CN102968506A - Personalized collaborative filtering recommendation method based on extension characteristic vectors - Google Patents

Abstract

The invention discloses a personalized collaborative filtering recommendation method based on extension characteristic vectors and belongs to the field of computer machine learning. The specific operation processes are as follows: 1, determining the extension characteristic vectors of users or objects; 2, calculating recommendation values of a candidate recommendation object; 3, sequencing the recommendation values from the larger to the smaller; and 4, selecting the objects in the first N to the users. Compared with a current personalized recommendation method, the method provided by the invention has the advantages that 1, with more calculation related information, a recommendation item list can be accurately offered to the user; 2, the characteristics of simplicity, feasibility and high efficiency are provided, and the application for currently widespread distributed calculation is realized; and 3, the recommendation can be made for new users according to the existing information of users and item per se attribute, so as to reduce the influence on the recommendation result due to the deficiency of preference information to a certain extent.

Description

A kind of personalized collaborative filtering recommending method of extension-based proper vector

Technical field

The present invention relates to a kind of personalized recommendation method, be specifically related to a kind of personalized collaborative filtering recommending method of extension-based proper vector, belong to the computer machine learning areas.

Background technology

The core concept of Web 2.0 is " colony's wisdom ", namely based on popular behavior, recommends for each user provides Extraordinary.This becomes the key that a Web uses success or failure so that how to allow the user obtain more accurately faster needed information.

The personalized recommendation engine utilizes special information filtering (Information Filtering) technology, and different content (such as film, music, books, news, picture, webpage etc.) is recommended may interested user.Generally, the realization of recommended engine is by user's personal like is compared with specific fixed reference feature, and attempts predictive user to some fancy grades of scoring item not.Choosing of fixed reference feature may be to extract from the information of project itself, or based on society or the corporate environment at user place.

The personalized recommendation algorithm mainly is divided three classes:

(1) based on demographic recommendation

Be a kind of recommend method that is easy to realize most based on demographic recommendation mechanisms, it just simply finds user's degree of correlation according to the essential information of system user, and other article of then similar users being liked are recommended the active user.At first, there is the modeling of a subscriber data in system to each user, comprising user's essential information, and age of user for example, sex etc. (information category is not limited only to this in the application of reality certainly); Then, according to user's material computation user's similarity, make afterwards recommendation.

(2) content-based recommendation

Content-based recommendation is the recommendation mechanisms that is most widely used at the beginning of recommended engine occurs, its core concept is according to the metadata of recommending article or content, find the correlativity of article or content, then based on user's hobby record in the past, recommend the similar article of user.

(3) based on the recommendation of collaborative filtering

Along with the development of Web 2.0, the Web website is advocated user's participation and user's contribution more, therefore gives birth to because of fortune based on the recommendation mechanisms of collaborative filtering.Its principle is: according to the preference of user to article, find the correlativity between the article, or find the correlativity between the user, and then recommend based on these relevances.

Recommendation mechanisms based on collaborative filtering is the recommendation mechanisms that is most widely used now, it has following advantage: it does not need article or user are carried out strict modeling, and the description that does not require article is machine understandable, so this method also is field independence.The recommendation that this method is calculated is open, can share other people experience, well supports the user to find potential interest preference.

But also there are some problems in this method:

1. the core of method is based on historical data, so the new article that has less preference information and new user there are the problem of " cold start-up ".

The effect of 2. recommending depends on what and accuracy of the historical preference data of user, and when less or character was relatively poor when the preference information in the system, recommendation results often was not fine.

For these problems, be necessary original collaborative filtering recommending mechanism is carried out some improvement, to adapt to a greater variety of production environments.

Summary of the invention

The objective of the invention is to propose a kind of personalized collaborative filtering recommending method of extension-based proper vector in order to overcome the deficiency of existing personalized recommendation method existence.

The objective of the invention is to be achieved through the following technical solutions.

A kind of personalized collaborative filtering recommending method of extension-based proper vector comprises: based on user's collaborative filtering recommending strategy with based on the collaborative filtering recommending strategy of article.

The specific operation process of described collaborative filtering recommending strategy based on the user is:

Step 1.1: the extension feature vector of determining the user.

Definition user's extension feature vector: user _i=(p _{(i, 1)}, p _{(i, 2)}..., p _{(i, m)}, a _{(i, 1)}, a _{(i, 2)}..., a _{(i, p)}).Wherein, user _iRepresent i the extension feature vector that the user is corresponding, 1≤i≤n, n are the total number of users in the website; p _{(i, j)}Represent i user to the preference value of j article, 1≤j≤m, m are the total number of items in the website; a _{(i, k)}Represent k the property value that i user itself has, 1≤k≤p, p are user's attribute number.

Described preference value be in website the user to scoring, the comment of article, buy and browse the information such as record.

