CN102609533B - Kernel method-based collaborative filtering recommendation system and method - Google Patents

Abstract

The invention provides a kernel method-based collaborative filtering recommendation system and a kernel method-based collaborative filtering recommendation method. The corresponding system comprises a data preparation module which is used for standardizing the original data and carrying out corresponding preprocessing, generating a user-project rating matrix and a project distance matrix to output; a user interest modeling module which is used for constructing an interest model for a user on a project space according to the user-project rating matrix and the project distance matrix as well as a kernel density estimation technology; and a recommendation result generation module which is used for computing the similarities among the users according to the interest model, generating a neighbor set of a target user, and predicting a score of the project rated by the user according to a predetermined recommendation strategy and returning the recommendation result. Through the recommendation system and the recommendation method provided by the invention, the user interest model can be better presented, the user similarity in the practical application is estimated more accurately, the performance of the recommendation system can be promoted considerably, and more stable recommendation result can be obtained.

Description

A kind of Collaborative Filtering Recommendation System based on kernel method and method

Technical field

The invention belongs to internet data to excavate and information retrieval field, relate to a kind of to the commercial product recommending system in ecommerce class website and method.

Background technology

Along with developing rapidly of Information technology and WEB 2.0 technology, internet information is increasingly huge and keep rapid growth.For Internet user, the problem that solve how efficiently to excavate oneself valuable information rapidly from magnanimity information; And for some websites such as social network sites, e-commerce website, more to consider how effectively to be presented to by web site contents user, to improve service quality.Personalized recommendation technology progressively grows up just in this context, the main thought of this technology is the historical behavior record by digging user, set up the interest model describing user's request, then this interest model is utilized to go to find user and the direct relevance of information, last based on this relevance, with certain recommended models prediction user preference, thus the information pushing meeting this preference or demand is selected for it to user.In e-commerce website, system or can be browsed record and the history score data of user to commodity and recommends it may interested product according to the purchase of user; To share etc. in multimedia sharing website at video, for user recommends its interested video to considerably increase browsing time of user and the viewed quantity of video; At social network sites, the recommendation of good friend has become the stickability of its adding users to website and the important means of satisfaction.

As the one of conventional recommendation technology, Collaborative Filtering Recommendation Algorithm, due to its simple feature efficiently, obtains more favor in practice, also receives the concern of a large amount of researcher simultaneously.Collaborative Filtering Recommendation Algorithm can be divided into based on neighbours' collection (Neighborhood-based) with based on two kinds of model (Model-based).Wherein, the proposed algorithm based on neighbours' collection can be divided into again based on user (User-based) and the algorithm based on project (Item-based).The core concept of Collaborative Filtering Recommendation Algorithm be the approximate project of the project utilizing approximated user or user to like to filter bulk information, thus for user filters out it may interested project.Particularly, this user is given in the project recommendation user similar to particular user interests liked based on the Collaborative Filtering Recommendation Algorithm of user; Project-based proposed algorithm is then filter out those projects similar to the project that user likes as recommendation results.Collaborative filtering based on model utilizes statistics and machine learning techniques to obtain a recommended models, and then produce recommendation results with this model.

Traditional Collaborative Filtering Recommendation Algorithm needs to consider ubiquitous drawback in various Collaborative Filtering Recommendation Algorithm, mainly comprise: first, only considered the data of common scoring during the tradition similarity of method for measuring similarity between computational item or user, cause the user of the project only having common scoring to have similar possibility, be not inconsistent with actual conditions; Second, Collaborative Recommendation is faced with the challenge of Sparse and cold start-up problem, when score data is sparse, how can reasonably calculate similarity between user, and then produce the key issue that accurate recommendation results has become the quality improving Collaborative Filtering Recommendation Algorithm.

Summary of the invention

The object of the invention is the deficiency overcoming traditional Collaborative Filtering Recommendation Algorithm, a better user interest modeling scheme and similarity calculating method corresponding are with it provided, thus more accurately digging user historical data, promote the raising of commending system performance.

The object of the invention is to be achieved through the following technical solutions:

Based on a Collaborative Filtering Recommendation System for kernel method, comprising:

Data preparation module, for by raw data standardization and corresponding pre-service, generates user-project rating matrix and project distance matrix and exports;

User interest MBM, for according to described user-project rating matrix and project distance matrix, and by Density Estimator technique construction user at project interest model spatially;

Recommendation results generation module, for according to described interest model, calculates the similarity between user, generate neighbours' collection of targeted customer, and returns recommendation results with predetermined Generalization bounds prediction user to the scoring of project.

