CN108549729A - Personalized user collaborative filtering recommending method based on Covering reduct - Google Patents

Abstract

The invention discloses a personalized user collaborative filtering recommendation method based on coverage reduction, which clearly defines the concept of redundant users of target users, and according to the function of removing redundant elements in coverage reduction in rough set of coverage, the target user's Redundant users are removed, thereby ensuring the quality of the target user's neighboring users, and using the scoring information of these high-quality neighboring users to provide high-precision and diverse personalized recommendations for the target user.

Description

Collaborative filtering recommendation method for personalized users based on coverage reduction

technical field

The invention relates to the field of recommendation systems, in particular to a personalized user collaborative filtering recommendation method based on coverage reduction that can improve recommendation accuracy and has diversity.

Background technique

The recommendation system can intelligently perceive the user's interests or needs through the user's personal information, realize high-quality recommendation of information, and effectively solve the problem of "information overload". User collaborative filtering algorithm is one of the most widely used and successful technologies in the field of recommendation systems. It assumes that if users have similar hobbies in the past, they may also have similar hobbies in the future. It has the advantages of simple calculation, high efficiency and high precision. . However, in the existing user collaborative filtering algorithm, the adjacent users of the target user tend to have the same hobbies, so the items with high predicted ratings obtained through these adjacent users are often concentrated in a small number of items, or even only popular items, so The diversity of its recommendations is often less than satisfactory.

Contents of the invention

The present invention aims to solve the above-mentioned technical problems existing in the prior art, and provides a personalized user collaborative filtering recommendation method based on coverage reduction that can improve recommendation accuracy and has diversity.

The technical solution of the present invention is: a personalized user collaborative filtering recommendation method based on coverage reduction, which is characterized in that the following steps are followed in turn:

Step 1. Statistically form a two-dimensional scoring information table:

Form the two-dimensional rating information table RM={U, I, R∪{*}} according to the rating information of the item by the user; in the two-dimensional rating information table RM , U represents the set of users, I represents the set of items, and R ∪{*} indicates the user's rating set for the item, where * indicates that the user has not rated the item;

Let user u∈U score r _u,i ∈R∪{*} on item i∈I , and the average score of user u is ; θ is the threshold of user rating, if r _u,i ≥ θ , it indicates that user u likes item i ; the set of items rated by user u is I _u ={i∈I|r _u,i ≠*} ; It is the collection of unrated items for user u ; the item attribute matrix is AM ; in the user set U, if there is a collection of items that user a likes that is included in the collection of items that user b likes, then user a is called the redundant user of the target user ;

Step 2. Use the coverage reduction algorithm to reduce redundant users:

Step 2.1 Let the item collection I be the domain of discourse I. In the domain of discourse I , each user’s favorite items form a set; extract the target user’s favorite attributes from the item attribute matrix AM :

(1)

In formula (1), m represents the number of attributes, at _m represents an attribute, and av _m represents the value of attribute at _m ;

Step 2.2. Use the acquired target user's favorite attributes to construct the target user's decision set D , which is composed of a collection of items with favorite attributes:

(2)

In formula (2), at _m (i) = av _m means that the value of item i on attribute at _m is av _m ;

Step 2.3. Reduce the domain of discourse I from the item set to the decision set D of the target user, that is, the domain of discourse D ; for each user u ∈ U , construct the favorite item set C _u of user u on the domain of discourse D :

(3)

Let C*=D-∪C _u , C={C ₁ ,C ₂ …C _n , C*} constitute a coverage C of the target user in the domain of discourse D ;

Step 2.4 Use the coverage reduction algorithm to reduce the redundant elements from the coverage C , and obtain the reduced coverage reduct ( C ) and the reduced user U ^r :

(4)

Step 3. Use the reduced user U _r to construct a candidate adjacent user u of the target user au ;

Step 4. Calculate the similarity between the target user and the candidate neighboring users, and select the neighboring users of the target user:

Use the Pearson similarity measure function (5) to calculate the similarity between the target user au and _{the candidate neighboring users u∈Ur} ,

(5)

In formula (5), sim(au,u) represents the similarity between the target user au and the candidate adjacent user u∈U _r , I _au ={i∈I| r _au,i ≠*} represents the evaluation of the target user au past collection of items, Indicates the average rating value of the target user;

Then select the first K candidate adjacent users with high similarity as the adjacent users N _au (k) of the target user;

Step 5. Predictive scoring of items not rated by the target user:

According to the scoring information of the target user's neighbor user N _au (k) , use the adjusted weighted sum function (6) to set the unrated item set of the target user au Perform predictive scoring to obtain the target user's predictive score table;

(6)

In formula (6), P _au,i represents the target user au ’s predicted score on item i , U _i ={u∈U|r _u,i ≠*} represents the set of users who have evaluated item i ; λ is used as a regularization factor:

(7)

Step 6. Select the top N items with high predicted scores as the recommendation result.