Described user property value is the attribute information that the user has, and comprising: user's sex, age, occupation, hour of log-on, liveness, their location, education degree etc.

Step 1.2: calculated candidate is recommended the recommendation of article.

The extension feature vector that obtains according to step 1.1 passes through the recommendation that formula (1) calculated candidate is recommended article.

$R_{(u,j)}=\frac{1}{\left|U_j\right|}\underset{v\∈U_j}{\mathrm{\Σ}}{\mathrm{sim}}_{(u,v)}{p}_{(v,j)}---\left(1\right)$

Wherein, R _{(u, j)}Expression article j is for user u recommendation; U _jExpression has provided user's set of preference value to article j; | U _j| expression set U _jIn element number; Sim _{(u, v)}Similarity between expression user u and the user v specifically refers to the similarity between the extension feature vector of user u and user v, and u, v are two different users in this website, and p _{(v, j)}Be the preference value of user v to article j.

The computing method of the similarity between described user u and the user v comprise: Pearson correlation coefficient, based on the similarity of Euclidean distance and this related coefficient of paddy etc.

Step 1.3: the recommendation of the candidate being recommended article sorts according to order from big to small.

Step 1.4: on the basis of step 1.3 ordering, choose the top n article and recommend user u, N is the artificial a certain positive integer of setting.

Through the operation of above-mentioned steps, namely finish the article of user u are recommended.

The specific operation process of described collaborative filtering recommending strategy based on article is:

Step 2.1: the extension feature vector of determining article.

The extension feature vector of definition article: item _j=(p _(1j), p _{(2, j)}..., p _{(n, j)}, b _{(j, 1)}, b _{(j, 2)}..., b _{(j, q)}).Wherein, item _jRepresent j the extension feature vector that article are corresponding, 1≤j≤m, m are the total number of items in the website; p _{(i, j)}Represent i user to the preference value of j article, 1≤i≤n, n are the total number of users in the website; b _{(j, l)}Represent l the property value that j article itself have, 1≤l≤q, q are the attribute number of the article in the website.

Described preference value be in website the user to scoring, the comment of article, buy and browse the information such as record.

Described goods attribute value is the attribute information that article have, and comprising: article content, classification, price, time, applicable crowd, the place of production etc.

Step 2.2: calculated candidate is recommended the recommendation of article.

The extension feature vector that obtains according to step 2.1 passes through the recommendation that formula (2) calculated candidate is recommended article.

$R_{(u,j)}=\frac{1}{\left|I_u\right|}\underset{j^{\′}\∈I_u}{\mathrm{\Σ}}{\mathrm{sim}}_{(j,j^{\′})}{p}_{(u,j^{\′})}---\left(2\right)$

Wherein, R _{(u, j)}Expression article j is for user u recommendation; I _uExpression user u provides the article set of preference value; | I _u| expression set I _uIn element number; Sim _{(j, j)}Similarity between expression article j and the article j' specifically refers to the similarity between the extension feature vector of article j and article j', and j, j' are two different article in this website, and p _{(u, j ')}Be the preference value of user u to article j'.

The computing method of the similarity between described article j and the article j ' comprise: Pearson correlation coefficient, based on the similarity of Euclidean distance and this related coefficient of paddy etc.

Step 2.3: the recommendation of the candidate being recommended article sorts according to order from big to small.

Step 2.4: on the basis of step 2.3 ordering, choose the top n article and recommend user u, N is the artificial a certain positive integer of setting.

Through the operation of above-mentioned steps, namely finish the article of user u are recommended.

Beneficial effect

The personalized collaborative filtering recommending method of a kind of extension-based proper vector that the present invention proposes is compared with existing personalized recommendation method, has following advantage:

1. because the information that participation is calculated is more, can provide recommended project tabulation for the user more accurately.

2. have simple, easy row, efficient characteristics, be fit to present pandemic Distributed Calculation and use.

3. can from existing information about user and project self attributes, for new user makes recommendation, reduce to a certain extent the impact of preference information shortage on recommendation results.

Description of drawings

Fig. 1 is the schematic flow sheet of the personalized collaborative filtering recommending method of the extension-based proper vector in the specific embodiment of the invention.

Embodiment

Below in conjunction with accompanying drawing and by 2 embodiment, describe the present invention.

Embodiment 1:

In website A, 500 of users are arranged, 2000 kinds of article, each user has age, sex, 3 kinds of attributes of occupation, now use the personalized collaborative filtering recommending method of extension-based proper vector that the 10th user in this website recommended article, its operating process synoptic diagram as shown in Figure 1.Because number of articles greater than number of users, adopts the collaborative filtering recommending strategy based on the user, operating process is as follows:

Step 1.1: the extension feature vector of determining the user.

