Optionally, the model parameter of feedback learning model can include: simultaneously each user in multiple users feedback to The weight of the feedback vector of each article, multistage user feedback matrix and multistage article feed back square in the weight of amount, multiple articles Battle array, correspondingly, corresponding the first interaction node of the target list of target user may include multistage the first interaction node of target list, Corresponding the second interaction node of the target list of each candidate item may include multistage the second interaction node of target list, accordingly , determine that the corresponding implicit feedback of target user and the treatment process of the corresponding implicit feedback of candidate item j can be such that mesh Mark data set in target user mark and corresponding the first interaction node of multistage target list, candidate item j mark and Corresponding the second interaction node of multistage target list, input feedback learning model, obtain the corresponding implicit feedback of target user and The corresponding implicit feedback of candidate item j.

In an implementation, server can select the corresponding each interaction of rank target first of target user in the manner described above The corresponding feedback vector of node listing and corresponding each the second interaction node of the rank target list of candidate item j it is corresponding feed back to Amount.Then, server can pass through the mesh in feedback learning model according to the mark of target user and the mark of candidate item j The weight for marking the feedback vector of user (can be usedIndicate) the corresponding feedback vector of target user k is weighted and is handled, The corresponding implicit feedback of target user is obtained, and (can by the weight of the feedback vector of the candidate item j in feedback learning model With withIndicate) the corresponding feedback vector of candidate item j is weighted and is handled, it is corresponding implicit anti-to obtain candidate item j Feedback.

Specifically, server can obtain the corresponding implicit feedback P of target user according to formula (7)_k,

Wherein,Indicate the feedback vector of target user kWeight.

Server can obtain the corresponding implicit feedback Q of candidate item j according to formula (8)_j,

Wherein,Indicate the feedback vector of candidate item jWeight.

Step 404, the marking according to target user to multiple candidate items determines that target recommends article.

In an implementation, after obtaining target user to the marking of multiple candidate items, server can be in multiple candidate items In, the marking according to target user to each candidate item of multiple candidate items determines that the target to be recommended to target user pushes away Article is recommended, in turn, target can be recommended to recommend article to target user.

Optionally, default recommendation condition has been can store in server, correspondingly, the treatment process of step 404 can be as Under: the marking according to target user to multiple candidate items determines that corresponding marking meets the target recommendation of default recommendation condition Article.

Wherein, it presets recommendation condition and can be and judge what whether certain article was recommended according to corresponding marking for server Condition.

In an implementation, default recommendation condition can be previously stored in server, server obtains target user to multiple , can be in multiple candidate items after the marking of candidate item, the target for choosing the corresponding default recommendation condition of marking satisfaction pushes away Recommend article.

Optionally, different based on default recommendation condition, determine that target recommends the processing mode of article can be varied, with Under give several feasible processing modes:

Mode one, the marking according to target user to multiple candidate items determine the corresponding maximum preset number of marking A target recommends article.

In an implementation, it after server determines target user to the marking of multiple candidate items, can be beaten according to corresponding Divide descending sequence, multiple candidate items are ranked up, in turn, the preset number candidate item that will sort forward, It is determined as target and recommends article.

Mode two, the marking according to target user to multiple candidate items determine that corresponding marking is greater than preset fraction threshold The target of value recommends article.

In an implementation, preset fraction threshold value can be previously stored in server.Server determines target user to more After the marking of a candidate item, the candidate that corresponding marking is greater than preset fraction threshold value can be chosen in multiple candidate items Determining candidate item can be determined as target and recommend article by article in turn.

The embodiment of the invention also provides a kind of training methods of scoring model, right below in conjunction with specific embodiment Process flow shown in fig. 5 is described in detail, and content can be such that

Step 501, the attribute data of multiple users, the attribute data of multiple articles and marking data are obtained.

It in an implementation, is training scoring model, server can predefine training dataset.Specifically, server can To obtain the attribute data of multiple users, the attribute data and marking data of multiple articles, wherein each user in multiple users Attribute data may include corresponding user mark, the attribute data of each article may include corresponding article in multiple articles Mark, marking data may include in multiple users each user an article or multiple articles in multiple articles are beaten Point.

Optionally, the attribute data of each user can also include one of following number breath or a variety of in multiple users: Gender, height, weight, age, occupation, income, hobby, education landscape, the attribute data of each article may be used also in multiple articles To include one of following data or a variety of: brand, color, size, price, comment, taste, shelf-life, icon；It grades According to can also include one of following data or a variety of: operating time, currently used equipment, discount situation.

Optionally, server can be recorded by obtaining multiple marking, to obtain attribute data, the multiple objects of multiple users The attribute data and marking data of product, correspondingly, the treatment process of step 501, which can be such that, obtains multiple marking records, it is more Each marking record includes the attribute data of user u, the attribute data of article i and user u to article i in a marking record Marking data, user u is any user beaten article i in excessive multiple users, and article i is any in multiple articles Article.

