CN113722608B - Collaborative filtering method and device based on association relation learning under guidance of iterative auxiliary information - Google Patents
Collaborative filtering method and device based on association relation learning under guidance of iterative auxiliary information Download PDFInfo
- Publication number
- CN113722608B CN113722608B CN202110825699.5A CN202110825699A CN113722608B CN 113722608 B CN113722608 B CN 113722608B CN 202110825699 A CN202110825699 A CN 202110825699A CN 113722608 B CN113722608 B CN 113722608B
- Authority
- CN
- China
- Prior art keywords
- user
- matrix
- heterogeneous
- users
- auxiliary information
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 37
- 238000001914 filtration Methods 0.000 title claims abstract description 13
- 239000011159 matrix material Substances 0.000 claims abstract description 85
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 14
- 210000005036 nerve Anatomy 0.000 claims abstract description 14
- 239000013598 vector Substances 0.000 claims description 16
- 238000010276 construction Methods 0.000 claims description 14
- 230000003993 interaction Effects 0.000 claims description 8
- 238000005065 mining Methods 0.000 claims description 7
- 238000004590 computer program Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 238000009825 accumulation Methods 0.000 claims description 5
- 238000012217 deletion Methods 0.000 claims description 2
- 230000037430 deletion Effects 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 4
- 238000012549 training Methods 0.000 description 19
- 238000012360 testing method Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000012804 iterative process Methods 0.000 description 4
- 238000007781 pre-processing Methods 0.000 description 3
- 230000006399 behavior Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000002360 explosive Substances 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 230000001771 impaired effect Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a collaborative filtering method and a collaborative filtering device based on association relation learning under the guidance of iterative auxiliary information, which integrate heterogeneous auxiliary information of a user and an article into a scoring record matrix; and recommending the articles for each user by using the generated preference prediction matrix. According to the invention, by distinguishing the importance of heterogeneous auxiliary information of different types of users and articles, the association relation among the heterogeneous auxiliary information of each type is mined by utilizing the nerve factor decomposition machine, and the effect of the heterogeneous auxiliary information of the users and the articles is fully exerted through iterative guidance, so that comprehensive understanding of the users and the articles is realized, and the accuracy of predicting the preference of the users to the articles is improved.
Description
Technical Field
The invention belongs to the technical field of software, and particularly relates to a collaborative filtering method and device based on association relation learning under the guidance of iterative auxiliary information.
Background
In recent years, with the rapid development of information technology, internet-based data has been shown to have explosive growth. In the face of the huge data volume, on one hand, people can acquire rich information resources through various channels, and enjoy convenience brought to life in the Internet era; on the other hand, to quickly and accurately obtain information required by a user from mass data becomes a difficult problem, and people inevitably waste time and energy on some irrelevant information. To help people find the desired information quickly, recommendation systems have emerged as an effective tool to filter and screen the information. The method can analyze collected user data (including the basic attributes of gender, occupation, hobbies and the like of a user), item data (including the basic attributes of category, appearance, use, price and the like of an item) and a user-item history interaction record, mine the implicit requirements of the user by means of the user portrait and the item portrait obtained by analysis, and then recommend the possibly interested item for the user by adopting different recommendation algorithms.
Collaborative filtering is one of the most widely used techniques in recommendation systems that model interactions between users and items. However, collaborative filtering presents serious data sparseness problems. For example, when a user only scores a few items, the performance of the recommendation system may drop dramatically. In order to alleviate the data sparseness problem, the recommendation system recommends by means of heterogeneous side information. Heterogeneous side information is typically stored in heterogeneous information networks, which are sources of additional information (e.g., age, gender, occupation, category, price, branding, etc.) for users and items. Since the heterogeneous side information can help the recommendation system to further understand the preference of the user and the attribute of the object, the recommendation based on the heterogeneous side information has recently attracted the interests of the scholars. For example, chinese patent (application number: CN202011624028.4, application publication number: CN 112784153A) first modeled secondary information features of users and attractions based on an attention mechanism; secondly, mapping heterogeneous semantic information of the user and the scenic spot, and learning and predicting the score of the user to the scenic spot through a knowledge representation translation mechanism; and finally, calculating the predictive scores of the users on the candidate scenic spots, sorting according to the scores, generating a recommended scenic spot list, and simultaneously explaining the scoring behaviors according to the auxiliary information of the users and the scenic spots. The method and the system integrate the feature attention of the auxiliary information of the scenic spot of the user with the heterogeneous semantic information of the scenic spot of the user to learn the interest preference of the user, explain the scoring behavior of the user to the scenic spot from the aspect of the auxiliary information of the user and the scenic spot, improve the interpretive of the recommendation of the scenic spot, and particularly provide great support in the aspect of the online recommendation prediction of the hot tourist attraction.
