CN105976229A - Collaborative filtering algorithm based on user and project mixing - Google Patents

Abstract

The invention discloses a collaborative filtering algorithm based on user and project mixing, and the algorithm comprises the steps: 1, carrying out the arrangement of a user-project scoring data set, and building a user-project scoring matrix U; 2, calculating the similarities of articles, and ordering the similarities from the big to the small; 3, generating a 'nearest neighbor N' of articles according to the similarity ordering of the articles; 4, calculating the similarities between a target user T and other users, and ordering the similarities from the big to the small; 5, generating a 'nearest neighbor N' of users according to the similarity ordering of the users. The algorithm gives consideration to the similarities of the users and the similarities of the projects, obtains a project prediction score (giving consideration to the similarities of the users and the similarities of the projects at the same time) through employing a weighting method, carries out recommendation according to the ordering of scores, can reduce the value of an MAE (mean average error), and improves the accuracy of a recommended algorithm.

Description

A kind of collaborative filtering mixed based on user and project

Technical field

The present invention relates to personalized recommendation technical field, particularly relate to a kind of association mixed based on user and project Same filter algorithm.

Background technology

Along with the fast development of network technology, people start to be increasingly dependent on the Internet, by network people Substantial amounts of information can be inquired, but this also means that, people have begun to enter information overload time Generation.

By network inquiry to information not necessarily you really want search content, therefore in the face of greatly During the data message measured, the most therefrom choose oneself information one-tenth interested and useful to oneself For an extremely difficult thing.For the Producer of information, how can allow oneself product Product are shown one's talent from substantial amounts of product library, find real interested in them and are ready to pay close attention to theirs User also becomes an extremely difficult thing.During this time the existence of commending system seems particularly has With.It will can be connected between user and project by certain computing, can either make the user discover that The article useful to them, it is also possible to make article that its information to be presented in the user face really liking them Before.

Personalized recommendation system is widely used on e-commerce website, nowadays as Amazon, when When the e-commerce website such as net rely on commending system be oneself obtain substantial amounts of sales volume and client.Sub-horse Inferior previous conviction scholar Greg Linden once represented, in the sale of Amazon, at least 35% comes from and pushes away Recommend algorithm, additionally, the front chief scientist Andreas Weigend of Amazon the most once revealed, it is recommended that System at least Amazon brings the sale of 20%～30% to be come.Can also use at film and video website Proposed algorithm, in this field, it is recommended that algorithm is most widely used is Netflix.Netflix is in publicity money Once mentioning in material, the user having 60% is that the mode selecting oneself film interested of lookup and video is By its commending system.Present commending system spreaded to include music, books, advertisement, mail, Social networks is in interior every field.The quality of proposed algorithm also determine the accuracy rate of commending system with reliable Property.

Conventional proposed algorithm includes rule-based proposed algorithm, content-based recommendation algorithm, works in coordination with Filtering recommendation algorithms etc., one of algorithm the most frequently used during wherein collaborative filtering is personalized recommendation system. The thought of collaborative filtering is once to collect or browsed article record by analyzing user, utilizes phase Like degree find the user similar to this user interest or may the article of this interest, and recommend this use Family.The article recommended than content-based recommendation system by the article of collaborative filtering recommending are had more newly Newness, and can be difficult to automatically analyze the article of content for percolator machine, such as the art work, music Etc. filtering, by sharing other people experience, it is to avoid because the content analysis inaccurate recommendation knot caused The most inaccurate.Collaborative filtering is widely used in commending system, but algorithm itself exists Certain shortcoming and defect, limit the application in systems of this algorithm.The most how optimized algorithm becomes The emphasis of a lot of scholar's research.

Although collaborative filtering is obtained for well utilization on a lot of e-commerce websites, but is as The continuous of ecommerce scale expands, and commodity amount and number of users are continuously increased, traditional collaborative filtering Technology starts to occur in that a series of problem:

The most openness.Theoretical thought according to coordination technique it is known that the realization of collaborative filtering first Need to set up user-project rating matrix, but in the application of practical situation, user and commodity are owned by Bigger quantity, not all commodity can be processed, and in such systems, general user purchases The total amount of the commodity bought only accounts for 1%～2% of the total volume of commodities in website, which results in user-project scoring Matrix is the most sparse.Such data volume greatly and rating matrix the most sparse in the case of, user or The calculating of item similarity needs to expend substantial amounts of resource, and also can searching nearest-neighbors set when Cause the loss of data message, greatly reduce the accuracy rate of proposed algorithm.

2. cold start-up.Traditional collaborative filtering is that behavior based on user is analyzed, so working as When there is a new user, owing to he does not has historical behavior record, system is not aware that his hobby interest, Therefore not having way by simple collaborative filtering recommending is that it recommends suitable project.Equally, website is worked as In add new project because not having user to evaluate it, system the most just cannot to it be predicted scoring, More it cannot be speculated.

3. extensibility.Along with being continuously increased of project and user, traditional algorithm can run into serious extension Property bottleneck problem, this influences whether the accuracy rate of collaborative filtering.Although algorithm based on model can solve The most certain problem, but the hobby that the precondition of algorithm based on model is user keeps not substantially Becoming, this premise also can the use of limit algorithm to a great extent.

The patent of the patent No. 201010613809.3 to some extent solves above openness problem, but It is marked by matrix blank and fills the method for user items matrix and solves forecasting accuracy problem, program More complicated, operation element amount is big, especially in the case of Sparse and data volume are huge, can grasp The property made is poor.It addition, and the method fill during user items matrix and there is certain error the most unavoidably, it is recommended that The accuracy rate of algorithm cannot ensure.And it seems on the solution cold start-up of prior art, scalability problem Not enough.

Summary of the invention

Because the drawbacks described above of prior art, the technical problem to be solved be to provide a kind of based on User and the collaborative filtering of project mixing, consider user's similarity and two sides of item similarity simultaneously Face, introduces controlling elements α and β, and the method utilizing weighting is pre-to the project obtained based on user's similarity Test and appraisal point and the project forecast scoring obtained based on item similarity are combined, and are considered user's phase simultaneously Like the project forecast scoring of degree and item similarity, and recommend according to the sequence of scoring.Confirmatory experiment is adopted Data set be MovieLens.Test result indicate that, this algorithm can reduce the value of mean error MAE, Improve the accuracy rate of proposed algorithm.

For achieving the above object, the invention provides a kind of collaborative filtering mixed based on user and project to calculate Method, it is characterised in that comprise the following steps:

Step 1: arrange user-project score data collection, sets up user-project rating matrix U；

Step 2: calculate Pearson's coefficient, calculates the similarity between article, and by similarity according to from greatly Being ranked up to little, Pearson's coefficient computing formula is:

$sim(i,j)=\frac{\underset{i,j\∈N}{\Σ}(R_{u,i}-{\stackrel{\‾}{R}}_u)(R_{u,j}-{\stackrel{\‾}{R}}_u)}{\sqrt{{\mathrm{\Σ}}_{i,j\∈N}{(R_{u,i}-{\stackrel{\‾}{R}}_u)}^2}\sqrt{{\mathrm{\Σ}}_{i,j\∈N}{(R_{u,j}-{\stackrel{\‾}{R}}_u)}^2}}$

Wherein N represents the project that user marks, and represents the user u scoring to project i, represents that user u is to project The meansigma methods of scoring；

Step 3: according to the sequencing of similarity between article, produce " nearest-neighbors N " of article, utilize The prediction user T scoring to other article of nearest-neighbors N；

Step 4: use the Pearson's coefficient that step 2 draws, calculates between targeted customer T and other users Similarity, and similarity is ranked up according to order from big to small；

Step 5: according to the sequencing of similarity between user, produce " nearest-neighbors K " of user, utilize The prediction user T scoring to other article of nearest-neighbors K；

Step 6: comprehensive utilization article similarity and the scoring utilizing user's similarity to obtain, if target is used Family to the score in predicting of article both in " nearest-neighbors N ", again in " nearest-neighbors K ", then in advance Test and appraisal point are weighted, and are obtained by article similarity and " nearest-neighbors X " of the combination of user's similarity With prediction scoring；

Step 7: according to the sequence of nearest-neighbors X, according to prediction scoring sequence output top-N from big to small Recommended project collection.

Above-mentioned a kind of collaborative filtering mixed based on user and project, it is characterised in that described step 6 introduce parameter alpha and β, set up dictionary pred, are used for depositing final nearest-neighbors information and prediction mark, The step introducing parameter alpha and β numerical value is:

Step 611: select data set；

Step 612: determine nearest-neighbors number；

Step 613: calculate mean absolute error MAE value by collaborative filtering based on user-project, The computing formula of MAE is as follows:

$MAE=\frac{{\mathrm{\Σ}}_{u,i\∈N}|r_{ui}-{\hat{r}}_{ui}|}{\left|N\right|}$

Wherein, N represents scoring total number of users mesh, r_uiRepresent the prediction scoring of algorithm,Represent the reality of user Mark in border；

Step 614: nearest-neighbors number keeps constant, changes the numerical value of α and β, when MAE value is minimum, Obtain parameter alpha and the value of β needed.

Above-mentioned a kind of collaborative filtering mixed based on user and project, it is characterised in that described step " nearest-neighbors X " calculation procedure of step 6 is:

Step 621: select data set；

Step 622: determine the value of parameter alpha and β；

Step 623: calculate mean absolute error MAE value by collaborative filtering based on user-project, The computing formula of MAE is as follows:

$MAE=\frac{{\mathrm{\Σ}}_{u,i\∈N}|r_{ui}-{\hat{r}}_{ui}|}{\left|N\right|}$

Wherein, N represents scoring total number of users mesh, r_uiRepresent the prediction scoring of algorithm,Represent the reality of user Mark in border；

Step 624: the value of parameter alpha and β keeps constant, changes the numerical value of nearest-neighbors, when MAE value Hour, obtain the nearest-neighbors number needed.

The invention has the beneficial effects as follows:

The present invention considers user's similarity and two aspects of item similarity simultaneously, introduce controlling elements α and β, the method utilizing weighting, the project forecast scoring to obtaining based on user's similarity is similar with based on project The project forecast scoring that degree obtains is combined, and is considered the item of user's similarity and item similarity simultaneously Mesh prediction scoring, and recommend according to the sequence of scoring.The data set that confirmatory experiment uses is MovieLens.Test result indicate that, this algorithm can reduce the value of mean error MAE, improves and recommends to calculate The accuracy rate of method.

Below with reference to accompanying drawing, the technique effect of design, concrete structure and the generation of the present invention is made furtherly Bright, to be fully understood from the purpose of the present invention, feature and effect.

Accompanying drawing explanation

Fig. 1 is the overall workflow figure of the present invention.

Fig. 2 is the flow chart that present invention determine that α and β numerical value.

Fig. 3 is the flow chart that present invention determine that nearest-neighbors number.

Fig. 4 is the coordinate diagram that MAE value is affected by nearest-neighbors number of the present invention.

Detailed description of the invention

As it is shown in figure 1, a kind of collaborative filtering mixed based on user and project, it is characterised in that bag Include following steps:

Step 1: arrange user-project score data collection, sets up user-project rating matrix U；

Step 2: calculate Pearson's coefficient, calculates the similarity between article, and by similarity according to from greatly Being ranked up to little, Pearson's coefficient computing formula is:

$sim(i,j)=\frac{\underset{i,j\∈N}{\Σ}(R_{u,i}-{\stackrel{\‾}{R}}_u)(R_{u,j}-{\stackrel{\‾}{R}}_u)}{\sqrt{{\mathrm{\Σ}}_{i,j\∈N}{(R_{u,i}-{\stackrel{\‾}{R}}_u)}^2}\sqrt{{\mathrm{\Σ}}_{i,j\∈N}{(R_{u,j}-{\stackrel{\‾}{R}}_u)}^2}}$

Wherein N represents the project that user marks, and represents the user u scoring to project i, represents that user u is to project The meansigma methods of scoring；

Step 3: according to the sequencing of similarity between article, produce " nearest-neighbors N " of article, utilize The prediction user T scoring to other article of nearest-neighbors N；