In an implementation, the available multiple marking of server record, and each marking record includes using in multiple marking records To the marking data of article i, article i is in multiple articles by the attribute data of family u, the attribute data of article i and user u Any article, user u are any user beaten article i in excessive multiple users, and the attribute data of user u includes user u Mark, the attribute data of article i includes the mark of article i, and user u may include user u to object to the marking data of article i The marking of product i, wherein marking record is alternatively referred to as intersection record, and (for example, user bought certain article, then corresponding marking is remembered 1) marking data in record can be.For example, multiple marking records are respectively (u₀, i₀, 1), (u₀, i₁, 1), (u₀, i₂, 1).

Step 502, the attribute data to the attribute data of multiple users, multiple articles and marking data are handled, Training dataset is obtained, training dataset includes the mark and corresponding first interaction node list, each article of each user The marking to articles one or more in multiple articles of mark and corresponding second interaction node list, each user, first Interaction node list is used to indicate the interactive information of user and other users or article, and the second interaction node list is used for expression thing The interactive information of product and other articles or user.

In an implementation, after getting the attribute data of multiple users, the attribute data of multiple articles and marking data, service Device can be handled it, obtain training dataset, wherein training dataset may include each user in multiple users The mark of each article and corresponding second interaction node column in mark and corresponding first interaction node list, multiple articles Table, marking of each user to articles one or more in multiple articles.

Optionally, for the case where multiple marking record is obtained, correspondingly, the treatment process of step 502 can be such that pair Multiple marking record is handled, and obtains training dataset, training data concentrate each training data include user u mark and Article i is beaten in corresponding first interaction node list, the mark of article i and corresponding second interaction node list, user u Point.

In an implementation, after getting multiple marking records, w, Ke Yigen are recorded for each marking in multiple marking record According to the marking record of marking record w and time of origin before marking records w, the corresponding training data g of marking record w is obtained. For example, the marking got first is recorded as w₀(u₀, i₀, 1), since marking records w₀It gets for the first time, therefore, user u₀ List is empty for corresponding first interaction node, article i₀List is empty for corresponding second interaction node, obtained marking record w₀It is right The training data g answered₀For the mark u of user u₀, article i mark i₀, list is empty, object for corresponding first interaction node of user u List is empty for corresponding second interaction node of product i, marking is 1；Next marking got is recorded as w₁(u₀, i₁, 1), thus may be used See, user u₀To article i₀It beats excessively, article i₁It is not beaten excessively by other users, therefore, user u₀Corresponding first interaction node List is i₀, article i₁List is empty for corresponding second interaction node, obtained marking record w₁Corresponding training data g₁For with The mark u of family u₀, article i mark i₁, the corresponding first interaction node list of user u is i₀, corresponding second interaction of article i Node listing is sky, marking is 1；Then the marking got is recorded as w₂(u₁, i₁, 1), it can be seen that, user u₁Not to other Article is beaten excessively, article i₁By user u₀It beats excessively, therefore, user u₁List is empty for corresponding first interaction node, article i₁It is right The the second interaction node list answered is u₀, obtain marking record w₂Corresponding training data g₂For the mark u of user u₁, article i Mark i₁, list is empty for corresponding first interaction node of user u, and the corresponding second interaction node list of article i is u₀, marking It is 1.

Step 503, according to training dataset, scoring model is trained.

In an implementation, after obtaining training dataset, server can be trained above-mentioned scoring model, it can air exercise Model parameter in sub-model is adjusted, the scoring model after being trained.

Optionally, feature learning model, feedback learning model and neural network model are included the case where for scoring model, Server can be unified to be trained feature learning model, feedback learning model and neural network model, correspondingly, step 503 treatment process can be such that the mark input feature vector learning model of the mark of user u, article i, and it is corresponding to obtain user u Feature vector and the corresponding feature vector of article i, and by the corresponding first interaction node list of user u, article i corresponding Two interaction node list input feedback learning models, obtain the corresponding implicit feedback of user u and the corresponding implicit feedback of article i； The corresponding feature vector of user u and the corresponding feature vector of article i, the corresponding implicit feedback of user u and article i is corresponding hidden Formula feed back input neural network model obtains prediction score；Marking according to prediction score and user u to article i, to feature Learning model, feedback learning model and neural network model are adjusted, the scoring model after being trained.