However, the existing recommendation method based on heterogeneous side information mainly has two disadvantages. First, since the effect of heterogeneous side information is inevitably weakened in the process of information extraction and utilization, the existing recommendation method based on heterogeneous side information cannot provide sufficient guidance for the user to predict the scoring of the item. Second, the existing methods cannot effectively utilize heterogeneous side information to understand users and items because correlations between various types of heterogeneous side information are ignored when learning representations of users and items. Thus, these methods fail to recommend the appropriate item to the user.
Disclosure of Invention
Aiming at the defects of the existing recommendation method, the collaborative filtering method and the collaborative filtering device based on the association relation learning under the guidance of the iterative auxiliary information are provided, and the importance of different types of heterogeneous auxiliary information is improved by the iterative heterogeneous auxiliary information guidance, the image construction based on the mutual relation learning and the learning by using the attention network.
The technical scheme adopted by the invention is as follows:
collaborative filtering method based on association relation learning under guidance of iterative auxiliary information
1) Construction of scoring record matrix R between m users and n articles 0 Wherein the score records matrix R 0 The method comprises a plurality of true score records and a plurality of missing score records;
2) Respectively mining association relations between users and heterogeneous auxiliary information in articles, and generating a user portrait matrix P users Article representation matrix Q items Calculating a preference prediction matrix R' between the user and the item;
3) Generating a preference prediction matrix based on the preference prediction matrix RAnd uses preference prediction matrix +.>Fill score record matrix R t-1 In (2) a deletion score record, generating a score record matrix R t Thereby integrating the preference prediction matrix R' into the score recording matrix R t Wherein t is the number of iterations;
4) When scoring record matrix R t-1 Each real score record and preference prediction matrix in the databaseWhen the corresponding prediction score record in (a) meets the setting, based on the trained preference prediction matrix +.>And recommending the articles to each user.
Further, the heterogeneous side information of the user includes: age, sex, and occupation.
Further, the heterogeneous side information of the article includes: category, price, and brand.
Further, a user portrait matrix P is generated by the steps of users Article representation matrix Q items ：
1) Using heterogeneous information network to obtain vector representation u of various heterogeneous auxiliary information in user and article j And i k ；
2) Representing u for vectors j And i k Learning the importance of heterogeneous auxiliary information of different types through an attention network to obtain the weight of each heterogeneous auxiliary information;
3) The weight and each vector are expressed as u j Or i k Performing dot multiplication operation to obtain weighted heterogeneous auxiliary information;
4) Inputting the weighted heterogeneous side information into a nerve factor decomposition machine to construct a user portrait p j Or the image q of the object k Thereby generating a user portrait matrix P users Or article representation matrix Q items 。
Further, the structure of the attention network includes: a multi-layer fully connected network.
Further, user portrayal p is constructed j The formula of (1) includes: wherein w is 0 Is global bias, w a Representation->Weight, v a And v b Row a and row b of the weight matrix v, respectively, < >>And->The a-th and b-th weighted heterogeneous side information of the user j are respectively represented, d represents the dimension of the heterogeneous side information, and the MLP is a fully connected network for learning Gao Jiete sign interactions between various types of heterogeneous side information of the user j and the article k; building an item representation q k The formula of (1) includes: /> Wherein w is g Representation->Weight, v g And v h Respectively representing the g th row and the h th row of the weight matrix v, ">And->Respectively represent the first of the articles kg and h weighted heterogeneous side information.
Further, a preference prediction matrix is generated by the steps of
1) Record the score of matrix R t-1 Inputting the multi-layer perceptron to generate a preference prediction matrix R' of the user on the articles t-1 ；
2) For preference prediction matrix R t-1 Element accumulation is carried out with the preference prediction matrix R' to generate the preference prediction matrix
Further, a scoring record matrix R is generated t The formula of (1) includes: r is R t ←β({p j } m ({q k } n ) T )+(1-β)MLP(R t-1 ) Wherein beta is a superparameter, p j User portrayal for the jth user, q k For the item representation of the kth item, MLP is a fully connected network used to learn high-level feature interactions between user j and various types of heterogeneous side information of item k, and ζ represents a data population process.