The present invention clearly defines the concept of redundant users of the target user, and removes the redundant users of the target user according to the function of removing redundant elements in the coverage rough set coverage reduction, thus ensuring the quality of the adjacent users of the target user , so as to use these high-quality rating information of neighboring users to provide high-precision and diverse personalized recommendations for target users.

Description of drawings

Fig. 1 is the schematic flow chart of the embodiment of the present invention;

Fig. 2 is a schematic diagram of corresponding results of the embodiment of the present invention and the comparative accuracy metrics (MAE and RMSE) as the number of neighboring users of the target user changes.

FIG. 3 is a schematic diagram of corresponding results of the diversity measure (Coverage) as the number of adjacent users of the target user changes according to the embodiment of the present invention and the comparative example.

Detailed ways

A kind of personalized user collaborative filtering recommendation method based on coverage reduction of the present invention, as shown in Figure 1, proceeds according to the following steps successively:

Step 1. Statistically form a two-dimensional scoring information table:

Form the two-dimensional rating information table RM={U, I, R∪{*}} according to the rating information of the item by the user; in the two-dimensional rating information table RM , U represents the set of users, I represents the set of items, and R ∪{*} indicates the user's rating set for the item, where * indicates that the user has not rated the item;

Let user u∈U score r _u,i ∈R∪{*} on item i∈I , and the average score of user u is ; θ is the threshold of user rating, if r _u,i ≥ θ , it indicates that user u likes item i ; the set of items rated by user u is I _u ={i∈I|r _u,i ≠*} ; It is the collection of unrated items for user u ; the item attribute matrix is AM ; in the user set U, if there is a collection of items that user a likes that is included in the collection of items that user b likes, then user a is called the redundant user of the target user ;

For example: user set U = {user 1, user 2, user 3, target user}, item set I = {item 1, item 2, item 3, item 4, item 5, item 6}, the value range of rating R is [1,5]. The two-dimensional scoring information table RM is shown in Table 1:

Table 1

Let the threshold value of the user's score equal to 3, and the items with a score greater than or equal to 3 are used as the user's favorite items, as can be seen from Table 1:

User 1's favorite items are {item 2, item 4, item 6};

User 2's favorite items are {item 4, item 6};

User 3's favorite items are {item 2, item 3, item 6};

The target user's favorite items are {item 1, item 3, item 4};

Step 2. Use the coverage reduction algorithm to reduce redundant users:

Step 2.1 Let the item collection I be the domain of discourse I. In the domain of discourse I , each user’s favorite items form a set; extract the target user’s favorite attributes from the item attribute matrix AM :

(1)

In formula (1), m represents the number of attributes, at _m represents an attribute, and av _m represents the value of attribute at _m ;

For example, let the item set I = {item 1, item 2, item 3, item 4, item 5, item 6} as the domain of discourse, Table 2 shows the attribute matrix AM of the item, according to Table 2 and the favorite item set of the target user, statistics Get the attribute value corresponding to the favorite item of the target user:

Comedy=3, Thriller=2, Action=1, Drama=1, Music=1,

Select the two attributes with the largest statistical value as the favorite attributes of the target user, then the favorite attributes of the target user are:

[comedy=1]∧[thriller=1]∧[action=0]∧[drama=0]∧[music=0]

Table 2

comedy thriller action drama music Item 1 1 0 1 1 0 Item 2 1 1 0 1 0 Item 3 1 1 0 0 0 Item 4 1 1 0 0 1 item 5 0 0 1 1 0 Item 6 1 1 1 0 1

Step 2.2. Use the acquired target user's favorite attributes to construct the target user's decision set D , which is composed of a collection of items with favorite attributes:

(2)

In formula (2), at _m (i) = av _m means that the value of item i on attribute at _m is av _m ;

For example, using the favorite attribute of the target user: [comedy=1]∧[thriller=1]∧[action=0]∧[drama=0]∧[music=0], construct the decision set D of the target user, which is composed of All items with (comedy, thriller) attributes, namely:

According to Table 2, the decision set D={item 2, item 3, item 4, item 6} can be obtained;