In an implementation, after obtaining training dataset, training data is concentrated the user u's in each training data by server Mark and article i mark input feature vector learning model, obtain the corresponding feature vector of user u and the corresponding feature of article i to Amount, and can by each training data the corresponding first interaction node list of user u and article i it is corresponding second interaction save Point list input feedback learning model obtains the corresponding implicit feedback of user u and the corresponding implicit feedback of article i, wherein obtain The concrete mode of the corresponding feature vector of user u and the corresponding feature vector of article i feature corresponding with target user is obtained to The mode for measuring feature vector corresponding with candidate item j is similar, obtains the corresponding implicit feedback of user u and article i is corresponding hidden The mode class of concrete mode corresponding with target user the is obtained implicit feedback and the corresponding implicit feedback of candidate item j of formula feedback Seemingly, it is no longer repeated herein.It is corresponding hidden to obtain the corresponding feature vector of user u, the corresponding feature vector of article i, user u After formula feeds back implicit feedback corresponding with article i, it can be inputted neural network model, obtain prediction score.Obtain user u After the prediction score of article i, the user u in each training data can be concentrated to object according to prediction score and training data The marking of product i is adjusted the model parameter of feature learning model, feedback learning model and neural network model, is instructed Scoring model after white silk, wherein user u can be leveled off to based on prediction score to the training philosophy of the marking of article i, to feature The model parameter of learning model, feedback learning model and neural network model is adjusted, the scoring model after being trained.

Optionally, the corresponding first interaction node list of user u may include multistage first interaction node list, article i Corresponding second interaction node list may include multistage second interaction node list, and the model parameter of feedback learning model can be with It include: multistage user feedback matrix and multistage article feedback matrix, wherein the rank of the corresponding first interaction node list of user u Number is identical as user feedback order of matrix number, order and the article feedback matrix of the corresponding second interaction node list of article i Order is identical, and odd-order the first interaction node list is used to indicate the interaction of user and article in multistage first interaction node list Information, even-order the first interaction node list is used to indicate the interaction of user and other users in multistage first interaction node list Information, odd-order the second interaction node list is used to indicate that article and user's to interact letter in multistage second interaction node list Breath, what even-order the second interaction node list was used to indicate article and other articles in multistage second interaction node list interacts letter Breath.For such situation, correspondingly, determining the specific processing of user u corresponding implicit feedback and the corresponding implicit feedback of article i Process can be such that the corresponding multistage first interaction node list of user u, the corresponding multistage second interaction node column of article i Table input feedback learning model obtains the corresponding implicit feedback of user u and the corresponding implicit feedback of article i.

In an implementation, server can also utilize corresponding multistage first interaction node of user u in training scoring model List, the corresponding multistage second interaction node list of article i.For such situation, server can be corresponding multistage by user u First interaction node list and the corresponding multistage second interaction node list input feedback learning model of article i, obtain u pairs of user The implicit feedback and the corresponding implicit feedback of article i answered.

Optionally, the model parameter of feedback learning model may include: the feedback vector of each user in multiple users Weight, the weight of the feedback vector of each article in multiple articles.In such cases, determine the corresponding implicit feedback of user u and The concrete processing procedure of the corresponding implicit feedback of article i can be such that the mark of user u and corresponding first interaction node column It is corresponding implicit anti-to obtain user u for table, the mark of article i and corresponding second interaction node list input feedback learning model Present implicit feedback corresponding with article i.

In an implementation, training data can be concentrated the mark of the user u in each training data and corresponding by server First interaction node list, the mark of article i and corresponding second interaction node list, input feedback learning model are used The corresponding implicit feedback of family u and the corresponding implicit feedback of article i.

Based on the same technical idea, the embodiment of the invention also provides a kind of devices for recommending article, as shown in fig. 6, The device includes:

Module 610 is obtained, for obtaining the attribute data of target user and the attribute data of multiple candidate items, the mesh The attribute data of mark user includes the mark of target user, and the attribute data of each candidate item includes the mark of corresponding candidate item Know, the acquisition function and other implicit steps in above-mentioned steps 401 specifically may be implemented.

Generation module 620, for by the attribute data of the attribute data of the target user and the multiple candidate item It is handled, generates target data set, the target data set includes the mark and corresponding target first of the target user The mark of each candidate item and the second interaction node of corresponding target column in interaction node list, the multiple candidate item Table, target the first interaction node list are used to indicate the interactive information of the target user and other users or article, institute Target the second interaction node list is stated for indicating the interactive information of candidate item and other articles or user, specifically may be implemented Systematic function and other implicit steps in above-mentioned steps 402.

Scoring modules 630 obtain the target user to described more for the target data set to be inputted scoring model The marking of a candidate item, wherein the scoring model according to the attribute data of the attribute data of multiple users, multiple articles with And giving a mark what data training obtained, the attribute data of each user includes the mark of corresponding user, institute in the multiple user The attribute data for stating each article in multiple articles includes the mark of corresponding article, and the marking data include the multiple use Marking of each user to articles one or more in the multiple article, specifically may be implemented in above-mentioned steps 403 in family Function of giving a mark and other implicit steps.

Determining module 640 determines that target is recommended for the marking according to the target user to the multiple candidate item Determination function and other implicit steps in above-mentioned steps 404 specifically may be implemented in article.

Optionally, the attribute data of the target user further includes one of following data or a variety of: gender, height, Weight, age, occupation, income, hobby, education landscape, the attribute data of each candidate item further include one in following data Kind is a variety of: brand, color, size, price, comment, taste, shelf-life, icon.

