CN110825978B

CN110825978B - Multitask collaborative filtering method based on neighbor user feature sharing

Info

Publication number: CN110825978B
Application number: CN201911026575.XA
Authority: CN
Inventors: 卢暾; 石宛露; 顾宁; 李东胜; 张鹏
Original assignee: Fudan University
Current assignee: Fudan University
Priority date: 2019-10-26
Filing date: 2019-10-26
Publication date: 2023-04-07
Anticipated expiration: 2039-10-26
Also published as: CN110825978A

Abstract

The invention belongs to the technical field of user interest preference prediction, and particularly relates to a multitask collaborative filtering method based on neighbor user feature sharing. The invention comprises two stages: a K neighbor user group generation stage, which generates K most similar neighbor users of the users as much as possible; in the interest prediction stage, the possible scoring condition of the current active user for the unevaluated article is predicted by means of the K neighbor user; based on the cognition that similar users have similar interests, the K neighbor user group of the users is determined according to the Pearson similarity with importance by taking the advantage of a multi-task feature learning method in processing sparse data as reference, local sharing is performed on the user feature space to different degrees in the range of the neighbor users, and finally feature parameters of collaborative filtering for interest prediction are obtained. According to the method, the feedback data of the similar users, namely the interest preferences of the similar users, are adopted to supplement the interest characteristics of the current user, so that the problem of data sparsity is alleviated, and the accuracy of recommendation prediction is improved.

Description

Multitask collaborative filtering method based on neighbor user feature sharing

Technical Field

The invention belongs to the technical field of user interest preference prediction, and particularly relates to a multitask collaborative filtering method based on neighbor user feature sharing.

Background

The data sparseness problem has been a difficult problem that recommended researchers address. With the constant popularity of the internet and the mobile internet, on one hand, the magnitude difference between the number of various network commodities provided for the increasing network users and the number of feedback evaluation information experienced by the users is further enlarged, so that the recommendation model can only depict favorite features which may deviate from the real interests of the users under the condition of insufficient data; on the other hand, the experience requirements of the network users for various recommendations are more and more biased to accurate personalization, and higher-accuracy user interest prediction capability support is needed. The contradiction between the data sparseness and the accurate recommendation provides more challenges and requirements for the recommendation model. Therefore, the invention provides a multitask collaborative filtering method based on feature sharing among adjacent users, which takes the general similar interest preference among the similar users as a basic cognition and footing point, and shares the interest features (directly embodied as shared feedback data) of the adjacent users in the K adjacent range of the active user in a multitask feature learning mode, thereby alleviating the problem of data sparsity of a single user per se, achieving the purpose of improving the prediction accuracy and finally improving the recommendation quality.

Disclosure of Invention

The invention aims to provide a multitask collaborative filtering method based on neighbor user feature sharing so as to alleviate and overcome the contradiction between data sparseness existing in the recommendation field for a long time and high prediction accuracy required by the recommendation user to be really interested in.

The multitask collaborative filtering method provided by the invention is based on the homogeneity phenomenon of similar users in a real environment and the advantages of multitask feature learning in the aspect of processing data sparsity, aims at the problem of data sparsity, and takes the feature sharing of adjacent users as a core. The method comprises two stages: a K neighbor user group generation stage, which is used for generating K most similar neighbor users of the users as much as possible; and an interest prediction stage, which is used for predicting the possible scoring condition of the current active user for the unevaluated goods by means of the K neighbor users.

K adjacent user group generation stage

The basic method is that based on the cognition that similar users have similar interest preferences, a K neighbor user group of the current active user is searched, the optimized inter-user similarity is taken as a weight to share a user characteristic space in the range of the neighbor user group, so that more similar users make greater contribution to the interest preference characterization of the active user, the inaccuracy of sparse user feedback on the interest characterization of the users is alleviated through the method, the accuracy of recommendation prediction is improved, and the recommendation quality is improved.

Generating similar users by adopting the Pearson similarity with the importance weight strategy, fully considering the deviation of the sparse common term to the Pearson similarity to measure the similarity of the users according to the analysis of sparse data, introducing the importance strategy, and only considering the first K users which are directly similar.

