CN103559626A - Individualized commodity recommendation method based on bigraph resource non-uniform distribution - Google Patents

Abstract

The invention relates to the field of artificial intelligence and electronic commerce, and discloses an individualized commodity recommendation method based on bigraph resource non-uniform distribution. The method includes the following specific steps: the pre-processing step, the resource spread step, the recommended commodity computation step and the individualized recommendation step. The method has the advantages that use preference can be effectively analyzed and predicated by introducing the bigraph resource non-uniform distribution method and simulating a resource spread mechanism in physics, the recommended commodities can have better accuracy and diversity, and therefore the method can have good application prospect especially in the field of the electronic commerce.

Description

Personalized commercial recommend method based on the non-homogeneous distribution of bigraph (bipartite graph) resource

Technical field

The present invention relates to artificial intelligence and e-commerce field, particularly a kind of personalized commercial recommend method based on the non-homogeneous distribution of bigraph (bipartite graph) resource.

Background technology

The transmission of information easily and information service that network brings are promoting the flourish of ecommerce, people are when enjoying bring thus huge pleasantly surprised gradually, also be faced with the challenge changing from traditional shopping way to network virtual shopping way: in the face of Web businessman numerous commodity like this, user cannot a glance understand all commodity by screen, cannot find easily own interested commodity, also cannot directly check the quality of commodity.The appearance of personalized recommendation technology has been solution problems, and it is according to user's Characteristic of Interest and buying behavior, to user, recommends its interested information and commodity.The personalized recommendation system that the personalized recommendation technology of take builds as core is to be based upon mass data to excavate a kind of senior business intelligence platform on basis, take and helps e-commerce website that decision support and the information service of complete personalization are provided as its client does shopping.For example: the commending system of shopping website is lead referral commodity, automatically complete the personalized process of selecting commodity, meet client's individual demand, conventionally recommend based on: website sell fast most commodity, client city of living in, the buying behavior in client's past and purchaser record, infer buying behavior that client is in the future possible etc.

Present stage, personalized recommendation method was a lot, mainly comprised collaborative filtering, content-based proposed algorithm etc.Collaborative filtering analysis user interest, in customer group, find similar (interest) user of designated user, the evaluation of comprehensive these similar users to a certain information, formation system is the fancy grade prediction to this information to this designated user, and its similarity calculating method often adopts cosine similarity and relevant similarity.Content-based recommendation is continuity and the development of collaborative filtering, it does not need the evaluation opinion to project according to user, but the product content information of having selected according to user is calculated the similarity between user, and then recommend accordingly, the method can overcome some problems that collaborative filtering exists, as: sparse property problem etc.Above-mentioned two kinds of methods respectively have its advantage, but also exist many deficiencies, for example, cannot by personalized service, improve user for the satisfaction of recommendation results well, to keeping existing customer, excavate the aspects such as potential customers and exist various weak points, cause the accuracy rate of recommending and diversity to be and cause anxiety.Therefore, be necessary in conjunction with researching and developing a kind of have better accuracy rate and multifarious recommend method for the personalized service of different user.

Summary of the invention

The present invention is directed to the personalized deficiency of ubiquity in prior art, analysis and prediction user preference rationally and effectively, the shortcoming of the accuracy rate of recommendation results and diversity deficiency, provides a kind of novel personalized commercial recommend method based on the non-homogeneous distribution of bigraph (bipartite graph) resource.

For achieving the above object, the present invention can take following technical proposals:

Personalized commercial recommend method based on the non-homogeneous distribution of bigraph (bipartite graph) resource, comprises following concrete steps:

Pre-treatment step: build respectively user and collect U={u ₁, u ₂..., u _m, commodity collection O={o ₁, o ₂..., o _n, wherein, m represents user's quantity, n represents the quantity of commodity; Set up user-commodity connection matrix A={a _ij, wherein, a _ijrepresent the connection between user i and commodity j, if user i selected commodity j, make a _ij=1, otherwise make a _ij=0;

Resource diffusing step: resource diffuses to user with asymmetric diffusion way by commodity, then, diffuses to commodity with same method by user, and concrete steps are as follows:

1) resource is diffused to user by commodity, its resource diffusion formula is:

wherein, p _ljrepresent user u _lfrom commodity o _jthe resource that place obtains, k (u _t) expression user u _tuser's degree, described user's kilsyth basalt shows that user selected the quantity of commodity, α is regulatory factor, for improving accuracy rate and the diversity of resource diffusion;