Further, item recommendation of q items is performed on p users, wherein p user sets e m user sets, q item sets e n item sets, by:
1) Respectively acquiring user IDs of p users and object IDs of q objects;
2) Predicting matrix from preferences using user IDs and item IDsSearching a predictive scoring value of the corresponding user on the article;
3) Q items are recommended to p users based on the predictive scoring values.
A storage medium having a computer program stored therein, wherein the computer program is arranged to perform the above method when run.
An electronic device comprising a memory and a processor, wherein the memory stores a program for performing the above-described method.
Compared with the prior art, the invention has the advantages that:
1. aiming at the problem that in the existing recommendation method based on heterogeneous auxiliary information, the effect of the heterogeneous auxiliary information is inevitably weakened in the process of information extraction and utilization, and the heterogeneous auxiliary information cannot provide sufficient guidance for scoring prediction of the user on the articles, the invention provides an image construction module based on interrelation learning. And the effective learning of the potential characteristics and semantic information of the heterogeneous auxiliary information of the user and the object is realized by means of the HIN2 vec. By distinguishing the importance of heterogeneous side information of different types of users and articles, the model is helped to accurately understand the users and the articles. And excavating the association relation between the heterogeneous auxiliary information of each type by utilizing a nerve factor decomposition machine so as to realize comprehensive understanding of users and articles, thereby improving the accuracy of preference prediction.
2. Aiming at the problem that the user and the article cannot be understood by effectively utilizing the heterogeneous auxiliary information when the user and the article are learned by ignoring the interrelationship among various types of heterogeneous auxiliary information in the conventional recommending method based on the heterogeneous auxiliary information, the invention provides an iterative heterogeneous auxiliary information guiding module which iteratively integrates the heterogeneous auxiliary information into a user-article scoring matrix and continuously retrieves the lost or weakened heterogeneous auxiliary information. The heterogeneous auxiliary information of the user and the article is fully exerted through iterative guidance, and the model is helped to improve the accuracy of article preference prediction of the user.
Drawings
Fig. 1 is a flow chart of event detection.
FIG. 2 is a diagram illustrating a scoring prediction architecture according to the present invention.
Fig. 3 is a schematic diagram of an iterative heterogeneous secondary information guiding module structure according to the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail with reference to the accompanying drawings.
The collaborative filtering method of the invention provides a complete scoring prediction framework, which mainly comprises a portrait construction module based on interrelation learning and an iterative heterogeneous auxiliary information guiding module: firstly, the iterative heterogeneous auxiliary information guiding module can guide the prediction of the object grading value by the user in an iterative way. Through the iterative coaching process, the model can continually incorporate heterogeneous side information into the user-item score record, and those that are lost or otherwise impaired will slowly integrate into the user-item score record. This iterative process does not terminate until the model predicts that the user's scoring value for the item is closest to the true scoring value. Thus, the role played by heterogeneous side information in predicting the user's scoring value for an item is greatly enhanced. Secondly, the portrait construction module based on interrelationship learning can capture interrelationships among various types of heterogeneous auxiliary information, and help the model to better understand users and articles. Because the nerve factor decomposition machine is good at capturing high-order characteristic interaction between various types of heterogeneous auxiliary information, the nerve factor decomposition machine is used for deeply knowing the user and the articles, so that the representations of the user and the articles can be comprehensively learned and the proper articles can be recommended to the user. In addition, in order to fully utilize heterogeneous side information, the method employs an attention network to learn the importance of different types of heterogeneous side information. Different types of heterogeneous auxiliary information can make different contributions to the learning of the characteristics of the user and the article, and different treatment of various heterogeneous auxiliary information is helpful for the accurate understanding of the user and the article by the model.
Through training the scoring prediction model, the portrait construction module based on interrelation learning firstly utilizes a heterogeneous information network embedded model (HIN 2 vec) to acquire heterogeneous side information of users and articles. The HIN2vec can map heterogeneous side information of users and articles into a low-dimensional vector space, can effectively learn potential characteristics and semantic information of Xi Yizhi side information, and has certain guiding significance for analysis of the heterogeneous side information. The attention network is then utilized to distinguish the importance of the heterogeneous secondary information of different types. Learning the importance of heterogeneous side information of different types can help the model to accurately understand users and items before mining the association between heterogeneous side information of each type. Because the nerve factor decomposition machine can learn the interrelationship between heterogeneous auxiliary information of different types of users and articles, the module adopts the nerve factor decomposition machine to construct portraits of the users and the articles so as to realize comprehensive understanding of the users and the articles.