Step 2.3. In order to eliminate the redundant users of the target user as much as possible, the domain I is reduced from the item set to the decision set D of the target user, that is, the domain D ; for each user u∈U , construct the user u in the domain A collection of favorites C _u on D :

(3)

Let C*=D-∪C _u , C={C ₁ ,C ₂ …C _n , C*} constitute a coverage C of the target user in the domain of discourse D ;

Such as C ₁ = {item 2, item 4, item 6};

C ₂ = {item 4, item 6};

C ₃ = {item 2, item 3, item 6};

Then C = { C ₁ , C ₂ , C ₃ } constitutes a coverage C on the target user decision set D ;

Step 2.4 uses the coverage reduction algorithm to reduce the redundant elements from the coverage C to obtain the coverage reduct ( C ) after the reduction is completed; the completion of the reduction of the redundant elements means that all redundant users of the target user have been deleted, thus User U ^r after reduction:

(4)

Since C ₂ ⊂ C ₁ , according to the cover reduction algorithm, C ₂ is called the redundant element removed from the cover C , so reduce ( C )={ C ₁ , C ₃ }; user 2 is called the target user The redundant users of are removed so the reduced user U ^r = {user 1, user 3};

Step 3. Use the reduced user U ^r to construct the candidate neighboring user u of the target user au , that is, the candidate neighboring users of the target user are {user 1, user 3};

Step 4. Calculate the similarity between the target user and the candidate neighboring users, and select the neighboring users of the target user:

_Use the Pearson similarity measure function (5) to calculate the similarity between the target user au and the candidate neighboring user u∈Ur ,

(5)

In formula (5), sim(au,u) represents the similarity between the target user au and the candidate adjacent user u∈U _r , I _au ={i∈I| r _au,i ≠*} represents the evaluation of the target user au past collection of items, Indicates the average rating value of the target user;

Then select the first K candidate adjacent users with high similarity as the adjacent users N _au (k) of the target user;

That is, use the Pearson similarity metric function to calculate the similarity between the target user and user 1, and the target user and user 3 respectively:

sim(target user, user1) = -0.76

sim(target user, user3) = -0.53

If two candidate adjacent users with the highest similarity are selected as the adjacent users of the target user, then the adjacent user N of _{the target user} (2) = {user 3, user 1};

Step 5. Predictive scoring of items not rated by the target user:

According to the scoring information of the target user's neighbor user N _au (k) , use the adjusted weighted sum function (6) to set the unrated item set of the target user au Perform predictive scoring to obtain the target user's predictive score table;

(6)

In formula (6), P _au,i represents the target user au ’s predicted score on item i , U _i ={u∈U|r _u,i ≠*} represents the set of users who have evaluated item i ; λ is used as a regularization factor:

(7)

According to the rating information of the target user’s neighbor users, the adjusted weighting function is used to predict and score the item 5 and item 6 that the target user has not rated. The results are as follows:

P _{target user, item 5} = 2.16

P _{target user, item 6} = 4.93

Step 6. If an item with the highest predicted score is selected as the recommendation result, item 6 will be recommended to the target user.

experiment:

(1) Use public datasets

Use the public dataset MovieLens, which is often used to test the performance of recommender systems in the recommender system field. The dataset contains 943 users, 1682 movies and 100000 ratings, the distribution of ratings is {1, 2, 3, 4, 5}, and each user has rated at least 20 movies.

(2) Evaluation metrics

The present invention uses mean absolute error MAE and root mean square error RMSE to measure the accuracy of the algorithm. Both MAE and RMSE measure the accuracy of the recommendation result by calculating the deviation between the user's actual rating and the predicted rating. Therefore, MAE and The smaller the RMSE, the higher the recommendation accuracy:

(8)

(9)

Coverage is used to measure the diversity of algorithms. Coverage refers to the ratio of the types of items that can be recommended to target users to the types of items rated by all target users. Therefore, the higher the coverage, the more diverse the recommendations.

(10)

In formula (10), , where S _u,i represents the neighboring users of user u who has rated item i.

(3) Parameter setting

The invention adopts the Pearson similarity measurement function to calculate the user similarity, and uses the adjustment weighting function to predict and score the items not rated by the target user. Select the first two item attributes that the target user likes most as the favorite attribute of the target user. In order to clearly compare the present invention with the traditional user collaborative filtering algorithm, the number of adjacent users of the target user is K∈{20,25,30,...,60}. Set the number of recommended items as {2, 4, 6, 8, 10, 12}.