The basic footholds and the starting points of the invention are the cognition of similar interest preferences for similar users, so the invention is very important for finding similar users of active users, and the accuracy of user interest prediction can be directly influenced. Based on the consideration of the similar measurement effect, the judgment of the similar user adopts the common Pearson similarity (PC):

as a basis. Meanwhile, the invention considers the pseudo-similar situation which may exist in the real scene of sparse data. For pseudo-similarity, the following definitions are given:

definition 1: pseudo-similarity means that users have extremely rare feedback data to the same article but the rare common feedback data are accidentally close to each other, resulting in extremely high pearson similarity between users.

In the case of sparse data, even if the feedback of the common intersection between users is very similar, it cannot be determined whether the users are really similar. If the pearson similarity shown in (equation 1) is used, it is likely that the pseudo-similar user is determined as the first K users most similar to the active user, and the truly similar user is pushed behind. In response to this problem, the present invention introduces other importance weighting strategies that have proven feasible through research work, namely:

where u and j represent any two users, I _uj Representing a common feedback item set between two users, s represents any item in the common feedback item set, | I _uj | represents the number of common feedback items, γ represents a threshold value for the number of common feedback items among users set for all users, and r _us And r _js Respectively representing the scores of the common items by the user u and the user j,

and &>

Respectively representing the average scores of the two users.

In addition, the pearson similarity can be evaluated similarly from the positive and negative angles, but the application scenario of the invention does not need to consider users with negative correlation, so that users with negative correlation and users with zero correlation can be filtered after the similarity between the users is calculated. And aiming at the users with positive correlation, only the first K users with high similarity are screened as a neighbor user group, and the users with low similarity are filtered out.

Specifically, the main steps of the generation of the K-neighbor user group are as follows:

step 1: obtaining a feedback article set of each user according to the user feedback data;

and 2, step: and traversing other residual users for each user, and judging whether the feedback article sets between the other residual users have intersection. For the condition that intersection exists, calculating the similarity of the two according to the formula 2, if the similarity is greater than 0, recording, otherwise, setting 0; for the case that no intersection exists, directly setting 0;

and step 3: and (3) according to the similarity of each user and other users obtained in the step (2), taking the top K users with the similarity ranked from high to low to form a K neighbor user group of the current user.

(II) user interest prediction stage

The basic method is that a collaborative filtering algorithm based on multi-task feature learning is provided, for interest preference portrayal of a current active user, according to feedback data of the active user and weighted reflection of interests of similar users, an interest feature space of the current active user is obtained through user feature sharing among adjacent users, and the problem of low prediction accuracy caused by data sparseness is solved.

Building user interest prediction model

For a classical SVD-based collaborative filtering algorithm, each user can train through historical feedback records to obtain a feature vector of the user, and the feature vector is the depiction of the user interest features. In a traditional single-task scene, the prediction accuracy and the feedback habit of a user are considered, and the scoring deviation of the user and the scoring deviation of an article are considered when the scoring value of the user is predicted, namely, an improved SVD algorithm is formed as follows:

where u represents an arbitrary user, i represents an arbitrary item,

the score prediction value of the user u to the user i, mu is the average score of the user u, b _u Is the score deviation of user u, b _i Is the score deviation, p, of item i _u Feature vector, q, representing user u _i Representing the feature vector of item i.

The invention migrates the improved SVD algorithm in the single task scene to the multi-task scene by virtue of the superiority of the multi-task characteristic learning in processing sparse data, and treats the matrix decomposition task of each user as an independent task like some research works at the present stage, and simultaneously considers that only the matrix decomposition tasks among similar users are relevant. Since similar users are users who show similar interests in the same field, it can be considered that user feature vectors obtained by matrix decomposition between neighboring users also have similarity to some extent.