At the same time, to enable heterogeneous side information to provide sufficient guidance for user preference prediction of items, the iterative heterogeneous side information guidance module attempts to iteratively integrate the heterogeneous side information into the user-item scoring matrix. Thus, such heterogeneous side information that has been lost or weakened may be continually integrated into the user-item scoring matrix. By predicting user preferences for items through iterative guidance, the contribution of heterogeneous side information can be effectively enhanced.
According to the design scheme provided by the invention, the collaborative filtering method based on the association relation learning under the guidance of the iterative auxiliary information comprises the following steps:
and step 1, data preprocessing. The historical data of users and articles are divided into a training set and a testing set. Wherein the training set includes a user's preference degree scoring for the item (scoring range of 1 to 5 points, higher score indicating the user is more interested in the item), heterogeneous side information of the user (e.g., age, gender, occupation, etc.) and heterogeneous side information of the item (e.g., category, price, brand, etc.). The test set includes the user and the item to be scored.
And 2, constructing a scoring prediction framework, wherein the framework consists of a portrait construction module based on interrelation learning and an iterative heterogeneous auxiliary information guiding module.
The portrait construction module based on the mutual relation learning comprises the acquisition of the auxiliary information of the user and the article, the importance learning of the heterogeneous auxiliary information of different types and the mining of the association relation between the heterogeneous auxiliary information of different types. The heterogeneous side information of the user and the article is obtained mainly through a heterogeneous information network embedded model (HIN 2 vec). The HIN2vec can map heterogeneous auxiliary information of users and articles into a low-dimensional vector space, can effectively learn potential characteristics and semantic information of Xi Yizhi auxiliary information, and has certain guiding significance for analysis of the heterogeneous auxiliary information. The attention network can realize the importance learning of heterogeneous auxiliary information of different types, and the learning of the importance of the heterogeneous auxiliary information of different types can help the model to accurately understand users and articles. The mining of the association relation between the heterogeneous side information of each type is realized by a nerve factor decomposition machine. The nerve factor decomposition machine can capture the high-order nonlinear relation between different types of heterogeneous auxiliary information of the user and the object, so that comprehensive understanding of the user and the object is realized.
The iterative heterogeneous side information guidance module attempts to iteratively integrate the heterogeneous side information into the user-item scoring matrix so that those heterogeneous side information that have been lost or weakened can be constantly retrieved. The preference of the user on the articles is predicted through the iterative guidance, so that the effect exerted by the heterogeneous auxiliary information is effectively enhanced, and sufficient guidance is provided for the preference prediction of the user on the articles.
And step 3, training a scoring prediction model.
Firstly, scoring records of users and articles in a training set and side information of the users and the articles are stored in a heterogeneous information network, and then the heterogeneous information network is input into a portrait construction module based on interrelation learning. After the heterogeneous side information of the user and the object is obtained, the importance of the heterogeneous side information of different types is learned through the attention network, and the weight of each heterogeneous side information is obtained. And performing dot multiplication operation on the weight and each heterogeneous auxiliary information to obtain weighted heterogeneous auxiliary information. The weighted heterogeneous side information is input into a nerve factor decomposition machine to excavate the association relation between various types of heterogeneous side information, so that the portrait of the user and the article is constructed. By means of portraits of the user and the articles, preference prediction of the user on the articles is achieved through matrix multiplication. Meanwhile, scoring records of users and objects in the training set are input into the multi-layer perceptron to predict the preference of the users for the objects. The two preference predictors are integrated together by means of element accumulation. The integrated preference predictors are considered as true values to continuously populate missing data in the user-item scoring records, thereby integrating heterogeneous side information of the user and item into the user and item scoring records. And predicting the preference of the user on the article again by using the filled user-article scoring data until the score predicted by the model is closest to the real score, ending the whole iterative process, and fully playing the guiding role of the heterogeneous auxiliary information. Calculating an error between the predicted value and the true value, updating network parameters, and training to obtain an optimal model;
and 4, recommending the articles to the user.
And 3, training to obtain an optimal model, obtaining optimal user preference prediction for the articles by utilizing scoring records of users and articles in a training set and heterogeneous side information of the users and the articles, searching corresponding user preference prediction values for the articles from preference prediction data of the users for the articles according to the users in a testing set and the articles to be scored, and recommending the articles with higher prediction values to the users.
In the step 2, the portrait construction module based on the mutual relation learning includes three parts of obtaining the auxiliary information of the user and the article, learning the importance of the heterogeneous auxiliary information of different types and mining the association relation between the heterogeneous auxiliary information of different types.