(4) Comparison and analysis of experimental results

The personalized user collaborative filtering recommendation method based on coverage reduction in the present invention is represented by CBCF, and the traditional user collaborative filtering algorithm is represented by UBCF. Fig. 2 shows the results of accuracy measurement MAE and RMSE. From the data in Figure 2, it can be seen that with the increase of the number of adjacent users of the target user, the MAE and RMSE results of the CBCF algorithm are always smaller than the results of the UBCF algorithm. Since the smaller the MAE and RMSE, the higher the recommendation accuracy, CBCF is able to recommend items with higher accuracy than UBCF. Figure 3 shows the results of the diversity measure Coverage. It can be seen from Figure 3 that as the number of adjacent users of the target user increases, the coverage of the CBCF algorithm is significantly greater than that of the UBCF. Since the higher the coverage, the more diverse the recommendations, so CBCF can recommend more diverse items than UBCF. From the comprehensive experimental results, it can be seen that the present invention can provide high-precision and diversified recommendation results at the same time, so as to realize personalized recommendation for target users.

Claims (1)

1. a kind of personalized user collaborative filtering recommending method based on Covering reduct, it is characterised in that successively in accordance with the following steps It carries out：

Step 1. statistics forms two-dimentional score information table：

Two-dimentional score information table is formed to the score information of article according to userRM={ U, I, R ∪ { * } }；The two dimension scoring letter Cease tableRMIn,UIndicate the set of user,IIndicate the set of article,R∪{*}Indicate scoring set of the user to article, wherein* Indicate that user does not score to article；

Enable useru∈UTo articlei∈IScoring ber _u,i ∈R∪{*}, and useruAverage score be；θIt is commented for user The threshold value divided, ifr _u,i ≥θ, show useruLike articlei；UseruThe article collection to have scored is combined intoI _u ={i∈I|r _u,i ≠*}；For useruThe article set not scored；Goods attribute matrix isAM；In user's set U, liked if there is user a The article set of love is contained in the favorite article set of user b, then user a is known as the redundant subscribers of target user；

Step 2. carries out yojan using Covering reduct algorithm to redundant subscribers：

Step 2.1 enables article setIAs domainI, in domainIIn, the article that each user likes forms a set；In object Product attribute matrixAMMiddle extraction target user's likes attribute：

(1)

In formula (1),mIndicate the number of attribute,at _m Indicate an attribute,av _m Indicate attributeat _m Value；

Step 2.2. likes attribute using the target user of acquisition, builds the decision set of target userD, decision setDBy having The article collection of attribute is liked to be combined into：

D={ i ∈ I|at ₁(i)= av ₁ ,at ₂ (i)=av ₂ ,…,at _m (i)=av _m (2)

In formula (2),at _m (i)= av _m Indicate articleiIn attributeat _m On value beav _m ；

Step 2.3. is by domainIThe decision set of target user is reduced to by article setD, i.e. domainD；For each useru ∈U, build useruIn domainDOn like article setC _u ：

(3)

It enablesC*=D-∪C _u , C={ C ₁ ,C ₂ …C _n , C*} Target user is constituted in domainDOne coveringC；

Step 2.4. utilizes Covering reduct algorithm, by redundant elements from coveringCMiddle yojan obtains the covering after yojanreduct(C) and yojan after userU ^r :

(4)

Step 3. utilizes the user after yojanU ^r Build target userauCandidate adjacent useru；

Step 4. calculates the similarity of target user and candidate adjacent user, chooses the adjacent user of target user：

Target user is calculated using Pearson's similarity metric function (5)auWith candidate adjacent useru∈U _r Between similarity,

(5)

In formula (5),sim(au,u)Indicate target userauWith candidate adjacent useru∈U _r Between similarity,I _au ={i∈I| r _au,i ≠*}Indicate target userauThe article set evaluated,Indicate the average score value of target user；

Then before selection similarity is highKAdjacent user of a candidate adjacent user as target userN _au (k)；

Step 5. article that do not score target user carries out prediction scoring：

According to the adjacent user of target userN _au (k)Score information, using adjustment weighted sum function (6) to target userauNot The article set of scoringPrediction scoring is carried out, the prediction grade form of target user is obtained；

(6)

In formula (6),P _au,i Indicate target userauTo articleiPrediction scoring,U _i ={u∈U|r _u,i ≠*}Article was evaluated in expressioniUser set；λAs a regularization factors：

(7)

Before step 6 selection prediction scorings are highNA article is as recommendation results.