CN108595598A - A kind of personalized recommendation method based on network reasoning - Google Patents

Abstract

The present invention discloses a kind of novel personalized recommendation method based on network reasoning, it introduces unfavorable ratings using article similitude, solves the problems, such as that the problem of user likes degree to the article of collection is lost and can not correctly be distinguished to the proposed algorithm unfavorable ratings for being mostly based on network.

Description

A kind of personalized recommendation method based on network reasoning

Technical field

The present invention relates to information filtering and Data Minings more particularly to a kind of personalization based on network reasoning to push away Recommend method.

Background technology

In recent years, with the explosion of internet, the growth rate of global information amount is significantly faster than us and handles letter The ability of breath, each user information pageview is less than 1% on internet, and this " information overload " problem becomes increasingly Seriously.Commending system is to solve the problems, such as " information overload " most promising information personalized technological development direction.Although recommending system System is successfully applied to many large scale systems and website, but still has many insoluble problems, and is possessed The space greatly improved is to promote recommendation accuracy.

So far, many different methods are suggested to achieve the purpose that personalized recommendation, this includes based on content Filter (CBF), collaborative filtering (CF), principal component analysis (PCA), hidden semantic model (LSM), the filtering based on two subnetworks Algorithm (NBI) etc..NBI algorithms (as shown in Figure 1) wherein based on mass diffusion theory and widely used collaborative filtering Compared to having shown higher accuracy rate and lower computation complexity.In NBI algorithms, user and article will be taken as difference The node of type finds the cause and effect between article by the process of random walk between node to generate a two subnetwork structures Relationship, and then recommend for user.

In some nearest years, many methods are suggested to improve the performance based on network recommendation algorithm, wherein most with For the purpose of optimizing the resource flow in network, such as Heats, CBI, IBA, UFCI etc..At the beginning of these algorithms all have ignored resource Influence of the beginningization to recommendation accuracy.In most of NBI algorithms, a kind of coarseness method, i.e. user's scoring is all used to be more than 3 Then think that user likes the article, the difference of degree is liked the article of all collections without distinguishing user.

Invention content

The present invention solves the problems, such as to be mostly based on unfavorable ratings in the algorithm of network reasoning using article similitude and loses, Favorable rating of the user to collection article is extracted from unfavorable ratings, it is proposed that one kind using article phase during network reasoning Like the personalized recommendation algorithm of property.

The present invention uses following technical scheme：A kind of personalized recommendation method based on network reasoning, which is characterized in that packet Include following steps：

(1) structure article collection O={ o₁,o₂,...,o_n, user collects U={ u₁,u₂,...,u_m, if user u_iAdopted object Product o_j, then adjacent degree a_ij=1, otherwise, a_ij=0, the adjacency matrix A={ a of a m × n can be obtained_ij}；

(2) for n commodity, the similarity matrix of n × n is built, matrix element is the similarity between two commodity；

(3) for user u_i, its column vector (a is extracted from A_i1,a_i2,...,a_in)^T；User u is extracted according to adjacent degree_i The article o adopted_j, calculate commodity o_jUnfavorable ratings functionWherein, K_iFor user u_iThe article that do not adopt Quantity, Sim_jqFor u_iThe commodity o adopted_jWith the commodity o not adopted_qBetween similarity.Quotient is obtained according to unfavorable ratings function Product o_jAttenuation functionUser u_iCorresponding goods o_jAdjacent degree be updated toFor with The article that family was not adopted, then adjacent degree is without update.Finally obtain user u_iUpdated adjacent degree column vector (a_i′₁, a_i′₂,...,a_i′_n)^T；

(4) user u_iInitial resource bex_jIndicate article o_jThe number of users being adopted, the money of article Source isy_iIndicate user u_iThe number of articles adopted；To obtain commodity resource matrix W=(f (o₁),f (o₂),...,f(o_n))^T。

(5) commodity resource matrix is represented by two multiplication of vectors, specific as follows：

W=W ' × (o₁,o₂,...,o_n)^T

(6) above-mentioned transition matrix W ' is utilized, user u is calculated_iArticle recommendation P, it is as follows：

P=W ' (a_i′₁,a_i′₂,...,a_i′_n)^T=(z_i1,z_i2,...,z_in)^T,

Wherein z_i1,z_i2,…,z_inFor commodity o₁,o₂,…,o_nTo user u_iRecommendation

(7) article that do not collected to all users is arranged according to descending, recommends user successively.

The beneficial effects of the present invention are：

(1) it solves the problems, such as to be mostly based on unfavorable ratings in network reasoning algorithm to lose；

(2) it solves to be mostly based on the article favorable rating that can not accurately distinguish user in network reasoning algorithm to collection Problem；

(3) so that resource original allocation is more reasonable in network-based proposed algorithm.

Description of the drawings

Fig. 1 is the proposed algorithm schematic diagram based on network reasoning.

