CN111324807A - Collaborative filtering recommendation method based on trust degree - Google Patents

Abstract

The invention discloses a collaborative filtering recommendation method based on credibility, which is expected to recommend items to the most mental apparatus of users in the large environment of Internet information blowout, thereby saving the precious time of the users and improving the information acquisition efficiency. The algorithm is characterized in that: on the basis of a traditional collaborative filtering algorithm, attribute features (including age factors, occupational factors and gender factors), interestingness (a certain category of items preferred by a user), trust (the satisfaction degree of the user on recommendation of a certain part of users) and other dimensions are fused to model the preference of the user, so that the cold start and sparsity problems of a user scoring matrix are relieved as much as possible, and the most valuable, most desirable and most surprised recommendation is made for the user.

Description

Trust-based collaborative filtering recommendation method

technical field

The invention relates to computer recommendation, in particular to a computer system in which a recommendation system, collaborative filtering, similarity modeling and trust modeling are applied to all recommendation scenarios.

Background technique

The 21st century is an era of information explosion. With the development of computer technology, massive amounts of information have created a huge challenge for information search. How to find the content you need in the huge ocean of information data and make the search more efficient can users Have a better online experience

Based on the above practical market demands, the recommendation system came into being. This is a method to discover new demands based on the user's previous behavior information. Currently, the personalized recommendation algorithm widely used in various e-commerce companies is collaborative filtering recommendation. algorithm, but the user's rating matrix is a sparse matrix after all. Therefore, the cold start problem of new users, data sparsity and recommendation accuracy still need continuous research and improvement.

SUMMARY OF THE INVENTION

The purpose of the present invention is to optimize the shortcomings of the existing recommendation algorithms, and to provide some methods for solving the problems of data sparsity, user cold start, and improving recommendation accuracy.

The contribution points of the present invention to similarity are as follows:

Combine attribute features. The attribute features involved in the present invention include age, occupation, and gender, and these features identify the user's living state. It also affects the user's preferences. By analyzing the similarity of attribute features, the recommendation accuracy can be improved.

Combine interest. Interest degree is the core recommendation basis of recommendation system, and a good interest degree model can greatly improve the accuracy of recommendation.

Combined with trust. In this paper, trust is divided into direct trust and indirect trust. Direct trust can improve the recommendation accuracy, and indirect trust can fill the scoring matrix, which can improve the accuracy and data sparsity.

Table 1 User-rating matrix R(m,n) of m*n based on m users with n items

I₁ I₂ … I_n U₁ R₁₁ R₁₂ … R_1n U₂ R₂₁ R₂₂ … R_2n … … … … … U_m R_m1 R_m2 … R_mn

The technical solution adopted in the present invention is a trust-based collaborative filtering recommendation method, and the implementation process of the method is as follows:

1) User attribute feature similarity modeling

The user attribute of the user attribute feature similarity model integrates the influence of gender, age, and occupation on the user's interest. The establishment process is as follows:

a) Gender similarity modeling

The interests of different genders may be very different. For example, women like cosmetics, clothing, and jewelry are obviously more interested than men, and men generally prefer electronic products. Assuming that the gender of user u is Su and the gender of user v is S _v , the gender similarity S( _u ,v) of user u and user v is as follows:

b) Age similarity modeling

Different ages have different hobbies. Therefore, an age model is introduced to optimize the similarity model. Let the age of user u be A _u and the age of user v to be A _v , then the age similarity of user u and user v A(u,v )as follows:

c) Occupational similarity modeling

The occupations are classified into a tree structure by type as shown in Figure 2. The length of any two nodes in Figure 2 is set to 1, and the height is used to represent the total length. The number of layers in the occupation tree of the nearest parent node of occupation a and occupation b is called Its height, remember H _{a, b} The occupation of user u and user v is represented by occupation a and occupation b respectively, then occupation similarity O(u, v) is expressed as follows:

To sum up, the user's attribute similarity sim _a (u, v) is obtained by comprehensively considering gender, age, and occupation factors as follows:

sim _a (u,v)=αS(u,v)+βA(u,v)+γO(u,v) (4)

Among them, α, β, γ are dynamically adjusted according to the specific system;