In step 2, the heterogeneous secondary information of the user and the article is obtained through a heterogeneous information network embedded model (HIN 2 vec). The HIN2vec can map heterogeneous side information of users and articles into a low-dimensional vector space, can effectively learn potential characteristics and semantic information of the heterogeneous side information, and has certain guiding significance for analysis of the heterogeneous side information.
In step 2, the attention network for learning the importance of heterogeneous secondary information of different types of users and articles is a multi-layered fully connected network.
In the step 2, the association relation between the heterogeneous sub-information of each type is mined by the nerve factor decomposition machine. The nerve factor decomposition machine can capture high-order nonlinear relations between different types of heterogeneous side information of users and articles.
In the step 2, the importance learning formula of the heterogeneous auxiliary information of the user and the article is as follows:
E 1 ＝σ 1 (W 1 U T +b 1 )
E 2 ＝σ 2 (W 2 E 1 +b 2 )
…
E L ＝σ L (W L E L-1 +b L )
and
F 1 ＝σ 1 (W 1 I T +b 1 )
F 2 ＝σ 2 (W 2 F 1 +b 2 )
…
F L ＝σ L (W L F L-1 +b L )
wherein U and I represent heterogeneous side information of the user and the article, respectively. W (W) L ，b L Sum sigma L The weight matrix, the bias vector and the activation function of the L-th full connection layer are respectively represented. E and F represent the attention weight matrix of the heterogeneous side information of the user and the item, respectively. d represents the dimension of the heterogeneous side information.
In the step 2, the mining formula of the association relationship between the user and the heterogeneous auxiliary information of the various types of the articles is as follows:
and
wherein w is 0 Is global bias, w a And w g Respectively representAnd->Is a weight of (a). v a And v g Respectively representing the a-th row and g-th row of the weight matrix v for +.>And->Is used for vector conversion of (a). Wherein (1)>A-th weighted heterogeneous side information indicating user j, < ->The g-th weighted heterogeneous side information for item k is represented. In addition, MLP is a fully connected network used to learn high-order feature interactions between various types of heterogeneous side information for user j and item k.
In the above step 2, the formula for integrating the heterogeneous side information into the user-item scoring matrix is:
R t ←β({p j } m ({q k } n ) T )+(1-β)MLP(R t-1 )
wherein { p } j } m Sum { q k } n Respectively representing a set of m user representations and n item representations. R is R t Is the user-item scoring matrix generated after the t-th iteration. Beta is a hyper-parameter used to balance heterogeneous side information and user-item scoring matrix weights. And ζ, represents a data population process.
The invention will be further described by taking the example of scoring n items by m users.
FIG. 1 is a flow chart of event detection, including four parts of data preprocessing, constructing a scoring prediction architecture, training a scoring prediction model, and recommending items to a user.
And step 1, data preprocessing. The historical data of users and articles are divided into a training set and a testing set. Wherein the training set includes a user's preference degree scoring for the item (scoring range of 1 to 5 points, higher score indicating the user is more interested in the item), heterogeneous side information of the user (e.g., age, gender, occupation, etc.) and heterogeneous side information of the item (e.g., category, price, brand, etc.). The test set includes the user and the item to be scored.