Based on a collaborative filtering recommending method for kernel method, comprising:

By raw data standardization and corresponding pre-service, generate user-project rating matrix and project distance matrix and export;

According to described user-project rating matrix and project distance matrix, and by Density Estimator technique construction user at project interest model spatially;

According to described interest model, calculate the similarity between user, generate neighbours' collection of targeted customer, and with predetermined Generalization bounds prediction user, recommendation results is returned to the scoring of project.

Compared with prior art, its remarkable result is in the present invention: not only consider the positive influences of data with existing to recommendation results during (1) digging user interest, also considers the impact of negative scoring simultaneously, can state user interest model better; (2) user's similarity is no longer dependent on limited common scoring item, but between abundant excavation project, potential contact and user interest, in project diffusion spatially, have estimated the user's similarity in practical application more accurately; (3) when Sparse, lifting larger in commending system performance can be promoted, and obtain more stable recommendation results.

Accompanying drawing explanation

The structural representation of the Collaborative Filtering Recommendation System based on kernel method that Fig. 1 provides for the specific embodiment of the present invention;

The user interest that Fig. 2 provides for the specific embodiment of the present invention is at project distribution schematic diagram spatially;

The schematic flow sheet of the collaborative filtering recommending method based on kernel method that Fig. 3 provides for the specific embodiment of the present invention.

Embodiment

This embodiment provides a kind of Collaborative Filtering Recommendation System based on kernel method, as shown in Figure 1, comprising:

Data preparation module 11, for by raw data standardization and corresponding pre-service, generates user-project rating matrix and project distance matrix and exports;

User interest MBM 12, for according to described user-project rating matrix and project distance matrix, and by Density Estimator technique construction user at project interest model spatially;

Recommendation results generation module 13, for according to described interest model, calculates the similarity between user, generate neighbours' collection of targeted customer, and returns recommendation results with predetermined Generalization bounds prediction user to the scoring of project.

Concrete, data preparation module 11 is responsible for preparing the data required for whole system, and carries out pre-service to partial data.System mainly uses two class data: the history score data of user to project and the category attribute information of project itself.Concrete enforcement is carried out as follows:

Steps A 1: initialising subscriber-project rating matrix

User-project rating matrix is the canonical representation of user to the history score data of project, and its form as shown in Table 1.The every a line of this matrix represents the scoring of some users to all items, and the scoring of all users to some projects is shown in each list.

Form 1

In matrix, the value of each comprises two kinds of score data, R _{m, n}represent original score data; R ^* _{m, n}represent the history score data of decentralization, scope median of namely marking is the situation of 0.The former is for the last score in predicting stage, and the latter is used for the structure stage that user builds model.If raw data is the data of decentralization, then both are identical.The situation of scoring disappearance is allowed to exist herein.

Steps A 2: similarity of classifying between computational item

Next pre-service to be carried out to the classified information of project itself.Suppose that project category set is contain several classifications preset in C, C, classification number with | C| represents.For project i, its category attribute subset of set C represents, is set to C _i, C _icomprise the one or more elements in C.The formula of the classification similarity between computational item i and project j is as follows:

${\mathrm{Sim}}_c(i,j)=\frac{{|c_i\∩c_j|}^2}{\left|c\right|*|c_i\∪c_j|}$

This formula rational number that [0,1] is interval weighs the similarity between two projects.To all items, calculate classification similarity between any two.The categorical attribute of project is relatively fixing, so this step can be carried out by off-line, and upgrades in time when there being new projects to add.

Steps A 3: the Pearson correlation coefficient between computational item

Similar with steps A 2, need here to calculate all items Pearson correlation coefficient between any two.Pearson correlation coefficient computing formula between project i and project j is as follows:

$\mathrm{Corr}(i,j)=\frac{{\mathrm{\Σ}}_{u\∈U_{i,j}}(R_{u,i}-\stackrel{\‾}{R_i})(R_{u,j}-\stackrel{\‾}{R_j})}{\sqrt{{\mathrm{\Σ}}_{u\∈U_{i,j}}{(R_{u,i}-\stackrel{\‾}{R_i})}^2}\sqrt{{\mathrm{\Σ}}_{u\∈U_{i,j}}{(R_{u,j}-\stackrel{\‾}{R_j})}^2}}$

Wherein U _{i, j}representing has the user of scoring to gather to project i and project j, R _{u, i}for the value in project rating matrix, represent that user u is to the scoring of project i, for the average of the score value that project i obtains.

Steps A 4: the distance in computational item space between projects

On the basis of steps A 2 and steps A 3, calculate the distance metric between projects, form project distance matrix as shown in Table 2.

Form 2

	Project 1	Project 2	......	Project N
					Project 1	D 1，1	D 1，2	......	D 1，N

Project 2	D 2，1	D 2，2	......	D 2，N
					......	......	......	......	......
Project N	D N，1	D N，2	......	D N，N

Wherein D _{n, N}distance between expression project N and project N, computing method are as follows:

D _i，j＝1-Sim _c(i，j)*Corr(i，j)

So far, the work of data preparation module completes, and exports the input as subsequent module of user-project rating matrix and project distance matrix.