2) Similarity modeling of user interest features;

The attribute features discussed in the above section provide a recommendation solution in the case of cold start, but after the user's comment footprint is enriched, the user's specific interest degree can be modeled. The influence of interest on selection is extremely important. The interest degree discussed in this method It is the user's interest in a certain type of item, that is, the user's evaluation ratio for a certain item is used as a measure of his interest in this type of item. Suppose NI _u,i represents the total number of evaluations by user u on items of type i, and NI _u represents the total number of items that the user has evaluated, then the degree of interest I _u,i of user u to items of type i is as follows:

用来标识用户u 对所有项目种类的平均兴趣度，根据皮尔逊相关相似度计算得到，两用户间的兴趣相似度sim_I(u,v)如下：Using formula (5), the set of interest degree of user u can be obtained as I _u =(I _u,1 ,I _u,2 ,...I _u,i ), where I _uv represents the item jointly rated by user u and user v number of species,

It is used to identify the average interest degree of user u to all item types, calculated according to the Pearson correlation similarity, and the interest similarity sim _I (u, v) between two users is as follows:

3) Similarity calculation:

The calculation of user similarity depends on the user-item rating table. Each user interest is represented by a rating vector to map with each row in Table 1. Therefore, the calculation of user similarity is essentially the calculation of the relationship between user rating vectors. the distance. There are many calculation methods for user similarity, and constrained Pearson correlation similarity and jaccard similarity are selected for correlation calculation.

代表用户u所有评分的均值，

标识用户v所有评分的均值，则用户 u和用户v的评分相似度用皮尔逊相关相似性表示如下：Constrained Pearson correlation similarity: R _uv represents the common rating set of user u and user v, R _u,i identifies user u’s rating for item i, R _v,i identifies user v’s rating for item i,

represents the mean of all ratings of user u,

Identifying the mean of all ratings of user v, then the similarity of ratings between user u and user v is expressed as Pearson correlation similarity as follows:

Jaccard similarity: Cosine similarity and constrained Pearson correlation similarity measure similarity based on the common rating items of users. In the case of sparse data, these two similarity calculation methods have a common problem, that is, when the data is sparse. It is sparse to the point that user u and user v have only one common rating item, and both of them rated this item with the same low score. At this time, it means that neither user u nor user v prefers this item, but the cosine similarity is related to the constraint Peel. However, the similarity between the two users is considered to be the maximum value of 1, so that the preferences of the two users are recommended to each other. In order to alleviate this situation, Jarccard similarity is introduced, and its similarity calculation method is based on the number of common scores among users, which can effectively avoid the occurrence of the above problems. Its calculation formula is as follows:

where |Iu _,v | represents the total number of items jointly rated by user u and user v, |I _u |, |I _v | represent the number of items evaluated by user u and user v respectively, and Jac _{(u, v)} takes The value range is [0, 1], and the larger the value, the higher the correlation between the two.

4) User trust modeling:

This method integrates the social trust factor into the collaborative filtering recommendation algorithm, and the definition of trust is: the target user's affirmation of the security, sincerity, and validity of the recommendation opinions put forward by other users. For all the recommended items n made by user a to user b, if user b accepts the recommendation, it indicates that he trusts the recommendation, and if he does not accept it, he does not trust the recommendation.

4.1) Direct trust calculation

Direct trust is the user's initiative to give the trust score of other users directly, but this scheme requires each user to give the trust score of other users, which is difficult to operate in practice. The direct trust degree is calculated as follows:

DTrust _u,v =|sim(u,v)|×Jac _(u,v) (9)

4.2) Indirect trust calculation

The trust relationship between users has the characteristics of subjectivity, weak transitivity, asymmetry and dynamics. Since the trust degree is transitive, the sparsity problem of the rating matrix is alleviated. The direct trust degree between two users depends on whether they have a common rating item. There is direct trust between users with common ratings, but in life It is often found that if user u trusts user i, and user i trusts user v, then user u often has a trust relationship with user v, which is the transitivity of trust. However, since the trust degree decays rapidly, in order to improve the accuracy, only two The order of confidence is calculated as follows:

Among them, PTrust _u,v represents the indirect trust degree between user u and user v, adj(u,v) represents the set of intermediate users that satisfy the condition DTrust _u,b ≥λ,DTrust _b,v ≥λ, and the λ parameter is An adjustable credit threshold, indicating that only intermediate users whose direct credit is greater than λ can be included in the indirect trust calculation.