And 2, constructing a scoring prediction framework. FIG. 2 is a schematic diagram of a scoring prediction architecture according to the present invention. The structure comprises a portrait construction module based on mutual relation learning and an iterative heterogeneous side information guiding module. In a portrait construction module based on interrelation learning, firstly, vector representations u of m users and n articles are obtained from a heterogeneous information network through a heterogeneous information network embedded model HIN2vec j And i k . And then learning by using the attention network to obtain the importance of heterogeneous auxiliary information of different types of users and articles. Then, the product operation is carried out on the weight and the vector of the user and the object to obtain the weighted heterogeneous auxiliary information of the user and the objectAnd->Next, we input the weighted heterogeneous side information into a neurofactorizer to construct a representation p of the user and the item j And q k . Portrayal matrix P with m users and n objects users ＝{p j } m And Q items ＝{q k } n Obtaining a user preference prediction matrix R' =p of the article by matrix multiplication users Q items T . Meanwhile, the scoring records R of the users and the articles in the training set are input into a multi-layer perceptron to obtain preference prediction R of the users on the articles. Integrating the two preference prediction results together in an element accumulation mode to obtain a preliminary user-article scoring prediction matrix +.>Will->The preference predictors of (c) are seen as true values to continuously populate the missing data in R, thereby integrating heterogeneous side information of users and items into R. By means of filled->The preference prediction of the user on the item is performed again, and the whole iterative process is not ended until the score predicted by the model is closest to the true score. FIG. 3 is a schematic diagram of an iterative heterogeneous secondary information guide module structure according to the method of the present invention, wherein R represents a scoring record of users and items, U and I represent vector sets of m users and n items, respectively, and P users And Q items Representing a portrait set of m users and n items respectively, MLP representing a multi-layer perceptron, R 'and R' representing a user preference prediction matrix for items, < >>Representing a user-item score prediction matrix;
and step 3, training a scoring prediction model. Firstly, storing scoring records of users and articles in a training set and auxiliary information of the users and the articles into a heterogeneous information network, inputting the heterogeneous information network into a portrait construction module based on interrelation learning, and obtaining vector representations u of m users and n articles j And i k . Obtaining the weight of heterogeneous auxiliary information of the user and the object through the attention network learning, and performing product operation on the weight and the vector of the user and the object to obtain the weighted heterogeneous auxiliary information of the user and the objectAnd->Next, the weighted heterogeneous side information is input into a nerve factor decomposition machine to construct an image matrix P of the user and the object users ＝{p j } m And Q items ＝{q k } n And obtaining a preference prediction matrix R' of the user on the articles through matrix multiplication. Meanwhile, the scoring records R of the users and the articles in the training set are input into the multi-layer perceptron to obtain preference prediction R of the users on the articles. Integrating the two preference prediction results together in an element accumulation mode to obtain a preliminary user-article scoring prediction matrix +.>Will->The preference predictors of (c) are seen as true values to continuously populate the missing data in R, thereby integrating the heterogeneous side information of the user and the item into R. And predicting the preference of the user on the article again by using the filled R until the score predicted by the model is closest to the true score, and ending the whole iterative process. Updating parameters of the network by calculating error between the predicted scoring value and the true scoring value, and training to obtain an optimal model;
step 4, obtaining a scoring prediction matrix of the optimal user-article by utilizing scoring records of the user and the article in the training set and heterogeneous side information of the user and the article on the basis of training in the step 3 to obtain the optimal modelFrom the user-item scoring prediction matrix +.>And searching a corresponding predictive scoring value of the user on the item, and recommending the item with the higher predictive scoring value to the user.
The above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and those skilled in the art may modify or substitute the technical solution of the present invention without departing from the spirit and scope of the present invention, and the protection scope of the present invention shall be defined by the claims.
Claims (6)
1. A collaborative filtering method based on association relation learning under the guidance of iterative auxiliary information,
1) Construction of scoring record matrix R between m users and n articles 0 Wherein the score records matrix R 0 The method comprises a plurality of real scoring records and a plurality of missing scoring records;
2) Respectively mining association relations between users and heterogeneous auxiliary information in articles, and generating a user portrait matrix P users Article representation matrix Q items Calculating a preference prediction matrix R' between the user and the item; wherein a user portrayal matrix P is generated users Article representation matrix Q items Comprises the following steps:
using heterogeneous information network to obtain vector representation u of various heterogeneous auxiliary information in user and article j And i k ；
Representing u for vectors j And i k Learning the importance of heterogeneous auxiliary information of different types through an attention network to obtain the weight of each heterogeneous auxiliary information;
the weight and each vector are expressed as u j Or i k Performing dot multiplication operation to obtain weighted heterogeneous auxiliary information;
inputting the weighted heterogeneous side information into a nerve factor decomposition machine to construct a user portrait p j Or the image q of the object k The method comprises the steps of carrying out a first treatment on the surface of the Wherein the user portraitsArticle portrait-> w 0 Is global bias, w a Representation->Weight, v a And v b Row a and row b of the weight matrix v, respectively, < >>And->The a-th and b-th weighted heterogeneous side information respectively representing user j, d representing the dimension of the heterogeneous side information, MLP being a fully connected network for learning high-order characteristic interactions between various types of heterogeneous side information of user j and item k, w g Representation->Weight, v g And v h Respectively representing the g th row and the h th row of the weight matrix v, ">And->The g-th and h-th weighted heterogeneous side information of the article k are respectively represented;
based on user portrayal p j And article representation q k Generating a user portrait matrix P users Or article representation matrix Q items ；
3) Generating a preference prediction matrix based on the preference prediction matrix RAnd uses preference prediction matrix +.>Fill score record matrix R t-1 In (2) a deletion score record, generating a score record matrix R t Thereby integrating the preference prediction matrix R' into the score recording matrix R t Wherein t is the number of iterations, generating a preference prediction matrix based on the preference prediction matrix R'>Comprising the following steps:
record the score of matrix R t-1 Inputting the multi-layer perceptron to generate a preference prediction matrix R' of the user on the articles t-1 ；
For preference prediction matrix R t-1 Element accumulation is carried out with the preference prediction matrix R' to generate the preference prediction matrix
4) When scoring record matrix R t-1 Each real score record and preference prediction matrix in the databaseWhen the corresponding prediction score record in (a) meets the setting, based on the trained preference prediction matrix +.>Recommending articles to each user; wherein the prediction matrix is +.>Recommending items to each user, comprising:
respectively acquiring user IDs of p users and object IDs of q objects; wherein, p user sets epsilon m user sets, q article sets epsilon n article sets;
predicting matrix from preferences using user IDs and item IDsSearching out the corresponding predicted grading value of the user to the article;
q items are recommended to p users based on the predictive scoring values.