Specific implementation mode

The present invention relates to a kind of personalized recommendation methods based on network reasoning, including following steps：

(1) structure article collection O={ o₁,o₂,...,o_n, user collects U={ u₁,u₂,...,u_m, if user u_iAdopted object Product o_j, then adjacent degree a_ij=1, otherwise, a_ij=0, the adjacency matrix A={ a of a n × m can be obtained_ij}；

" adopting " of the present invention be bought, used, post-consumer etc.；For example, being pushed away for TV play, film It recommends, " adopting " can be regarded as " having seen "；Recommendation for commodity, " adopting " can be regarded as " purchase "；For vegetable Recommend, " adopting " can be regarded as " eating " or " purchase "；In commercial applications, " adopted " and may generally be expressed as " disappearing Took " or " purchase ".

(2) for n commodity, the similarity matrix of n × n is built, matrix element is the similarity between two commodity；

(3) for user u_i, its column vector (a is extracted from A_i1,a_i2,...,a_in)^T；User u is extracted according to adjacent degree_i The article o adopted_j, calculate commodity o_jUnfavorable ratings functionWherein, K_iFor user u_iThe article that do not adopt Quantity, Sim_jqFor u_iThe commodity o adopted_jWith the commodity o not adopted_qBetween similarity.Quotient is obtained according to unfavorable ratings function Product o_jAttenuation functionUser u_iCorresponding goods o_jAdjacent degree be updated toFor with The article that family was not adopted, then adjacent degree is without update.Finally obtain user u_iUpdated adjacent degree column vector (a_i′₁, a_i′₂,...,a_i′_n)^T；

The update of adjacent degree allows for recommendation Weight, although such as user like article A and B, to two The degree of liking of article may be different, since the article E that article A and user do not like is similar, one can consider that with Family is none so fond of article A, and therefore, when being recommended, article A and B will be with different recommendation weights.

The calculating of similarity can according to different field using CN101079026, CN1434400, CN107610715A, In the open files such as CN104346796A, CN104504055A, CN103631858A, CN107506456A, CN105426916A The method of institute's publicity is calculated.

(4) as shown in Figure 1, user u_iInitial resource be expressed asx_jIndicate article o_jThe user being adopted The resource representation of quantity, article isy_iIndicate user u_iThe number of articles adopted；To obtain commodity money Source matrix W=(f (o₁),f(o₂),...,f(o_n))^T。

(5) commodity resource matrix is represented by two multiplication of vectors, specific as follows：

W=W ' × (o₁,o₂,...,o_n)^T

(6) above-mentioned transition matrix W ' is utilized, user u is calculated_iArticle recommendation P, it is as follows：

P=W ' (a_i′₁,a_i′₂,...,a_i′_n)^T=(z_i1,z_i2,...,z_in)^T,

Wherein z_i1,z_i2,…,z_inFor commodity o₁,o₂,…,o_nTo user u_iRecommendation

(7) article that do not collected to all users is arranged according to descending, recommends user successively.

Claims (1)

1. a kind of personalized recommendation method based on network reasoning, which is characterized in that including following steps：

(1) structure article collection O={ o₁,o₂,...,o_n, user collects U={ u₁,u₂,...,u_m, if user u_iAdopted article o_j, Then adjacent degree a_ij=1, otherwise, a_ij=0, the adjacency matrix A={ a of a m × n can be obtained_ij}；

(2) for n commodity, the similarity matrix of n × n is built, matrix element is the similarity between two commodity；

(3) for user u_i, its column vector (a is extracted from A_i1,a_i2,...,a_in)^T；User u is extracted according to adjacent degree_iAdopt The article o crossed_j, calculate commodity o_jUnfavorable ratings functionWherein, K_iFor user u_iThe number of articles that do not adopt, Sim_jqFor u_iThe commodity o adopted_jWith the commodity o not adopted_qBetween similarity.Commodity o is obtained according to unfavorable ratings function_j Attenuation functionUser u_iCorresponding goods o_jAdjacent degree be updated toNot for user The article adopted, then adjacent degree is without update.Finally obtain user u_iUpdated adjacent degree column vector (a '_i1,a ′_i2,...,a′_in)^T；

(4) user u_iInitial resource bex_jIndicate article o_jThe resource of the number of users being adopted, article isy_iIndicate user u_iThe number of articles adopted；To obtain commodity resource matrix W=(f (o₁),f (o₂),...,f(o_n))^T。

(5) commodity resource matrix is represented by two multiplication of vectors, specific as follows：

W=W ' × (o₁,o₂,...,o_n)^T

(6) above-mentioned transition matrix W ' is utilized, user u is calculated_iArticle recommendation P, it is as follows：

P=W ' (a '_i1,a′_i2,...,a′_in)^T=(z_i1,z_i2,...,z_in)^T,

Wherein z_i1,z_i2,…,z_inFor commodity o₁,o₂,…,o_nTo user u_iRecommendation

(7) article that do not collected to all users is arranged according to descending, recommends user successively.