The trust degree between user u and user v is expressed as Trust _{u, v} , then:

5) In summary:

Integrating user attributes, user interest, and scoring similarity, the comprehensive similarity of trust degree sim _z (u, v) is expressed by the following formula: δ, ε, μ, ρ can vary dynamically according to the degree of dependence of the specific system on each factor parameter tuning

sim _z (u,v)=δsim(u,v)+εsim _I (u,v)+μsim _a (u,v)+ρTrust _u,v (12)

6) Generate recommended results:

According to the comprehensive similarity sim _z (u, v) of user u and user v,

P_u,i表示根据v的评分推测的用户u对项目i的评分，N_u为所有预测评分中评分最高的 top-N个集合，即为最终推荐的项目组合Let the rating of user v to item i be R _v,i , and the average rating of user v to all items is

P _u,i represents the user u's rating for item i based on the rating of v, and N _u is the top-N set with the highest rating among all the predicted ratings, which is the final recommended item combination

7) Evaluation criteria for this method:

The standard mean error (MAE) is used as the rating standard to evaluate the recommendation quality of the algorithm. In the research of recommendation algorithms, MAE is recognized as the most widely used evaluation index. Absolute deviation. The smaller the MAE value, the higher the recommendation accuracy of the recommendation system. The calculation formula is as follows:

Among them: R _u,i represents the actual rating of user u to item i, and P _u,i is the predicted rating of user u to item i obtained according to this method.

Description of drawings

Figure 1 is a flow chart of the implementation of the method.

Figure 2 is a career tree diagram.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

The trust-based collaborative filtering recommendation method is characterized in that: on the basis of the traditional collaborative filtering algorithm, it integrates attributes, interest, trust and other dimensions to model the user's hobbies, so as to reduce user ratings as much as possible. The cold start of the matrix, the sparsity problem, makes the most valuable, most desirable and most surprising recommendations for users.

Based on the traditional collaborative filtering recommendation algorithm, the user's attribute similarity is integrated, and attributes such as age, occupation, and gender that may affect the user's choice are integrated into the algorithm's recommendation process.

Based on the traditional collaborative filtering algorithm, it combines the similarity of user attributes and features, and fully considers the user's preference for a certain type of item into the algorithm.

Based on the traditional collaborative filtering algorithm, the modeling of user trust uses two dimensions of direct trust and indirect trust, which alleviates the problem of the sparsity of the user rating matrix.

The calculation method of the final similarity can dynamically adjust the proportion parameters of different influencing factors according to different scenarios, so as to maximize the items that best meet the needs of users.

Claims (1)

1. The collaborative filtering recommendation method based on the trust degree is characterized by comprising the following steps: the method comprises the following implementation processes:

1) user attribute feature similarity modeling

The user attribute of the user attribute feature similarity model integrates the influence of gender, age and occupation on the interest of the user, and the establishing process is as follows:

a) gender similarity modeling

Suppose user u has a gender S_uGender of user v is S_vThen, the gender similarity S (u, v) of user u and user v) As follows:

b) age similarity modeling

Let the age of user u be A_uAge of user v is A_vThe age similarity a (u, v) of user u and user v is as follows:

c) occupational similarity modeling

Classifying careers into a tree structure according to categories, setting the length of any two nodes to be 1, expressing the total length of the layer number of the nearest parent node of career a and career b in the career tree by Height to be called the Height thereof, and recording H_a,bThe professions of user u and user v are denoted by profession a and profession b, respectively, and the profession similarity O (u, v) is expressed as follows:

in conclusion, the attribute similarity sim of the user is obtained by comprehensively considering the gender, the age and the occupational factors_a(u, v) are as follows:

sim_a(u,v)＝αS(u,v)+βA(u,v)+γO(u,v) (4)

α, adjusting the parameters of gamma according to the specific system;