2) resource is diffused to commodity by user, its resource diffusion formula is:

wherein, q _ilrepresent commodity o _ifrom user u _lthe resource that place obtains, k (o _i) expression commodity o _icommodity degree, described commodity kilsyth basalt shows the number of times that commodity are selected by user, v _lirepresent user u _lto commodity o _iinterest-degree, described interest-degree represents the degree of concern of user to commodity, k (o _s) expression commodity o _scommodity degree, v _lsrepresent user u _lto commodity o _sinterest-degree;

3) by resource

from commodity o _jthrough user u _ldiffuse to commodity o _i, its resource diffusion formula is: $w_{\mathrm{ij}}^l=p_{\mathrm{lj}}{q}_{\mathrm{il}}=\frac{a_{\mathrm{lj}}{a}_{\mathrm{li}}{k}^{\mathrm{\α}}\left(o_i\right)v_{\mathrm{li}}}{{\mathrm{\Σ}}_{t=1}^m{a}_{\mathrm{tj}}{k}^{\mathrm{\α}}\left(u_t\right){\mathrm{\Σ}}_{s=1}^n{a}_{\mathrm{ls}}{k}^{\mathrm{\α}}\left(o_s\right)v_{\mathrm{ls}}};$

4) by resource by commodity o _jthrough all users, diffuse to commodity o _i, its diffusion formula is: $w_{\mathrm{ij}}=$ ${\mathrm{\Σ}}_{l=1}^m{w}_{\mathrm{ij}}^l={\mathrm{\Σ}}_{l=1}^m{p}_{\mathrm{lj}}{q}_{\mathrm{il}}=\frac{1}{{\mathrm{\Σ}}_{t=1}^m{a}_{\mathrm{tj}}{k}^{\mathrm{\α}}\left(u_t\right)}{\mathrm{\Σ}}_{l=1}^m\frac{a_{\mathrm{lj}}{a}_{\mathrm{li}}{k}^{\mathrm{\α}}\left(o_i\right)v_{\mathrm{li}}}{{\mathrm{\Σ}}_{s=1}^n{a}_{\mathrm{ls}}{k}^{\mathrm{\α}}\left(o_s\right)v_{\mathrm{ls}}};$

Recommendations calculation procedure: calculate final resource allocation vector: W '=W+ β W ²,

wherein, represent final resource allocation vector, W={w _ijrepresent the resource matrix obtain through resource diffusing step, and β is customized parameter, for eliminating similarity redundancy,

the initial resource that represents user;

Personalized recommendation step: all commodity that targeted customer was not yet selected, respectively according to final resource allocation vector

in descending sequence of corresponding element, get one or more forward commodity of rank and recommend user as Recommendations.

As preferably, further comprising the steps of, user resources initialization step: for given targeted customer, the initial resource of the commodity of the selected mistake of this targeted customer is made as to 1, the initial resource of the commodity of the non-selected mistake of this targeted customer is made as to 0, the value that obtains a n dimension is 0 or 1 vector, i.e. this targeted customer's initial resource

The present invention, owing to having adopted above technical scheme, has significant technique effect:

Adopted resource diffusion mechanism, can effectively overcome the deficiencies in the prior art, can meet user's individual demand, recommendation results has higher accuracy rate, and compared to existing technology, recommendation results presents better diversity.Further, targeted customer is adopted to initial resource matrix

can, better in conjunction with user's existing preference, improve accuracy rate and the diversity of recommendation results.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of resource diffusing step.

Fig. 2 is the schematic flow sheet of personalized commercial recommend method of the present invention.

Embodiment

Below in conjunction with embodiment, the present invention is described in further detail.

Embodiment 1

Personalized commercial recommend method based on the non-homogeneous distribution of bigraph (bipartite graph) resource, as shown in Figure 1-2, comprises the following steps:

Data pre-service: suppose to have m user, n commodity, user integrates as U={u ₁, u ₂..., u _m, commodity integrate as O={o ₁, o ₂..., o _n.If user i selected commodity j, so just between i and j, connect a limit a _ij, and make a _ij=1 (i=1,2 ..., m; J=1,2 ..., n), on the contrary a _ij=0.Build according to this user-commodity connection matrix A={a _ij.