Optionally, the generation module 620, is used for:

According to the mark of the target user, the corresponding target of mark of each user in pre-recorded multiple users In first interaction node list, corresponding the first interaction node of the target list of the target user is determined, and according to each candidate The mark of article, corresponding the second interaction node of target of the mark of each candidate item in pre-recorded multiple candidate items In list, corresponding the second interaction node of the target list of each candidate item is determined；

According to the mark of the target user, corresponding the first interaction node of the target list of the target user, Mei Yihou The mark and corresponding the second interaction node of the target list of each candidate item of article are selected, target data set is generated.

Optionally, the scoring model includes feature learning model, feedback learning model and neural network model；

Wherein, scoring modules 630 are used for:

The mark for the target user that the target data is concentrated and the mark of candidate item j input the feature learning mould Type, obtains the corresponding feature vector of the target user and the corresponding feature vector of the candidate item j, and by the number of targets It is handed over according to corresponding target the first interaction node list of the target user of concentration and the corresponding target second of the candidate item j Mutual node listing inputs the feedback learning model, obtains the corresponding implicit feedback of the target user and the candidate item j Corresponding implicit feedback, wherein the article j is any candidate item in the multiple candidate item；

By the corresponding feature vector of the target user, the corresponding feature vector of the candidate item j, the target user Corresponding implicit feedback and the corresponding implicit feedback of the candidate item j input neural network model, obtain the target user Marking to candidate item j.

Optionally, the first interaction node of target list includes multistage the first interaction node of target list, each candidate Corresponding the second interaction node of the target list of article includes multistage the second interaction node of target list, the multistage interaction of target first section The the first interaction node list of odd-order target is used to indicate the interactive information of target user and article in point list, multistage target the What the first interaction node list of even-order target was used to indicate target user and other users in one interaction node list interacts letter It ceases, the second interaction node list of odd-order target is for indicating candidate item and user in multistage the second interaction node of target list Interactive information, the second interaction node list of even-order target is for indicating candidate in multistage the second interaction node of target list The interactive information of product and other articles；

The scoring modules 630, are used for:

Corresponding multistage target the first interaction node list of the target user that the target data is concentrated and candidate Corresponding multistage the second interaction node of the target list of article j, inputs the feedback learning model, and it is corresponding to obtain the target user Implicit feedback and the corresponding implicit feedback of candidate item j.

Optionally, the model parameter of the feedback learning model include: each user in the multiple user feedback to The weight of amount, the weight of the feedback vector of each article in the multiple article；

The scoring modules 630, are used for:

The mark for the target user that the target data is concentrated and the list of the first interaction node of corresponding target are waited The mark and the second interaction node list of corresponding target for selecting article j, input the feedback learning model, obtain the target and use The corresponding implicit feedback in family and the corresponding implicit feedback of candidate item j.

Optionally, the determining module 640, is used for:

Marking according to the target user to the multiple candidate item determines that corresponding marking meets default recommendation item The target of part recommends article.

Optionally, the determining module 640, is used for:

Marking according to the target user to the multiple candidate item determines the corresponding maximum preset number of marking A target recommends article；Alternatively,

Marking according to the target user to the multiple candidate item determines that corresponding marking is greater than preset fraction threshold The target of value recommends article.

Optionally, as shown in fig. 7, the acquisition module 610, is also used to:

Obtain the attribute data of the multiple user, the attribute data of the multiple article and the marking data；

The generation module 620, is also used to:

At the attribute data of the multiple user, the attribute data of the multiple article and the marking data Reason, obtains training dataset, and the training dataset includes the mark of each user and corresponding first interaction node list, every The mark of one article and corresponding second interaction node list, each user are to articles one or more in the multiple article Marking, the first interaction node list are used to indicate the interactive information of user and other users or article, second interaction Node listing is used to indicate the interactive information of article and other articles or user；

Described device further include:

Training module 650, for being trained to scoring model according to the training dataset.

Optionally, the attribute data of each user further includes one of following number breath or a variety of in the multiple user: Gender, height, weight, age, occupation, income, hobby, education landscape, the attribute data of each article in the multiple article It further include one of following data or a variety of: brand, color, size, price, comment, taste, shelf-life, icon；It is described to beat Divided data further includes one of following data or a variety of: operating time, currently used equipment, discount situation.

Optionally, the acquisition module 610, is used for:

Multiple marking records are obtained, each marking record includes attribute data, the article i of user u in multiple marking records Attribute data and user u to the marking data of article i, the user u is beat the article i excessively the multiple Any user in user, the article i are any article in multiple articles；

The generation module 620, is used for:

Multiple marking record is handled, obtains training dataset, it includes using that training data, which concentrates each training data, The mark of family u and the mark and corresponding second interaction node list, u pairs of user of corresponding first interaction node list, article i The marking of article i.