User interest MBM 12 utilizes the method for Density Estimator to estimate in the distribution spatially of whole project user interest.Density Estimator is the nonparametric technique of density Estimation.In the method, kernel function and bandwidth thereof have multiple choices.This instructions illustrates the concrete implementation step of user interest modeling scheme in the present invention for nucleus vestibularis triangularis function.

Step B1: calculate the kernel function value of user on all items

For a user, in existing score data, contain the distribution of its interest.Suppose that user u comments undue project set to be I _u.I _uit is the subset of project set I.Before the computation, first to preset the bandwidth h of kernel function, such as, establish h=0.4.For project i, user's kernel function value thereon calculates as follows.To I _uin project j, calculate the scoring of j to the impact of project i:

If $\left|D_{i,j}\right|<\sqrt{2}h,$ Then $K_i\left(j\right)=R_{u,j}\&times;(\sqrt{2}h-|D_{i,j}\left|\right)/2h;$

Otherwise, K _i(j)=0;

Calculate kernel function value for all items as above method of pressing, acquisition user interest is as shown in Figure 2 at project distribution plan spatially.

Step B2: the interest density of estimating user on all items

On the basis of steps A 1, estimating user-project interest density matrix, as shown in Table 3.

Form 3

In table represent the interest density of user u on project i, calculate according to formula below:

${\hat{f}}_u\left(i\right)=\frac{1}{\left|I_u\right|*h}{\mathrm{\&Sigma;}}_{j{\&Element;I}_u}{K}_i\left(j\right)$

Repeat as above to calculate to all users, until whole user-project interest density matrix is filled up.So far, the task of user interest MBM completes, and exports user-project interest density matrix for recommendation results generation module.

The user that recommendation results generation module 13 utilizes the first two module to produce-project rating matrix and the score data of user-project interest density Estimation matrix to disappearance are predicted, and recommendation results is returned to user.Here to be recommended as to targeted customer u the workflow that example introduces this module.Concrete enforcement is carried out according to as described below four steps.

Step C1: the interest distribution similarity calculating targeted customer and other users

In the user shown in form 3-project interest density matrix, every a line can be regarded as one and simplifies the user interest model of version, and the interest distribution similarity calculating user u and user v is equivalent to the similarity calculating in user-project interest density matrix two corresponding row vectors to a certain extent.In the specific embodiment of the invention, different from traditional similarity calculation method, weigh the similarity between user interest distribution with the volume of two distribution laps.Particularly, for user u and user v, consider each project i and the interest density on it with

If ${\hat{f}}_u\left(i\right)*{\hat{f}}_v\left(i\right)>0,$ Then $\mathrm{sim}(u,v)=\mathrm{sim}(u,v)+\min \left(\right|{\hat{f}}_u\left(i\right)|,|{\hat{f}}_v\left(i\right)\left|\right),$

Otherwise, $\mathrm{sim}(u,v)=\mathrm{sim}(u,v)+\min ({\hat{f}}_u\left(i\right),{\hat{f}}_v\left(i\right))$

All items travels through the similarity just obtaining two user u and v after a time.Repeat to calculate the similarity that can obtain between targeted customer u and other all users of data centralization above.

Step C2: generate targeted customer neighbours' collection according to interest distribution similarity

For targeted customer u, those users being greater than 0 with its similarity form its neighbours and collect, and similarity is that the user of negative value forms its negative neighbours' collection.

Step C3: prediction disappearance score data

For destination item i, target of prediction user u is shown below to its scoring:

$P_{u,i}={\stackrel{\&OverBar;}{R}}_u+\frac{{\mathrm{\&Sigma;}}_{v\&Element;N\left(u\right)\&cap;U_i}\mathrm{sim}(u,v)*(R_{v,i}-{\stackrel{\&OverBar;}{R}}_v)}{{\mathrm{\&Sigma;}}_{v\&Element;N\left(u\right)\&cap;U_i}\left|\mathrm{sim}(u,v)\right|}$

Wherein, N (u) represents neighbours' collection of user u, U _ithe project i of being expressed as comments undue user's set.

Step C4: produce recommendation results and return to targeted customer

For the project that user does not mark, predict that it is marked according to the method for step 3, then according to predicting that scoring order is from high to low to entry sorting.According to the needs in practical application, a front K project is returned to user, K is preassigned value.