2. The method of claim 1, wherein the heterogeneous side information of the user comprises: age, sex, and occupation; the heterogeneous side information of the article includes: category, price, and brand.
3. The method of claim 1, wherein the structure of the attention network comprises: a multi-layer fully connected network.
4. The method of claim 1, wherein a scoring record matrix R is generated t The formula of (1) includes: r is R t ←β({p j } m ({q k } n ) T )+(1-β)MLP(R t-1 ) Wherein beta is a superparameter, p j User portrayal for the jth user, q k For the item representation of the kth item, MLP is a fully connected network used to learn high-level feature interactions between user j and various types of heterogeneous side information of item k, and ζ represents a data population process.
5. A storage medium having a computer program stored therein, wherein the computer program is arranged to perform the method of any of claims 1-4 when run.
6. An electronic device comprising a memory, in which a computer program is stored, and a processor arranged to run the computer program to perform the method of any of claims 1-4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110825699.5A CN113722608B (en) | 2021-07-21 | 2021-07-21 | Collaborative filtering method and device based on association relation learning under guidance of iterative auxiliary information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110825699.5A CN113722608B (en) | 2021-07-21 | 2021-07-21 | Collaborative filtering method and device based on association relation learning under guidance of iterative auxiliary information |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113722608A CN113722608A (en) | 2021-11-30 |
| CN113722608B true CN113722608B (en) | 2024-03-22 |
Family
ID=78673705
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110825699.5A Active CN113722608B (en) | 2021-07-21 | 2021-07-21 | Collaborative filtering method and device based on association relation learning under guidance of iterative auxiliary information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113722608B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106682114A (en) * | 2016-12-07 | 2017-05-17 | 广东工业大学 | Personalized recommending method fused with user trust relationships and comment information |
| KR20180093741A (en) * | 2017-02-14 | 2018-08-22 | 주식회사 아이디어랩스 | Method for collaboratively filtering information in use of personalized regression with auxiliary information to predict preference given by user of item to the item and computing appatarus apparatus using the same |
| CN109271582A (en) * | 2018-08-20 | 2019-01-25 | 东南大学 | It is a kind of based on the recommendation method for personalized information with attribute member path |
| CN110020883A (en) * | 2018-12-12 | 2019-07-16 | 阿里巴巴集团控股有限公司 | The method and device that unknown scoring in a kind of pair of rating matrix is predicted |
| CN110245285A (en) * | 2019-04-30 | 2019-09-17 | 中国科学院信息工程研究所 | A kind of personalized recommendation method based on Heterogeneous Information network |
| CN110263256A (en) * | 2019-06-21 | 2019-09-20 | 西安电子科技大学 | Personalized recommendation method based on multi-modal heterogeneous information |
| CN111061961A (en) * | 2019-11-19 | 2020-04-24 | 江西财经大学 | Multi-feature-fused matrix decomposition interest point recommendation method and implementation system thereof |
| CA3118083A1 (en) * | 2018-11-01 | 2020-05-07 | Netflix, Inc. | Techniques for recommending items to users |
-
2021
- 2021-07-21 CN CN202110825699.5A patent/CN113722608B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106682114A (en) * | 2016-12-07 | 2017-05-17 | 广东工业大学 | Personalized recommending method fused with user trust relationships and comment information |
| KR20180093741A (en) * | 2017-02-14 | 2018-08-22 | 주식회사 아이디어랩스 | Method for collaboratively filtering information in use of personalized regression with auxiliary information to predict preference given by user of item to the item and computing appatarus apparatus using the same |
| CN109271582A (en) * | 2018-08-20 | 2019-01-25 | 东南大学 | It is a kind of based on the recommendation method for personalized information with attribute member path |