Wherein, the training module 650, is used for:

The mark of the user u, the mark of the article i are inputted into the feature learning model, obtain u couples of the user The feature vector and the corresponding feature vector of the article i answered, and by the corresponding first interaction node list of the user u, institute It states the corresponding second interaction node list of article i and inputs the feedback learning model, obtain the corresponding implicit feedback of the user u Implicit feedback corresponding with the article i；

The corresponding feature vector of the user u and the corresponding feature vector of the article i, the user u is corresponding hidden Formula feeds back implicit feedback corresponding with the article i and inputs the neural network model, obtains prediction score；

Marking according to the prediction score and the user u to the article i, to the feature learning model, institute It states feedback learning model and the neural network model is adjusted, the scoring model after being trained.

Optionally, the corresponding first interaction node list of the user u includes multistage first interaction node list, the object The corresponding second interaction node list of product i includes multistage second interaction node list, the model parameter of the feedback learning model It include: multistage user feedback matrix and multistage article feedback matrix, wherein the corresponding first interaction node list of the user u Order it is identical as the user feedback order of matrix number, the order of the corresponding second interaction node list of the article i and institute The order for stating article feedback matrix is identical, and odd-order the first interaction node list is for indicating in multistage first interaction node list The interactive information of user and article, even-order the first interaction node list is for indicating user in multistage first interaction node list With the interactive information of other users, odd-order the second interaction node list is for indicating article in multistage second interaction node list With the interactive information of user, even-order the second interaction node list is for indicating article and its in multistage second interaction node list The interactive information of his article；

The training module 650, is used for:

By the corresponding multistage first interaction node list of the user u, corresponding multistage second interaction node of the article i List inputs the feedback learning model, obtains the corresponding implicit feedback of the user u and the article i is corresponding implicit anti- Feedback.

The training module 650, is used for:

By the mark of the user u and the mark and corresponding second of corresponding first interaction node list, the article i Interaction node list inputs the feedback learning model, obtains the corresponding implicit feedback of the user u and the article i is corresponding Implicit feedback.

It should be noted that above-mentioned acquisition module 610, generation module 620, scoring modules 630, determining module 640, training Module 650 can be realized by processor or processor cooperates memory to realize, alternatively, processor executes in memory Program instruction is realized or processor cooperation memory, transmitter are realized.

It should be understood that it is provided by the above embodiment recommend article device when recommending article, only with above-mentioned each function Can module division progress for example, in practical application, can according to need and by above-mentioned function distribution by different functions Module is completed, i.e., the internal structure of server is divided into different functional modules, to complete whole described above or portion Divide function.In addition, the embodiment of the method for the device provided by the above embodiment for recommending article and recommendation article belongs to same design, Its specific implementation process is detailed in embodiment of the method, and which is not described herein again.

Based on the same technical idea, the embodiment of the invention also provides a kind of training devices of scoring model, such as Fig. 8 institute Show, which includes:

Obtain module 810, for obtain the attribute data of the multiple user, the multiple article attribute data and Acquisition function and other implicit steps in above-mentioned steps 501 specifically may be implemented in the marking data.

Generation module 820, attribute data and institute for attribute data, the multiple article to the multiple user It states marking data to be handled, obtains training dataset, the training dataset includes the mark and corresponding the of each user One interaction node list, the mark of each article and corresponding second interaction node list, each user are to the multiple article The marking of middle one or more article, the first interaction node list are used to indicate the interaction of user and other users or article Information, the second interaction node list are used to indicate the interactive information of article and other articles or user, specifically may be implemented Systematic function and other implicit steps in above-mentioned steps 502.

Training module 830, for being trained to scoring model, specifically may be implemented according to the training dataset State the training function and other implicit steps in step 503.

Optionally, the acquisition module 810, is used for:

The generation module 820, is used for:

Wherein, the training module 830, is used for:

The training module 830, is used for:

It should be noted that above-mentioned acquisition module 810, generation module 820, training module 830 can be realized by processor, Or processor cooperation memory is realized, alternatively, processor executes the program instruction in memory to realize or processor Memory, transmitter is cooperated to realize.

The training device of scoring model provided by the above embodiment is in training scoring model, only with above-mentioned each functional module Division progress for example, in practical application, can according to need and above-mentioned function distribution is complete by different functional modules At the internal structure of server being divided into different functional modules, to complete all or part of the functions described above. In addition, the training device of scoring model provided by the above embodiment and the training method embodiment of scoring model belong to same structure Think, specific implementation process is detailed in embodiment of the method, and which is not described herein again.

Those of ordinary skill in the art will appreciate that realizing that all or part of the steps of above-described embodiment can pass through hardware It completes, relevant hardware can also be instructed to complete by program, the program can store in a kind of computer-readable In storage medium, storage medium mentioned above can be read-only memory, disk or CD etc..

The foregoing is merely one embodiment of the invention, all in spirit herein and original not to limit the application Within then, any modification, equivalent replacement, improvement and so on be should be included within the scope of protection of this application.