The core idea of the classical multi-task feature learning model is to share features between related tasks, in other words, the features of each task are composed of the weighted sum of a set of common features, while the common features between similar tasks have an intersection. Therefore, by means of the core idea, the feature vectors of the users are shared in the range of the K neighbor users, and the feature vector finally used for prediction of each user is the weighted sum of the features of the current user and the neighbor users. Considering that the importance degree of the user feature vector of the active user and the user feature vector of the adjacent user to the description of the interest of the active user is inconsistent, the similarity between the users is used as the embodiment of feature contribution and is used as a weight parameter for feature weighting. Finally, the scoring prediction formula for the current active user is as follows:

wherein N (u) represents K neighbor user set of user u, j represents any neighbor user, S _uj Representing the similarity value between user u and the neighboring user j. Comparing (equation 4) with (equation 3) clearly shows that: when the possible scoring condition of the user is predicted, the method does not rely on the feature vector of the current active user any more, but takes the feature vector of the adjacent user as feature characterization supplement of the current user to a certain extent under the condition of data sparsity. This is beneficial to reduce the biased impact of sparse feedback data on user interest preference prediction.

Without loss of generality, when the user interest prediction model is constructed according to the formula 4, a regular penalty term is introduced to prevent overfitting and underfitting. Therefore, the loss function of the algorithm at this stage is:

wherein R denotes the observed score data set, R _uj Representing the observed true score of user u on item i, and λ represents the regularization coefficient, whichThe remaining parameters have the same meanings as above.

The optimal parameters are obtained in a random gradient descent manner by minimizing the loss function (equation 5), and each gradient update equation is as follows:

/>

specifically, the main steps of user interest prediction are as follows:

training phase

Step 1: randomly extracting 20% of user data sets as training sets, and taking the rest as test sets, and internally disordering the training sets and the test sets to prevent data skew. The generalization capability can be improved by randomly carrying out a plurality of times;

step 2: and (3) training parameters by adopting a random gradient descent method, updating the formula according to the formula (6-10), and iterating until the algorithm converges.

Prediction phase

The prediction score value is calculated according to (formula 4) using the corresponding user feature vector and the corresponding item feature vector.

The invention has the beneficial effects that:

aiming at the problem of data sparseness faced by the actual recommendation field, the invention considers the possibility of homogeneity of interests among similar users in the same field on the basis of the existing model and algorithm, introduces multi-task feature learning, and provides a multi-task collaborative filtering method based on neighbor user feature sharing, which is used for alleviating the contradiction between serious data sparseness existing in user feedback data and recommendation quality of users, improving the accuracy of recommendation prediction and finally enhancing the user experience of article recommendation.

Drawings

FIG. 1 is a core schematic diagram of the method of the present invention.

Fig. 2 is a matrix formed by user similarities according to the present invention.

Detailed Description

According to the method, a comparison experiment is carried out on a broad bean movie data set (an experiment set consisting of 3000 users with the data density of 0.261% obtained after data cleaning) captured in 2016 from the end of 8 months to the beginning of 9 months. Comparative experiments used the Root Mean Square Error (RMSE):

as an evaluation standard for measuring the prediction capability and accuracy of the model, the method is also one of the common standard indexes of most research works in evaluating the prediction accuracy of the recommended model. The BiasedMF model is selected as the experimental control model, and the BiasedMF model and the method are mainly considered only by means of sparse scoring records of the user. The pseudo code for a specific implementation of the algorithm of the present invention is shown in the appendix. The complexity of the training algorithm is

N is R ^* Based on the data amount of (4), is greater than or equal to>

Is the number of neighbors of the user. Under normal conditions, the algorithm may converge after hundreds of iterations, and->

Although the range of (A) varies according to the data set, most of the algorithms have relatively good effect between 20 and 50, so that the algorithm has the advantages ofThe training complexity is acceptable.