Resource diffusion: commodity are distributed to user according to the asymmetric diffusion way of resource by resource, and user is spread resource to commodity with same method again.Initialization user resources, for a given targeted customer, the initial resource of the commodity that he was selected is made as 1, and that did not select is made as 0.Can obtain like this 0/1 vector of a n dimension, represent that this user's initial resource allocation forms

user-commodity connection matrix A={a _ijattractive force and user's degree of each node, commodity degree and user's preference is relevant, resource diffusing step is as follows:

First, commodity are distributed to user by resource, and each user is relevant with its user's degree (user's kilsyth basalt shows that this user selects too much Shaoshang's product) by obtaining how many resources, and its resource diffusion formula can be expressed as:

$p_{\mathrm{lj}}=\frac{a_{\mathrm{lj}}}{{\mathrm{\Σ}}_{t=1}^m{a}_{\mathrm{tj}}{k}^{\mathrm{\α}}\left(u_t\right)}---\left(1\right),$

Wherein, p _ljrepresent user u _lfrom commodity o _jthe resource that place obtains, k (u _t) expression user u _tuser's degree, α is regulatory factor, and α span is [0.9 ,-0.5], and this is preferred value, is used for improving accuracy rate and the diversity of algorithm.

Secondly, user spreads resource in a similar fashion to commodity, and it is relevant with user preference (interest-degree) to its commodity degree (kilsyth basalt of commodity shows these commodity by how many users were selected) that each commodity can obtain how many resources, and its resource diffusion formula can be expressed as:

$q_{\mathrm{il}}=\frac{s_{\mathrm{li}}{k}^{\mathrm{\α}}\left(o_i\right)v_{\mathrm{li}}}{{\mathrm{\Σ}}_{s=1}^n{a}_{\mathrm{ls}}{k}^{\mathrm{\α}}\left(o_s\right)v_{\mathrm{ls}}}---\left(2\right),$

Wherein, q _ilrepresent commodity o _ifrom user u _lthe resource that place obtains, k (o _i) expression commodity o _idegree, α is regulatory factor, with identical in formula (1), v _lirepresent user u _lto commodity o _iinterest-degree (interest-degree is obtained by user marking conventionally, the degree of concern of reflection user to certain commodity); In like manner, k (o _s) expression commodity o _scommodity degree, α is regulatory factor, span is [0.9 ,-0.5], v _lsrepresent user u _lto commodity o _sinterest-degree.

Again, resource is from commodity o _jthrough user u _ldiffusion is to commodity o _i, its formula can be expressed as:

$w_{\mathrm{ij}}^l=p_{\mathrm{lj}}{q}_{\mathrm{il}}\frac{a_{\mathrm{lj}}{a}_{\mathrm{li}}{k}^{\mathrm{\α}}\left(o_i\right)v_{\mathrm{li}}}{{\mathrm{\Σ}}_{t=1}^m{a}_{\mathrm{tj}}{k}^{\mathrm{\α}}\left(u_t\right){\mathrm{\Σ}}_{s=1}^n{a}_{\mathrm{ls}}{k}^{\mathrm{\α}}\left(o_s\right)v_{\mathrm{ls}}}---\left(3\right),$

Wherein,

represent commodity o _jthrough user u _lfinally spread to commodity o _iresource.

Finally, resource is by commodity o _jthrough all users, can spread to commodity o _i, its resource diffusion formula can be expressed as:

$w_{\mathrm{ij}}={\mathrm{\Σ}}_{l=1}^m{w}_{\mathrm{ij}}^l={\mathrm{\Σ}}_{l=1}^m{p}_{\mathrm{lj}}{q}_{\mathrm{il}}=\frac{1}{{\mathrm{\Σ}}_{t=1}^m{a}_{\mathrm{tj}}{k}^{\mathrm{\α}}\left(u_t\right)}{\mathrm{\Σ}}_{l=1}^m\frac{a_{\mathrm{lj}}{a}_{\mathrm{li}}{k}^{\mathrm{\α}}\left(o_i\right)v_{\mathrm{li}}}{{\mathrm{\Σ}}_{s=1}^n{a}_{\mathrm{ls}}{k}^{\mathrm{\α}}\left(o_s\right)v_{\mathrm{ls}}}\left(4\right),$

Wherein, w _ijrepresent commodity o _jthrough all users, finally spread to commodity o _iresource.

Recommendations calculate: for different users, it is different that its initialization resource forms, and is designated as vector

it is user's initial resource.The final resource allocation vector obtaining after spreading by resource can be expressed as:

W′=W+βW ² (5)，

${\stackrel{\→}{f}}^{\′}=W^{\′}\stackrel{\→}{f}---\left(6\right),$

Wherein, represent final resource allocation vector, W={w _ijrepresent resource matrix, and β is customized parameter, and span is [0.5 ,-0.1], and this value is preferred value, is used for eliminating similarity redundancy.