Claims

1. a kind of method for recommending article, which is characterized in that the described method includes:

The attribute data of target user and the attribute data of multiple candidate items are obtained, the attribute data of the target user includes The mark of target user, the attribute data of each candidate item include the mark of corresponding candidate item；

The attribute data of the attribute data of the target user and the multiple candidate item is handled, target data is generated Collection, the target data set include the mark of the target user and the list of the first interaction node of corresponding target, the multiple The mark of each candidate item and the second interaction node list of corresponding target, first interaction node of target in candidate item List is used to indicate the interactive information of the target user and other users or article, and target the second interaction node list is used In the interactive information for indicating candidate item and other articles or user；

The target data set is inputted into scoring model, obtains marking of the target user to the multiple candidate item, In, the scoring model is obtained according to the attribute data of multiple users, the attribute data of multiple articles and marking data training , the attribute data of each user includes the mark of corresponding user, each object in the multiple article in the multiple user The attribute data of product includes the mark of corresponding article, the marking data include in the multiple user each user to described The marking of one or more articles in multiple articles；

Marking according to the target user to the multiple candidate item determines that target recommends article.

2. the method according to claim 1, wherein the attribute data of the target user further includes following data One of or it is a variety of: gender, height, weight, age, occupation, income, hobby, education landscape, the attribute of each candidate item Data further include one of following data or a variety of: brand, color, size, price, comment, taste, shelf-life, icon.

3. the method according to claim 1, wherein described by the attribute data of the target user and described more The attribute data of a candidate item is handled, and target data set is generated, comprising:

According to the mark of the target user, the corresponding target first of the mark of each user in pre-recorded multiple users In interaction node list, corresponding the first interaction node of the target list of the target user is determined, and according to each candidate item Mark, corresponding the second interaction node of the target list of the mark of each candidate item in pre-recorded multiple candidate items In, determine corresponding the second interaction node of the target list of each candidate item；

According to the mark of the target user, corresponding the first interaction node of the target list of the target user, each candidate The mark of product and corresponding the second interaction node of the target list of each candidate item generate target data set.

4. the method according to claim 1, wherein the scoring model includes feature learning model, feedback Practise model and neural network model；

Wherein, described that the target data set is inputted into scoring model, the target user is obtained to the multiple candidate item Marking, comprising:

The mark for the target user that the target data is concentrated and the mark of candidate item j input the feature learning model, Obtain the corresponding feature vector of the target user and the corresponding feature vector of the candidate item j, and by the target data The corresponding interaction of target second of corresponding target the first interaction node list of the target user and the candidate item j concentrated Node listing inputs the feedback learning model, obtains j pairs of the corresponding implicit feedback of the target user and the candidate item The implicit feedback answered, wherein the article j is any candidate item in the multiple candidate item；

The corresponding feature vector of the target user, the corresponding feature vector of the candidate item j, the target user is corresponding Implicit feedback and the corresponding implicit feedback of the candidate item j, input neural network model, obtain the target user to time Select the marking of article j.

5. according to the method described in claim 4, it is characterized in that, the target the first interaction node list includes multistage target First interaction node list, corresponding the second interaction node of the target list of each candidate item include the multistage interaction of target second section Point list, in multistage the first interaction node of target list the first interaction node list of odd-order target for indicate target user with The interactive information of article, the first interaction node list of even-order target is for indicating mesh in multistage the first interaction node of target list The interactive information of user and other users are marked, the second interaction node of odd-order target column in multistage the second interaction node of target list Table is for indicating candidate item and the interactive information of user, and even-order target second is handed in multistage the second interaction node of target list Mutual node listing is used to indicate the interactive information of candidate item Yu other articles；

Corresponding target the first interaction node list of the target user that the target data is concentrated and candidate item j Corresponding the second interaction node of target list, inputs the feedback learning model, and it is corresponding implicit anti-to obtain the target user Present implicit feedback corresponding with candidate item j, comprising:

Corresponding multistage target the first interaction node list of the target user that the target data is concentrated and candidate item j Corresponding the second interaction node of multistage target list, inputs the feedback learning model, it is corresponding hidden to obtain the target user Formula feeds back implicit feedback corresponding with candidate item j.

6. according to the method described in claim 4, it is characterized in that, the model parameter of the feedback learning model includes: described The weight of the feedback vector of each user in multiple users, the weight of the feedback vector of each article in the multiple article；

The mark and the list of the first interaction node of corresponding target, candidate for the target user that the target data is concentrated The mark of product j and the second interaction node list of corresponding target, input the feedback learning model, obtain the target user couple The implicit feedback and the corresponding implicit feedback of candidate item j answered.

7. the method according to claim 1, wherein it is described according to the target user to the multiple candidate The marking of product determines that target recommends article, comprising:

Marking according to the target user to the multiple candidate item determines that corresponding marking meets default recommendation condition Target recommends article.