| CA3118083A1 (en) * | 2018-11-01 | 2020-05-07 | Netflix, Inc. | Techniques for recommending items to users |
| CN110020883A (en) * | 2018-12-12 | 2019-07-16 | 阿里巴巴集团控股有限公司 | The method and device that unknown scoring in a kind of pair of rating matrix is predicted |
| CN110245285A (en) * | 2019-04-30 | 2019-09-17 | 中国科学院信息工程研究所 | A kind of personalized recommendation method based on Heterogeneous Information network |
| CN110263256A (en) * | 2019-06-21 | 2019-09-20 | 西安电子科技大学 | Personalized recommendation method based on multi-modal heterogeneous information |
| CN111061961A (en) * | 2019-11-19 | 2020-04-24 | 江西财经大学 | Multi-feature-fused matrix decomposition interest point recommendation method and implementation system thereof |
Non-Patent Citations (6)
| Title |
|---|
| Collaborative Filtering with Entity Similarity Regularization in Heterogeneous Information Networks;Xiao Yu;Citeseer;1-6 * |
| Deep Collaborative Filtering with Multi-Aspect Information in Heterogeneous Networks;Chuan Shi等;IEEE;1413-1425 * |
| 基于异质信息网络的深度学习推荐算法研究;温玉娇;中国优秀硕士;I138-1318 * |
| 基于用户与物品信息挖掘的矩阵分解推荐算法研究;王东;中国优秀硕士论文全文数据库 信息科技辑;I138-3530 * |
| 融合用户动态标签和信任关系的协同过滤算法;吴鸿玲;程耕国;;计算机工程与应用(19);43-49 * |
| 融合社交网络特征的协同过滤推荐算法;郭宁宁;王宝亮;侯永宏;常鹏;;计算机科学与探索(02);208-217 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113722608A (en) | 2021-11-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Huo et al. | Knowledge modeling via contextualized representations for LSTM-based personalized exercise recommendation | |
| CN111222332B (en) | Commodity recommendation method combining attention network and user emotion | |
| CN109376222B (en) | Question-answer matching degree calculation method, question-answer automatic matching method and device | |
| Pan et al. | BIM log mining: Learning and predicting design commands | |
| CN110533974B (en) | Intelligent volume assembling method and system and computer readable storage medium | |
| CN108921221A (en) | Generation method, device, equipment and the storage medium of user characteristics | |
| CN110377814A (en) | Topic recommended method, device and medium | |
| CN111242310B (en) | Feature validity evaluation method and device, electronic equipment and storage medium | |
| CN110991645A (en) | Self-adaptive learning method, system and storage medium based on knowledge model | |
| US20200401948A1 (en) | Data sampling for model exploration | |
| CN110245285A (en) | A kind of personalized recommendation method based on Heterogeneous Information network | |
| Lian et al. | Mutual reinforcement of academic performance prediction and library book recommendation | |
| CN111523738A (en) | System and method for predicting learning effect based on user online learning behavior pattern | |
| CN115329746A (en) | Event extraction method, device and equipment | |
| Doko et al. | Sequential Pattern Mining Model to Identify the Most Important or Difficult Learning Topics via Mobile Technologies. | |
| Segou et al. | The use of stochastic optimization in ground motion prediction | |
| CN113722608B (en) | Collaborative filtering method and device based on association relation learning under guidance of iterative auxiliary information | |
| CN111401525A (en) | Adaptive learning system and method based on deep learning | |
| Shujrah et al. | Measurement of e-learners’ level of interest in online course using support vector machine | |
| Kaensar et al. | Analysis and Prediction of Student Performance Based on Moodle Log Data using Machine Learning Techniques | |
| Baralis et al. | Highlighter: Automatic highlighting of electronic learning documents | |
| Zhao et al. | State and tendency: an empirical study of deep learning question&answer topics on Stack Overflow | |
| Han | Research on personalized recommendation method of educational resources based on learner behavior analysis | |
| CN112559905A (en) | Conversation recommendation method based on dual-mode attention mechanism and social similarity | |
| JP2021536050A (en) | Reducing instances of data inclusion associated with hindsight bias in data training sets for machine learning systems |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |