CN108460619A - A kind of fusion shows the Collaborative Recommendation model of implicit feedback - Google Patents

Abstract

The invention discloses the Collaborative Recommendation models that a kind of fusion shows implicit feedback, from the angle of the true shopping process of analog subscriber, our recommendation task is completed by two steps, after being modeled first by the user concealed feedback data of the recommended models of sequencing-oriented analysis, the article selected user and most possibly browsed is concentrated from article to be sorted, then pass through user's explicit feedback data using the recommended models towards scoring, learn user characteristics matrix and article characteristics matrix for predicting scoring, it is then based on scoring and sequencing is rearranged to the article collection selected before, target user is returned to as last recommendation list, to greatly improve the accuracy of recommendation, it is worth with extensive commercial application.

Description

Collaborative recommendation model integrating explicit and implicit feedback

Technical Field

The invention belongs to the technical field of computer software, and particularly relates to a collaborative recommendation model integrating explicit-implicit feedback.

Background

In most current online shopping platforms, the historical behavior data (such as rating, praise, browsing history) of the user implicitly contains the preference information of the user, and many recommendation systems use the preference information as one of important sources of input data. According to the feedback mechanism of the user, the online behavior data of the user can be divided into explicit feedback data and implicit feedback data, and the two data are concurrent with the current e-commerce field, and the explicit feedback data or the implicit feedback data rarely independently exist. However, most of the prior personalized recommendation technologies focus on the single explicit feedback or implicit feedback, which is difficult to meet the urgent needs of the current e-commerce for recommendation technologies, and how to fuse the explicit and implicit feedback for recommendation tasks is one of the difficulties of research. In addition, the sequence in the finally generated recommended item set influences the purchasing intention of people, and it is generally considered that the former item is the most desirable item to purchase by the user, so the sequence between different items in the recommendation result should be considered by the recommendation task.

In the present invention, we consider the purchase flow of the user in the real scene as follows: the user browses the commodities for a long time and selects the commodities which the user wants to buy. The user leaves much implicit feedback data (such as clicks, browsing, purchasing records and the like) in the browsing process, and the evaluation information after purchasing is explicit feedback data (such as scores, ratings and the like) of the user, so the final recommendation result should not depend on the explicit feedback data or the implicit feedback data alone, and the recommendation task can be divided into two steps: the method comprises the steps of firstly generating an item set which is most likely to be browsed by a user, and then reordering the obtained item set to generate an item set which is most likely to be purchased by the user.

Disclosure of Invention

The invention aims to provide a collaborative recommendation model fusing explicit-implicit feedback. The model has good recommendation effect and can realize personalized recommendation.

The technical scheme adopted by the invention for solving the technical problem comprises the following steps:

step 1, training a required ranking-oriented recommendation model by using implicit feedback data, and selecting an article set I which is most likely to be browsed by a user;

step 2, adding explicit feedback data and implicit feedback data into the scoring-oriented model to train the model, and learning a user characteristic matrix and an article characteristic matrix for predicting scoring; reordering the item set I obtained in the step 1 to obtain an item set II which is most likely to be purchased by the user, and returning the item set II to the target user as a final recommendation result;

through the steps, the implicit feedback data and the explicit feedback data of the user are used for personalized recommendation, and the purpose of improving the personalized recommendation effect is achieved.

The step 1 is specifically realized as follows:

the recommendation algorithm used by the ranking-oriented recommendation model in step 1 may adopt a more excellent algorithm bpr (bayesian qualified ranking) applied to implicit feedback data in the current ranking learning algorithm, and the algorithm is implemented as follows:

firstly, carrying out data pair preprocessing on historical scoring data of a user: the BPR algorithm processes the user's score data for an item (the item with which the user has interacted is denoted as "1", and the item without interaction is denoted as "0") into a set of pair < i, j >, where i is the item with a score of 1 and j is the item with a score of 0. Expressed in a triplet < u, i, j >: user "u" likes item "i" more than item "j".

Preconditions of the recommendation model:

① the preference behavior between users is independent.

② the order bias relationship of the same user to different items is independent.

Thus, a target is defined that needs maximization:

Πp(Θ|i＞_uj)∝Πp(i＞_uj|Θ)p(Θ)(1)

wherein Θ is the recommended model parameter that is sought, including: a user's feature matrix U and an item's feature matrix V. p (Θ | i >)_uj) Representing the posterior probability, p (i >)_uj | Θ) represents the likelihood part, p (Θ) represents the prior probability, i >_uj indicates that user "u" likes item "i" more than item "j".