8. the method according to the description of claim 7 is characterized in that it is described according to the target user to the multiple candidate The marking of product determines that corresponding marking meets the target recommendation article of default recommendation condition, comprising:

Marking according to the target user to the multiple candidate item determines the corresponding maximum preset number mesh of marking Mark recommends article；Alternatively,

Marking according to the target user to the multiple candidate item determines that corresponding marking is greater than preset fraction threshold value Target recommends article.

9. method according to claim 1-8, which is characterized in that the scoring model is trained by the following method It obtains:

The attribute data of the multiple user, the attribute data of the multiple article and the marking data are handled, Obtain training dataset, the training dataset includes the mark of each user and corresponding first interaction node list, each Articles one or more in the multiple article are beaten in the mark of article and corresponding second interaction node list, each user Point, the first interaction node list is used to indicate the interactive information of user and other users or article, the second interaction section Point list is used to indicate the interactive information of article and other articles or user；

According to the training dataset, scoring model is trained.

10. according to the method described in claim 9, it is characterized in that, the attribute data of each user is also in the multiple user Including it is following number breath one of or it is a variety of: gender, height, weight, age, occupation, income, hobby, education landscape, it is described more The attribute data of each article further includes one of following data or a variety of in a article: brand, size, price, is commented color By, taste, shelf-life, icon；The marking data further include one of following data or a variety of: the operating time currently makes With equipment, discount situation.

11. according to the method described in claim 9, it is characterized in that, the attribute data for obtaining the multiple user, described The attribute data of multiple articles and the marking data, comprising:

It obtains multiple marking to record, each marking record includes the category of the attribute data of user u, article i in multiple marking records Property data and user u to the marking data of article i, the user u is beat the article i the multiple user excessively In any user, the article i be multiple articles in any article；

It is described to the attribute data of the multiple user, the attribute data of the multiple article and the marking data at Reason, obtains training dataset, comprising:

Multiple marking record is handled, obtains training dataset, it includes user u that training data, which concentrates each training data, Mark and corresponding first interaction node list, the mark of article i and corresponding second interaction node list, user u are to article i Marking.

12. according to the method for claim 11, which is characterized in that the scoring model includes feature learning model, feedback Learning model and neural network model；

Wherein, described according to the training dataset, scoring model is trained, comprising:

The mark of the user u, the mark of the article i are inputted into the feature learning model, it is corresponding to obtain the user u Feature vector and the corresponding feature vector of the article i, and by the corresponding first interaction node list of the user u, the object The corresponding second interaction node list of product i inputs the feedback learning model, obtains the corresponding implicit feedback of the user u and institute State the corresponding implicit feedback of article i；

The corresponding feature vector of the user u and the corresponding feature vector of the article i, the user u is corresponding implicit anti- It presents implicit feedback corresponding with the article i and inputs the neural network model, obtain prediction score；

According to the marking to the article i of prediction score and the user u, to the feature learning model, described anti- Feedback learning model and the neural network model are adjusted, the scoring model after being trained.

13. according to the method for claim 12, which is characterized in that the corresponding first interaction node list packet of the user u Include multistage first interaction node list, the corresponding second interaction node list of the article i includes multistage second interaction node column Table, the model parameter of the feedback learning model include: multistage user feedback matrix and multistage article feedback matrix, wherein institute The order for stating the corresponding first interaction node list of user u is identical as the user feedback order of matrix number, and the article i is corresponding The second interaction node list order it is identical as the order of the article feedback matrix, it is odd in multistage first interaction node list Number rank the first interaction node list is used to indicate the interactive information of user and article, even-order in multistage first interaction node list First interaction node list is used to indicate the interactive information of user and other users, odd-order in multistage second interaction node list Second interaction node list is for indicating article and the interactive information of user, even-order second in multistage second interaction node list Interaction node list is used to indicate the interactive information of article Yu other articles；

It is described to input the corresponding first interaction node list of the user u, the corresponding second interaction node list of the article i The feedback learning model obtains the corresponding implicit feedback of the user u and the corresponding implicit feedback of the article i, comprising:

By the corresponding multistage first interaction node list of the user u, the corresponding multistage second interaction node list of the article i The feedback learning model is inputted, the corresponding implicit feedback of the user u and the corresponding implicit feedback of the article i are obtained.

14. according to the method for claim 12, which is characterized in that the model parameter of the feedback learning model includes: institute The weight of the feedback vector of each user in multiple users is stated, the weight of the feedback vector of each article in the multiple article；

By the mark of the user u and the mark and corresponding second interaction of corresponding first interaction node list, the article i Node listing inputs the feedback learning model, obtains the corresponding implicit feedback of the user u and the article i is corresponding implicit Feedback.