Definition user's extension feature vector: user _i=(p _{(i, 1)}, p _{(i, 2)}..., p _{(i, 2000)}, a _{(i, 1)}, a _{(i, 2)}..., a _{(i, 3)}).Wherein, user _iRepresent i the extension feature vector that the user is corresponding, 1≤i≤1000; p _{(i, j)}Represent i user to the preference value of j article, 1≤j≤2000; a _{(i, k)}Represent k the property value that i user itself has, 1≤k≤3, respectively corresponding age, sex, 3 kinds of attributes of occupation.

Described preference value is that the user is to the score information of article in website, and score value is 1 to 5.

Described user property value is the attribute information that the user has in the website; Age attribute value be 1 to 6:1 corresponding below 20 years old, 2 corresponding 20-29 year, 3 corresponding 30-39 the year, 4 corresponding 40-49 year, 5 corresponding 50-59 the year, 6 corresponding more than 60 years old; Sex attribute value is 1 and the corresponding male sex of 2:1,2 corresponding women; Occupation attribute value be 1 to 5:1 corresponding student, the 2 corresponding employees of enterprise and institution, 3 corresponding peasants, 4 corresponding office clerks, 5 corresponding other.

Step 1.2: calculated candidate is recommended the recommendation of article.

The extension feature vector that obtains according to step 1.1 passes through the recommendation that formula (3) calculated candidate is recommended article.

$R_{(10,j)}=\frac{1}{\left|U_j\right|}\underset{v\∈U_j}{\mathrm{\Σ}}{\mathrm{sim}}_{(10,v)}{p}_{(v,j)}---\left(3\right)$

Wherein, R _{(10, j)}Expression article j is for user's 10 recommendations; U _jExpression has provided user's set of preference value to article j; | U _j| expression set U _jIn element number; Sim _{(10, v)}Similarity between expression user 10 and the user v specifically refers to the similarity between the extension feature vector of user 10 and user v, and v is other users in this website, and

p _{(v, j)}Be the preference value of user v to article j.

The computing method of the similarity between described user 10 and the user v are Pearson correlation coefficient.

Step 1.3: the recommendation of the candidate being recommended article sorts according to order from big to small.

Step 1.4: on the basis of step 1.3 ordering, choose front 20 article and recommend user 10.

Through the operation of above-mentioned steps, namely finish the 10th user's article are recommended.

Embodiment 2:

In website B, 1000 of users are arranged, and 200 kinds of article, each article have price, time, 3 kinds of attributes of classification, now use the personalized collaborative filtering recommending method of extension-based proper vector that the 10th user in this website recommended article, its operating process synoptic diagram as shown in Figure 1.Because number of users greater than number of articles, adopts the collaborative filtering recommending strategy based on article, its operating process is as follows:

Step 2.1: the extension feature vector of determining article.

The extension feature vector of definition article: item _j=(p _{(1, j)}, p _{(2, j)}..., p _{(1000, j)}, b _{(j, 1)}, b _{(j, 2)}..., b _{(j, 3)}).Wherein, item _jRepresent j the extension feature vector that article are corresponding, 1≤j≤2000; p _{(i, j)}Represent i user to the preference value of j article, 1≤i≤1000; b _{(j, l)}Represent l the property value that j article itself have, 1≤l≤3 are respectively to dutiable value, time, 3 attributes of type.

Described preference value is that the user is to the score information of article in website, and score value is 1 to 5.

Described goods attribute value is the attribute information that article have; The price attribute: the concrete price of article rounds up; Time attribute: be the concrete article productive year; The categorical attribute value be 1 to 5:1 corresponding commodity, 2 corresponding audio-visual products, 3 corresponding household electrical appliances, 4 corresponding clothes, 5 corresponding other.

Step 2.2: calculated candidate is recommended the recommendation of article.

The extension feature vector that obtains according to step 2.1 passes through the recommendation that formula (4) calculated candidate is recommended article.

$R_{(10,j)}=\frac{1}{\left|I_{10}\right|}\underset{j^{\′}\∈I_{10}}{\mathrm{\Σ}}{\mathrm{sim}}_{(j,j^{\′})}{p}_{(u,j^{\′})}---\left(4\right)$

Wherein, R _{(10, j)}Expression article j is for user's 10 recommendations; I ₁₀Expression user 10 provides the article set of preference value; | I ₁₀| expression set I ₁₀In element number; Sim _{(j, j ')}Similarity between expression article j and the article j' specifically refers to the similarity between the extension feature vector of article j and article j', and j, j' are two different article in this website, and

p _{(10, j ')}Preference value for 10 couples of article j' of user.

The computing method of the similarity between described article j and the article j ' are Pearson correlation coefficient.

Step 2.3: the recommendation of the candidate being recommended article sorts according to order from big to small.

Step 2.4: on the basis of step 2.3 ordering, choose front 20 article and recommend user 10.