2) modeling the similarity of user interest characteristics;

provided with NI_u,iIndicates the total number of evaluations, NI, of user u on the i-class item_uRepresenting the total number of the items evaluated by the user, the interest degree I of the user u in the I-type items_u,iThe following were used:

obtaining the interest degree set I of the user u by using the formula (5)_u＝(I_u,1,I_u,2,...I_u,i) Is shown in the formula I_uvThe number of categories of items that represent user u and user v to score together,

the method is used for identifying the average interest degree of the user u to all the item types, and the interest degree sim between the two users is obtained through calculation according to the Pearson correlation similarity_I(u, v) are as follows:

3) similarity calculation:

the calculation of the user similarity relies on the user-item score table, each user interest is represented by a score vector to map with each row in table 1, so the calculation of the user similarity essentially calculates the distance between the user score vectors; the method for calculating the user similarity is numerous, and restricted Pearson correlation similarity and jaccard similarity are selected for correlation calculation;

and (3) constraining the Pearson correlation similarity: r_uvRepresenting a common set of scores for user u and user v, R_u,iIdentifying a rating, R, for item i by user u_v,iThe rating of item i by user v is identified,

represents the mean of all the scores of the user u,

identifying the mean of all scores of user v, the score similarity of user u and user v is expressed in terms of pearson's relative similarity as follows:

jaccard similarity: the calculation formula is as follows:

in the formula I_u,vI represents the total number of commonly scored items for user u and user v, I_u|，|I_vI represents the number of items, Jac, evaluated by user u and user v, respectively_(u,v)Has a value range of [0, 1 ]]A larger value indicates a higher correlation between the two;

4) modeling the user trust:

fusing trust factors in social contact into a collaborative filtering recommendation algorithm, wherein if the user a receives a recommendation, the recommendation is not accepted as untrustworthy, and the trust degree of the user b on the user a is the ratio of the accepted number of strokes to n, and all recommended items n made by the user a on the user b are recommended;

4.1) direct confidence calculation

The direct trust degree is that a user actively and directly gives trust scores of other users, but the scheme requires that each user gives trust scores of other users, and the actual operation is difficult, so the method adopts direct trust degree calculation based on a common user score project, and comprises the following steps:

DTrust_u,v＝|sim(u,v)|×Jac_(u,v)(9)

4.2) Indirect confidence computation

The direct trust between two users depends on whether the two users have a common scoring item, the direct trust exists between the users with common scoring, if the user u trusts the user i, and the user i trusts the user v, the user u often has a trust relationship to the user v, which is the transitivity of trust, and only the second-order trust is taken, and the calculation formula is as follows:

DTrust_u,b≥λ,DTrust_b,v≥λ (10)

wherein, PTrust_u,vRepresents the indirect trust between the user u and the user v, and adj (u, v) represents the satisfaction of the condition DTrust_u,b≥λ,DTrust_b,vIn ≧ λThe lambda parameter is an adjustable credit threshold value, which indicates that only the intermediate users with direct credit larger than lambda can be brought into indirect trust calculation;

the Trust level between user u and user v is denoted Trust_u,vThen:

5) in summary, the following steps:

integrating user attributes and user interest degrees, score similarity and trust degree_z(u, v) is expressed by the following formula: the delta, epsilon, mu and rho can be dynamically adjusted according to different dependence degrees of a specific system on various factors

sim_z(u,v)＝δsim(u,v)+εsim_I(u,v)+μsim_a(u,v)+ρTrust_u,v(12)

6) Generating a recommendation result:

according to the comprehensive similarity sim of the user u and the user v_z(u, v) let user v score item i be R_v,iThe average of the scores of the user v for all the items is

P_u,iDenotes the rating, N, of item i by user u based on the rating inference of v_uThe top-N sets with the highest scores in all the prediction scores are the final recommended item combination

7) Evaluation criteria for the method:

the recommendation quality of the algorithm is evaluated by taking the standard average error MAE as a rating standard, the average absolute deviation of the prediction score of the computing system to the item and the actual score of the user to the item is calculated, the smaller the MAE value is, the higher the recommendation precision of the recommendation system is, and the calculation formula is as follows:

wherein: r_u,iRepresents the actual rating, P, of user u for item i_u,iAnd (4) the predicted scores of the user u on the item i are obtained according to the method.

Images