Personalized recommendation step: all commodity that targeted customer was not selected, according to final resource allocation vector

the size of middle corresponding element sorts, and the corresponding element here can be the resource apportioning cost that commodity are corresponding, and the probability that this user of the larger explanation of value may like is larger.By the forward commercial product recommending of rank to user.

In a word, the foregoing is only preferred embodiment of the present invention, all equalizations of doing according to the present patent application the scope of the claims change and modify, and all should belong to the covering scope of patent of the present invention.

Claims (2)

1. the personalized commercial recommend method based on the non-homogeneous distribution of bigraph (bipartite graph) resource, is characterized in that, comprises following concrete steps:

Pre-treatment step: build respectively user and collect U={u ₁, u ₂..., u _m, commodity collection O={o ₁, o ₂..., o _n, wherein, m represents user's quantity, n represents the quantity of commodity; Set up user-commodity connection matrix A={a _ij, wherein, a _ijrepresent the connection between user i and commodity j, if user i selected commodity j, make a _ij=1, otherwise make a _ij=0;

Resource diffusing step: resource diffuses to user with asymmetric diffusion way by commodity, then, diffuses to commodity with same method by user, and concrete steps are as follows:

1) resource is diffused to user by commodity, its resource diffusion formula is: wherein, p _ljrepresent user u _lfrom commodity o _jthe resource that place obtains, k (u _t) expression user u _tuser's degree, described user's kilsyth basalt shows that user selected the quantity of commodity, α is regulatory factor, for improving accuracy rate and the diversity of resource diffusion;

2) resource is diffused to commodity by user, its resource diffusion formula is:

wherein, q _ilrepresent commodity o _ifrom user u _lthe resource that place obtains, k (o _i) expression commodity o _icommodity degree, described commodity kilsyth basalt shows the number of times that commodity are selected by user, v _lirepresent user u _lto commodity o _iinterest-degree, described interest-degree represents the degree of concern of user to commodity, k (o _s) expression commodity o _scommodity degree, v _lsrepresent user u _lto commodity o _sinterest-degree;

3) by resource from commodity o _jthrough user u _ldiffuse to commodity o _i, its resource diffusion formula is:

$w_{\mathrm{ij}}^l=p_{\mathrm{lj}}{q}_{\mathrm{il}}=\frac{a_{\mathrm{lj}}{a}_{\mathrm{li}}{k}^{\mathrm{\α}}\left(o_i\right)v_{\mathrm{li}}}{{\mathrm{\Σ}}_{t=1}^m{a}_{\mathrm{tj}}{k}^{\mathrm{\α}}\left(u_t\right){\mathrm{\Σ}}_{s=1}^n{a}_{\mathrm{ls}}{k}^{\mathrm{\α}}\left(o_s\right)v_{\mathrm{ls}}};$

4) by resource by commodity o _jthrough all users, diffuse to commodity o _i, its diffusion formula is: $w_{\mathrm{ij}}={\mathrm{\Σ}}_{l=1}^m{w}_{\mathrm{ij}}^l=$ ${\mathrm{\Σ}}_{l=1}^m{p}_{\mathrm{lj}}{q}_{\mathrm{il}}=\frac{1}{{\mathrm{\Σ}}_{t=1}^m{a}_{\mathrm{tj}}{k}^{\mathrm{\α}}\left(u_t\right)}{\mathrm{\Σ}}_{l=1}^m\frac{a_{\mathrm{lj}}{a}_{\mathrm{li}}{k}^{\mathrm{\α}}\left(o_i\right)v_{\mathrm{li}}}{{\mathrm{\Σ}}_{s=1}^n{a}_{\mathrm{ls}}{k}^{\mathrm{\α}}\left(o_s\right)v_{\mathrm{ls}}};$

Recommendations calculation procedure: calculate final resource allocation vector: W '=W+ β W ², wherein,

represent final resource allocation vector, W={w _ijrepresent the resource matrix obtain through resource diffusing step, and β is customized parameter, for eliminating similarity redundancy,

the initial resource that represents user; Personalized recommendation step: all commodity that targeted customer was not yet selected, respectively according to final resource allocation vector

in descending sequence of corresponding element, get one or more forward commodity of rank and recommend user as Recommendations.

2. the personalized commercial recommend method based on the non-homogeneous distribution of bigraph (bipartite graph) resource according to claim 1, it is characterized in that, further comprising the steps of, user resources initialization step: for given targeted customer, the initial resource of the commodity of the selected mistake of this targeted customer is made as to 1, the initial resource of the commodity of the non-selected mistake of this targeted customer is made as to 0, and the value that obtains a n dimension is 0 or 1 vector, i.e. this targeted customer's initial resource

Images