Wherein with respect to the likelihood part: p (i >)_uj|Θ)＝δ(x_u，i-x_u，j)，x_u，i＝p_u·p_iLet the prior probability obey the following distribution: theta to N (0, lambda)_ΘI) Then the density function of the prior probability is:

based on the above assumptions, the optimization objective is further expanded to obtain:

Πp(i＞_uj|Θ)p(Θ)∝ln(p(i＞_uj|Θ)p(Θ))＝∑(lnδ(x_u，i-x_u，j)+lnp(Θ))

＝∑(lnδ(x_u，i-x_u，j)-λ_Θ||Θ||²)

＝∑(lnδ(x_u，i-x_u，j)-λ_Θ||p_u||²-λ_Θ||q_i||²-λ_Θ||q_j||²)

＝∑(lnδ(p_u·q_i-p_u·q_j)-λ_Θ||p_u||²-λ_Θ||q_i||²-λ_Θ||q_j||²)(3)

wherein x_u，i<p_u，q_i>The preference degree of the user u for the item i is expressed in the form of an inner product of two vectors, and theta is a parameter of the recommendation model. In order to maximize the function value of the above expression (3), it is solved by a random gradient descent algorithm (SGD). The invention can obtain the item set I which is most likely to be browsed by the user through the method.

The step 2 is realized as follows:

for the recommendation model facing the score in step 2, the SVD + + model may be used to complete the screening of the item set ii most likely to be purchased by the user. The SVD + + model is improved by integrating implicit feedback data on the basis of a Singular Value Decomposition (SVD) model. The core of SVD is a matrix decomposition technology, and the basic idea is to extract some features according to a user's historical behavior data set (such as user's score data) as the basis of a recommendation task. For movie recommendations, these features can be understood as the degree of fun, the degree of horror, the degree of adventure, etc. of the movie.

In the SVD model, each user has an item that is not scored for which the user may actually have a potential interest, but the SVD incorrectly believes that the user is not interested in the item. The SVD + + model incorporates the fact that the user is potentially interested.

Where N (u) represents the user u scoring item set, and W is called the potential item feature matrix. The objective function is defined as follows:

wherein,vector b_iRepresenting the deviation of the score of movie i from the mean score, vector b_uRepresenting the deviation of the score made by user u from the average score, the average score is scored as μ.

The invention has the following beneficial effects:

the invention discloses a collaborative recommendation model integrating explicit-implicit feedback. In the traditional recommendation technology, the rating-oriented recommendation model only utilizes the explicit feedback data of the users, so that the capability of capturing the like similarity between the users is lost, and the ranking-oriented recommendation model only considers whether the preferences of the users to the product pairs are consistent and ignores the important information of the preference degree between the users. The invention analyzes and researches two traditional recommendation models and finds respective defects. Therefore, from the perspective of simulating the real shopping process of the user, the invention combines the advantages of the two recommendation models together by two steps to complete the recommendation task, firstly, the recommendation model facing the sequencing analyzes the implicit feedback data of the user to model, then, the most likely browsed items of the user are selected from the item set to be sequenced, then, the recommendation model facing the grading learns the user characteristic matrix and the item characteristic matrix for predicting the grading through the explicit feedback data of the user, and then, the selected item set is rearranged in the sequence based on the grading and is returned to the target user as the final recommendation list.

Drawings

FIG. 1 is a flow chart of the present invention.

Detailed Description

The invention is further illustrated by the following figures and examples.

A collaborative recommendation model fusing explicit and implicit feedback is shown in the overall framework of fig. 1. The method comprises the following steps:

step 1, training a required ranking-oriented recommendation model (such as BPR) by using implicit feedback data, and selecting an item set I which is most likely to be browsed by a user.

And 2, adding explicit feedback data and implicit feedback data into the rating-oriented model (such as SVD + +) to train the model, reordering the item set I obtained in the step 1 to obtain an item set II which is most likely to be purchased by the user, and taking the item set II as a final recommendation result.

Through the steps, the implicit feedback data and the explicit feedback data of the user are used for personalized recommendation, and the purpose of improving the personalized recommendation effect is achieved.

The step 1 is specifically realized as follows:

the recommendation algorithm used by the ranking-oriented recommendation model in step 1 may adopt a more excellent algorithm bpr (bayesian qualified ranking) applied to implicit feedback data in the current ranking learning algorithm, and the algorithm is implemented as follows:

firstly, carrying out data pair preprocessing on historical scoring data of a user: the BPR algorithm processes the user's score data for an item (the item with which the user has interacted is denoted as "1", and the item without interaction is denoted as "0") into a set of pair < i, j >, where i is the item with a score of 1 and j is the item with a score of 0. Expressed in a triplet < u, i, j >: user "u" likes item "i" more than item "j".

Preconditions of the recommendation model:

① the preference behavior between users is independent.

② the order bias relationship of the same user to different items is independent.

Thus, a target is defined that needs maximization:

Πp(Θ|i＞_uj)∝Πp(i＞_uj|Θ)p(Θ)(1)

wherein Θ is the recommended model parameter that is sought, including: a user's feature matrix U and an item's feature matrix V. p (Θ | i >)_uj) Representing the posterior probability, p (i >)_uj | Θ) represents the likelihood part, p (Θ) represents the prior probability, i >_uj indicates that user "u" likes item "i" more than item "j".