According to some of the previous experiments, the present invention used the following parameters in the comparative experiments:

under the condition that the variation range of the characteristic dimension k is {10,20,30,40 and 50}, the finally obtained results are visible through 5 groups of random training sets and test set experiments, when the characteristic dimension is less than 30, the prediction accuracy of the method has a better prediction effect than that of a BiasedMF model, but as the characteristic dimension is greater than 30, the scoring prediction accuracy of the method is close to that of the BiasedMF model. This is caused by inaccuracy of the neighbor user set sharing the user feature vector obtained under the experimental set. The feedback information of the user to the common scoring movie is crucial to the generation of the neighboring user set. The number of the movies scored by the user together in the experimental data set of the comparison experiment is rare (mostly less than 10), and thus the generated K-neighbor user set cannot guarantee that the users are really most similar. With the increase of the feature dimension, the user interest feature is divided more finely, and then the feature vector is shared in a range of the neighboring users which are not similar, so that the effect is not as good as expected.

The above description is only exemplary of the present invention and should not be taken as limiting the invention, as any modification, equivalent replacement, or improvement made within the spirit and scope of the present invention is also included in the present invention.

Appendix

/>

/>

Claims

1. A multitask collaborative filtering method based on neighbor user feature sharing is characterized by comprising two stages: a K neighbor user group generation stage for generating K most similar neighbor users of the users; an interest prediction stage, which is used for predicting the possible grading condition of the current active user for the unevaluated goods by means of the K neighbor user; wherein:

k neighbor user group generation stage

Based on the cognition that similar users have similar interest preference, a K neighbor user group of the current active user is searched, and the optimized inter-user similarity is taken as weight to share a user feature space in the range of the neighbor user group, so that more similar users make greater contribution to the interest preference characterization of the active user;

generating similar users by adopting the Pearson similarity with importance weight strategy, fully considering the bias of sparse common terms to the Pearson similarity to measure the similarity of the users according to the analysis of sparse data, introducing the importance strategy, and only considering the first K users which are just similar;

the method comprises the following specific steps:

step 2: for each user, traversing other remaining users, and judging whether the feedback article sets between the other remaining users have intersection or not; for the case of intersection, calculating the similarity of the two according to a formula 2, if the similarity is greater than 0, recording, and otherwise, setting 0; for the case that no intersection exists, 0 is directly set;

where u and j represent any two users, I _uj Representing a common feedback item set between two users, s representing any item in the common feedback item set, | I _uj | represents the number of common feedback items, γ represents a threshold value for the number of common feedback items among users set for all users, and r _us And r _js Respectively representing the scores of the common items by the user u and the user j,

and &>

Respectively representing the average scoring conditions of the two users;

and step 3: according to the similarity of each user and other users obtained in the step 2, taking K neighbor user groups of the current user, wherein the similarity comprises the top K users in high-to-low order;

(II) user interest prediction stage

By adopting a collaborative filtering algorithm based on multi-task feature learning, interest preference of the current active user is depicted, and according to feedback data of the active user and the interest of similar users reflected by weight, an interest feature space of the current active user is obtained through user feature sharing among neighboring users, so that the problem of low prediction accuracy caused by data sparsity is solved;

the method comprises the following specific steps:

(1) Constructing a user interest prediction model, wherein a score prediction formula for a current active user is as follows:

where u represents an arbitrary user, i represents an arbitrary item,

is the predicted value of the user u's score to the user i, mu is the average score value of the user u, b _u Is the score deviation of user u, b _i Is the score deviation, p, of item i _u Feature vector, q, representing user u _i A feature vector representing item i; n (u) represents K neighbor user set of user u, and j represents any neighbor user; s _uj Representing the similarity value between the user u and the adjacent user j;

introducing a regular penalty term to prevent over-fitting and under-fitting; the loss function of the algorithm is:

wherein R is ^* Representing the observed set of scoring data, r _ui Representing the observed real score of the user u on the item i, and lambda represents a regular coefficient;

the optimal parameters are obtained in a random gradient descent mode by minimizing the loss function of formula 5, and each gradient update formula is as follows:

(2) Training model

Step 1: randomly extracting 20% of user data sets as training sets, using the rest as test sets, and disordering the sequence inside the training sets and the test sets to prevent data from tilting; the generalization capability is improved by randomly carrying out a plurality of times;

and 2, step: training parameters by adopting a random gradient descent method, updating a formula according to the formula (6) to the formula (10) by a gradient updating formula, and iterating until the algorithm converges;

(3) Prediction