15. a kind of device for recommending article, which is characterized in that described device includes:

Module is obtained, for obtaining the attribute data of target user and the attribute data of multiple candidate items, the target user Attribute data include target user mark, the attribute data of each candidate item includes the mark of corresponding candidate item；

Generation module, for will be at the attribute data of the attribute data of the target user and the multiple candidate item Reason, generates target data set, and the target data set includes the mark and the interaction section of corresponding target first of the target user The mark of each candidate item and the second interaction node list of corresponding target, described in point list, the multiple candidate item Target the first interaction node list is used to indicate the interactive information of the target user and other users or article, the target the Two interaction node lists are used to indicate the interactive information of candidate item and other articles or user；

Scoring modules obtain the target user to the multiple candidate for the target data set to be inputted scoring model The marking of article, wherein the scoring model is according to the attribute data of multiple users, the attribute data of multiple articles and marking Data training obtains, and the attribute data of each user includes the mark of corresponding user in the multiple user, the multiple The attribute data of each article includes the mark of corresponding article in article, and the marking data include every in the multiple user Marking of one user to articles one or more in the multiple article；

Determining module determines that target recommends article for the marking according to the target user to the multiple candidate item.

16. device according to claim 15, which is characterized in that the attribute data of the target user further includes following number According to one of or it is a variety of: gender, height, weight, age, occupation, income, hobby, education landscape, the category of each candidate item Property data further include one of following data or a variety of: brand, color, size, price, comment, taste, shelf-life, icon.

17. device according to claim 15, which is characterized in that the generation module is used for:

18. device according to claim 15, which is characterized in that the scoring model includes feature learning model, feedback Learning model and neural network model；

Wherein, scoring modules are used for:

19. device according to claim 18, which is characterized in that the first interaction node of target list includes multistage mesh Mark the first interaction node list, corresponding the second interaction node of the target list of each candidate item includes the multistage interaction of target second Node listing, the first interaction node list of odd-order target is for indicating target user in multistage the first interaction node of target list With the interactive information of article, the first interaction node list of even-order target is for indicating in multistage the first interaction node of target list The interactive information of target user and other users, the second interaction node of odd-order target in multistage the second interaction node of target list List is for indicating candidate item and the interactive information of user, even-order target second in multistage the second interaction node of target list Interaction node list is used to indicate the interactive information of candidate item Yu other articles；

The scoring modules, are used for:

20. device according to claim 18, which is characterized in that the model parameter of the feedback learning model includes: institute The weight of the feedback vector of each user in multiple users is stated, the weight of the feedback vector of each article in the multiple article；

The scoring modules, are used for:

21. device according to claim 15, which is characterized in that the determining module is used for:

22. device according to claim 21, which is characterized in that the determining module is used for:

23. the described in any item devices of 5-22 according to claim 1, which is characterized in that the acquisition module is also used to:

The generation module, is also used to:

Described device further include:

Training module, for being trained to scoring model according to the training dataset.

24. device according to claim 23, which is characterized in that the attribute data of each user is also in the multiple user Including it is following number breath one of or it is a variety of: gender, height, weight, age, occupation, income, hobby, education landscape, it is described more The attribute data of each article further includes one of following data or a variety of in a article: brand, size, price, is commented color By, taste, shelf-life, icon；The marking data further include one of following data or a variety of: the operating time currently makes With equipment, discount situation.

25. device according to claim 23, which is characterized in that the acquisition module is used for:

The generation module, is used for:

26. device according to claim 25, which is characterized in that the scoring model includes feature learning model, feedback Learning model and neural network model；

Wherein, the training module, is used for:

27. device according to claim 26, which is characterized in that the corresponding first interaction node list packet of the user u Include multistage first interaction node list, the corresponding second interaction node list of the article i includes multistage second interaction node column Table, the model parameter of the feedback learning model include: multistage user feedback matrix and multistage article feedback matrix, wherein institute The order for stating the corresponding first interaction node list of user u is identical as the user feedback order of matrix number, and the article i is corresponding The second interaction node list order it is identical as the order of the article feedback matrix, it is odd in multistage first interaction node list Number rank the first interaction node list is used to indicate the interactive information of user and article, even-order in multistage first interaction node list First interaction node list is used to indicate the interactive information of user and other users, odd-order in multistage second interaction node list Second interaction node list is for indicating article and the interactive information of user, even-order second in multistage second interaction node list Interaction node list is used to indicate the interactive information of article Yu other articles；

The training module, is used for:

28. device according to claim 26, which is characterized in that the model parameter of the feedback learning model includes: institute The weight of the feedback vector of each user in multiple users is stated, the weight of the feedback vector of each article in the multiple article；

The training module, is used for:

29. a kind of equipment, which is characterized in that the equipment includes processor and memory, and processor is configured as executing storage The instruction stored in device；Processor executes instruction so that the equipment is realized as described in claim 1-14 any claim Method.

30. a kind of computer readable storage medium, which is characterized in that including instruction, when the computer readable storage medium exists When being run on computer, so that the computer executes method described in any claim in the claim 1-14.