Through the operation of above-mentioned steps, namely finish the 10th user's article are recommended.

Above-described specific descriptions; purpose, technical scheme and beneficial effect to invention further describe; institute is understood that; the above only is specific embodiments of the invention; be used for explaining the present invention, the protection domain that is not intended to limit the present invention, within the spirit and principles in the present invention all; any modification of making, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (1)

1. the personalized collaborative filtering recommending method of an extension-based proper vector is characterized in that: it comprises based on user's collaborative filtering recommending strategy with based on user's collaborative filtering recommending strategy;

The specific operation process of described collaborative filtering recommending strategy based on the user is:

Step 1.1: the extension feature vector of determining the user;

Definition user's extension feature vector: user _i=(p _{(i, 1)}, p _{(i, 2)}..., p _{(i, m)}, a _{(i, 1)}, a _{(i, 2)}..., a _{(i, p)}); Wherein, user _iRepresent i the extension feature vector that the user is corresponding, 1≤i≤n, n are the total number of users in the website; p _{(i, j)}Represent i user to the preference value of j article, 1≤j≤m, m are the total number of items in the website; a _{(i, k)}Represent k the property value that i user itself has, 1≤k≤p, p are user's attribute number;

Described preference value be in website the user to scoring, the comment of article, buy and browse the information such as record;

Described user property value is the attribute information that the user has, and comprising: user's sex, age, occupation, hour of log-on, liveness, their location, education degree etc.;

Step 1.2: calculated candidate is recommended the recommendation of article;

The extension feature vector that obtains according to step 1.1 passes through the recommendation that formula (1) calculated candidate is recommended article;

$R_{(u,j)}=\frac{1}{\left|U_j\right|}\underset{v\∈U_j}{\mathrm{\Σ}}{\mathrm{sim}}_{(u,v)}{p}_{(v,j)}---\left(1\right)$

Wherein, R _{(u, j)}Expression article j is for user u recommendation; U _jExpression has provided user's set of preference value to article j; | U _j| expression set U _jIn element number; Sim _{(u, v)}Similarity between expression user u and the user v specifically refers to the similarity between the extension feature vector of user u and user v, and u, v are two different users in this website, and

p _{(v, j)}Be the preference value of user v to article j;

The computing method of the similarity between described user u and the user v comprise: Pearson correlation coefficient, based on the similarity of Euclidean distance and this related coefficient of paddy etc.;

Step 1.3: the recommendation of the candidate being recommended article sorts according to order from big to small;

Step 1.4: on the basis of step 1.3 ordering, choose the top n article and recommend user u, N is the artificial a certain positive integer of setting;

Through the operation of above-mentioned steps, namely finish the article of user u are recommended;

The specific operation process of described collaborative filtering recommending strategy based on article is:

Step 2.1: the extension feature vector of determining article;

The extension feature vector of definition article: item _j=(p _{(1, j)}, p _{(2, j)}..., p _{(n, j)}, b _{(j, 1)}, b _{(j, 2)}..., b _{(j, q)}); Wherein, item _jRepresent j the extension feature vector that article are corresponding, 1≤j≤m, m are the total number of items in the website; p _{(i, j)}Represent i user to the preference value of j article, 1≤i≤n, n are the total number of users in the website; b _{(j, l)}Represent l the property value that j article itself have, 1≤l≤q, q are the attribute number of the article in the website;

Described preference value be in website the user to scoring, the comment of article, buy and browse the information such as record;

Described goods attribute value is the attribute information that article have, and comprising: article content, classification, price, time, applicable crowd, the place of production etc.;

Step 2.2: calculated candidate is recommended the recommendation of article;

The extension feature vector that obtains according to step 2.1 passes through the recommendation that formula (2) calculated candidate is recommended article;

$R_{(u,j)}=\frac{1}{\left|I_u\right|}\underset{j^{\′}\∈I_u}{\mathrm{\Σ}}{\mathrm{sim}}_{(j,j^{\′})}{p}_{(u,j^{\′})}---\left(2\right)$

Wherein, R _{(u, j)}Expression article j is for user u recommendation; I _uExpression user u provides the article set of preference value; | I _u| expression set I _uIn element number; Sim _{(j, j)}Similarity between expression article j and the article j' specifically refers to the similarity between the extension feature vector of article j and article j', and j, j' are two different article in this website, and

p _{(u, j ')}Be the preference value of user u to article j';

The computing method of the similarity between described article j and the article j ' comprise: Pearson correlation coefficient, based on the similarity of Euclidean distance and this related coefficient of paddy etc.;

Step 2.3: the recommendation of the candidate being recommended article sorts according to order from big to small;

Step 2.4: on the basis of step 2.3 ordering, choose the top n article and recommend user u, N is the artificial a certain positive integer of setting;

Through the operation of above-mentioned steps, namely finish the article of user u are recommended.

Images