Wherein with respect to the likelihood part: p (i >)_uj|Θ)＝δ(x_u，i-x_u，j)，x_u，i＝p_u·p_iLet the prior probability obey the following distribution: theta to N (O, lambda)_θI) Then the density function of the prior probability is:

based on the above assumptions, the optimization objective is further expanded to obtain:

Πp(i＞_uj|Θ)p(Θ)∝ln(p(i＞_uj|Θ)p(Θ))＝∑(lnδ(x_u，i，x_u，j)+ln p(Θ))

＝∑(lnδ(x_u，i-x_u，j)-λ_Θ||Θ||²)

＝∑(lnδ(x_u，i-x_u，j)-λ_Θ||p_u||²-λ_Θ||q_i||²-λ_Θ||q_j||²)

＝∑(lnδ(p_u·q_i-p_u·q_j)-λ_Θ||p_u||²-λ_Θ||q_i||²-λ_Θ||q_j||²)(3)

wherein x_u，i<p_u，q_i>The preference degree of the user u for the item i is expressed in the form of an inner product of two vectors, and theta is a parameter of the recommendation model. In order to maximize the function value of the above expression (3), it is solved by a random gradient descent algorithm (SGD). The invention can obtain the item set I which is most likely to be browsed by the user through the method.

The step 2 is realized as follows:

for the recommendation model facing the score in step 2, the SVD + + model may be used to complete the screening of the item set ii most likely to be purchased by the user. The SVD + + model is improved by integrating implicit feedback data on the basis of a Singular Value Decomposition (SVD) model. The core of SVD is a matrix decomposition technology, and the basic idea is to extract some features according to a user's historical behavior data set (such as user's score data) as the basis of a recommendation task. For movie recommendations, these features can be understood as the degree of fun, the degree of horror, the degree of adventure, etc. of the movie.

In the SVD model, each user has an item that is not scored for which the user may actually have a potential interest, but the SVD incorrectly believes that the user is not interested in the item. The SVD + + model incorporates the fact that the user is potentially interested.

Where N (u) represents the user u scoring item set, and W is called the potential item feature matrix. The objective function is defined as follows:

wherein,vector b_iRepresenting the deviation of the score of movie i from the mean score, vector b_uRepresenting the deviation of the score made by user u from the average score, the average score is scored as μ.

Claims (2)

1. A collaborative recommendation model integrating explicit-implicit feedback is characterized in that from the perspective of simulating the real shopping process of a user, the collaborative recommendation model comprises the following specific steps:

step 1, training a required ranking-oriented recommendation model by using implicit feedback data, and selecting an article set I which is most likely to be browsed by a user;

step 2, adding explicit feedback data and implicit feedback data into the scoring-oriented model to train the model, and learning a user characteristic matrix and an article characteristic matrix for predicting scoring; reordering the item set I obtained in the step 1 to obtain an item set II which is most likely to be purchased by the user, and returning the item set II to the target user as a final recommendation result;

the recommendation algorithm used by the ranking-oriented recommendation model in the step 1 adopts a BPR algorithm applied to implicit feedback data in the current ranking learning algorithm, and is specifically realized as follows:

firstly, carrying out data pair preprocessing on historical scoring data of a user: the BPR algorithm processes the scoring data of the user to the articles into a set of pair < i, j >, wherein i is the article with the score of 1, and j is the article with the score of 0; expressed in triads < u, i, j >: user "u" likes item "i" more than item "j"; the item with interaction by the user is marked as '1', and the item without interaction is marked as '0';

the preconditions of the ranking-oriented recommendation model are:

① the preference behavior among users is independent;

② the partial order relations of different articles of the same user are independent;

thus, a target is defined that needs maximization:

Πp(Θ|i＞_uj)∝Πp(i＞_uj|Θ)p(Θ) (1)

wherein Θ is the recommended model parameter that is sought, including: a characteristic matrix U of the user and a characteristic matrix V of the article; p (Θ | i >)_uj) Representing the posterior probability, p (i >)_uj | Θ) represents the likelihood part, p (Θ) represents the prior probability, i >_uj indicates that user "u" likes item "i" more than item "j";

wherein with respect to the likelihood part: p (i >)_uj|Θ)＝δ(x_u，i-x_u，j)，x_u，i＝p_u·p_iLet the prior probability obey the following distribution: theta to N (0, lambda)_ΘI) Then the density function of the prior probability is:

based on the above assumptions, the optimization objective is further expanded to obtain:

wherein x_u，i＜p_u，q_iThe preference degree of the user u to the item i is expressed in an inner product form of two vectors, and theta is a recommendation model parameter; in order to maximize the function value of the above expression (3), solving by a random gradient descent algorithm, the item set I most likely to be browsed by the user is obtained.

2. The collaborative recommendation model with explicit and implicit feedback fusion as claimed in claim 1, wherein for the recommendation model facing the score in step 2, the SVD + + model is used to complete the screening of the item set II most likely to be purchased by the user, and the SVD + + model incorporates the fact that the user has potential interest;

wherein N (u) represents a user u scoring item set, and W is called a potential item feature matrix; the objective function is defined as follows:

wherein,vector b_iRepresenting the deviation of the score of movie i from the mean score, vector b_uRepresenting the deviation of the score made by user u from the average score, the average score is scored as μ.