Adopt the technical scheme that this embodiment provides, compared with prior art, its remarkable result is: not only consider the positive influences of data with existing to recommendation results during (1) digging user interest, also considers the impact of negative scoring simultaneously, can state user interest model better; (2) user's similarity is no longer dependent on limited common scoring item, but between abundant excavation project, potential contact and user interest, in project diffusion spatially, have estimated the user's similarity in practical application more accurately; (3) when Sparse, lifting larger in commending system performance can be promoted, and obtain more stable recommendation results.

The specific embodiment of the present invention additionally provides a kind of collaborative filtering recommending method based on kernel method, as shown in Figure 3, comprising:

Step 31, by raw data standardization and corresponding pre-service, generates user-project rating matrix and project distance matrix and exports;

Step 32, according to described user-project rating matrix and project distance matrix, and by Density Estimator technique construction user at project interest model spatially;

Step 33, according to described interest model, calculates the similarity between user, generate neighbours' collection of targeted customer, and returns recommendation results with predetermined Generalization bounds prediction user to the scoring of project.

Optionally, corresponding pre-service comprises: the score data comprising negative value that to be all converted into 0 by described raw data be median, retains original score data simultaneously.

Optionally, corresponding project distance obtains by the following method: obtain the distance of project spatially between two projects by the distance metric function of the classification similarity between project and Pearson correlation coefficient.

Optionally, user builds by the following method at project interest model spatially: adopt the interest of the method establishment user of Density Estimator at project model spatially.

Optionally, the similarity between user obtains by the following method: adopt the overlapping part of two projects probability distribution spatially to estimate the similarity of two users.

The implementation of the above-mentioned various method steps based on comprising in the collaborative filtering recommending method of kernel method describes in system embodiment before, in this no longer repeated description.

Adopt the technical scheme that this embodiment provides, compared with prior art, its remarkable result is: not only consider the positive influences of data with existing to recommendation results during (1) digging user interest, also considers the impact of negative scoring simultaneously, can state user interest model better; (2) user's similarity is no longer dependent on limited common scoring item, but between abundant excavation project, potential contact and user interest, in project diffusion spatially, have estimated the user's similarity in practical application more accurately; (3) when Sparse, lifting larger in commending system performance can be promoted, and obtain more stable recommendation results.

The above; be only the present invention's preferably embodiment, but protection scope of the present invention is not limited thereto, is anyly familiar with those skilled in the art in the technical scope that the present invention discloses; the change that can expect easily or replacement, all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claims.

Claims (6)

1. based on a Collaborative Filtering Recommendation System for kernel method, it is characterized in that, comprising:

Data preparation module, for by raw data standardization and corresponding pre-service, generates user-project rating matrix and project distance matrix and exports;

User interest MBM, for according to described user-project rating matrix and project distance matrix, and by Density Estimator technique construction user at project interest model spatially;

Recommendation results generation module, for according to described interest model, calculates the similarity between user, generate neighbours' collection of targeted customer, and returns recommendation results with predetermined Generalization bounds prediction user to the scoring of project; Wherein, described recommendation results generation module comprises: similarity measurement submodule, for the similarity adopting the overlapping part of two projects probability distribution spatially to estimate two users;

Comprise in data preparation module:

Pre-service submodule, for the score data comprising negative value that to be all converted into 0 by described raw data be median, retains original score data simultaneously.

2. the Collaborative Filtering Recommendation System based on kernel method according to claim 1, is characterized in that, comprises in user interest MBM:

Distance calculating sub module, obtains the distance of project spatially between two projects for the distance metric function by the classification similarity between project and Pearson correlation coefficient.

3. the Collaborative Filtering Recommendation System based on kernel method according to claim 1, is characterized in that, also comprises in user interest MBM:

Submodule set up by model, for adopting the interest of the method establishment user of Density Estimator at project model spatially.

4. based on a collaborative filtering recommending method for kernel method, it is characterized in that, comprising:

By raw data standardization and corresponding pre-service, generate user-project rating matrix and project distance matrix and export;

According to described user-project rating matrix and project distance matrix, and by Density Estimator technique construction user at project interest model spatially;

According to described interest model, calculate the similarity between user, generate neighbours' collection of targeted customer, and with predetermined Generalization bounds prediction user, recommendation results is returned to the scoring of project; Wherein, the similarity between described user obtains by the following method: adopt the overlapping part of two projects probability distribution spatially to estimate the similarity of two users;

Described pre-service comprises:

The score data comprising negative value that to be all converted into 0 by described raw data be median, retains original score data simultaneously.

5. the collaborative filtering recommending method based on kernel method according to claim 4, is characterized in that, described project distance obtains by the following method:

The distance of project spatially between two projects is obtained by the distance metric function of the classification similarity between project and Pearson correlation coefficient.

6. the collaborative filtering recommending method based on kernel method according to claim 4, is characterized in that, described user builds by the following method at project interest model spatially:

Adopt the interest of the method establishment user of Density Estimator at project model spatially.