Step 4: use the Pearson's coefficient that step 2 draws, calculates between targeted customer T and other users Similarity, and similarity is ranked up according to order from big to small；

Step 5: according to the sequencing of similarity between user, produce " nearest-neighbors K " of user, utilize The prediction user T scoring to other article of nearest-neighbors K；

Step 6: comprehensive utilization article similarity and the scoring utilizing user's similarity to obtain, if target is used Family to the score in predicting of article both in " nearest-neighbors N ", again in " nearest-neighbors K ", then in advance Test and appraisal point are weighted, and are obtained by article similarity and " nearest-neighbors X " of the combination of user's similarity With prediction scoring；

Step 7: according to the sequence of nearest-neighbors X, according to prediction scoring sequence output top-N from big to small Recommended project collection.

The specific algorithm of the present invention is accomplished by

1. the information in data set is converted to user-project rating matrix.

Def loadData():

Set up dictionary trainSet, put into the user id in u1.base, project id and Score information.

Set up dictionary testSet, put into the user id in u1.test, project id and commenting Divide information.

2. calculate article similarity Itemsim

Def Itemsim():

Set up dictionary Itemsim, preserve result of calculation, including user id, project id and Prediction mark.

Utilize formula 1, calculate Pearson's coefficient.

3. produce nearest-neighbors N

Def itemrecommendations():

A=sort []

Traversal dictionary Itemsim, is ranked up according to the size of article similarity.

Set up dictionary pred1, preserve information and the prediction mark of nearest-neighbors.

4. calculate user's similarity Usersim

Def Usersim():

Set up dictionary Usersim, be used for preserving result of calculation, including user id, project Id and prediction mark.

Utilize formula 1, calculate Pearson's coefficient.

5. produce nearest-neighbors K

Def userrecommendations():

A=sort []

Traversal dictionary usersim, is ranked up according to the size of user's similarity.

Set up dictionary pred2, preserve information and the prediction mark of nearest-neighbors.

6. calculate prediction mark

Def ratings():

Introduce parameter alpha and β, set up dictionary pred, be used for depositing final nearest-neighbors Information and prediction mark.

Traversal dictionary pred1 and pred2, if item1==item2 and Userid1==userid2, then

Pred [userid] [item]=α * pred1 [userid1] [item1]+β * pred2 [userid2] [ item2]

7. generate top-N recommended project collection

Def recommendations():

Traversal dictionary pred, is ranked up according to prediction mark size.

Output top-N recommends collection.

As in figure 2 it is shown, be the flow chart determining α and β numerical value.In the present embodiment, described step 6 introduces Parameter alpha and β, set up dictionary pred, is used for depositing final nearest-neighbors information and prediction mark, introduces The step of parameter alpha and β numerical value is:

Step 611: select data set；

Step 612: determine nearest-neighbors number；

Step 613: calculate mean absolute error MAE value by collaborative filtering based on user-project, The computing formula of MAE is as follows:

$MAE=\frac{{\mathrm{\Σ}}_{u,i\∈N}|r_{ui}-{\hat{r}}_{ui}|}{\left|N\right|}$

Wherein, N represents scoring total number of users mesh, r_uiRepresent the prediction scoring of algorithm,Represent the reality of user Mark in border；

Step 614: nearest-neighbors number keeps constant, changes the numerical value of α and β, when MAE value is minimum, Obtain parameter alpha and the value of β needed.

Table 1 α and the β impact on MAE

As shown in table 1, for α and β MAE affected Exemplary statistical data, can be seen by experimental result Going out, when the value of α and β is all higher than 0.5 or respectively less than 0.5 when, the value of MAE can be relatively large, Prediction can be made to mark outside the scope that 1-5 divides this is because value is all higher than 0.5 or respectively less than 0.5, Error is increased.By above table it can be seen that when α takes 0.2, the when that β taking 0.9, can obtain Obtain minimum MAE value, say, that when α=0.2, during β=0.9, it is possible to obtain best recommendation results.

As it is shown on figure 3, determine the flow chart of nearest-neighbors number.In the present embodiment, described step step 6 " nearest-neighbors X " calculation procedure be:

Step 621: select data set；

Step 622: determine the value of parameter alpha and β；

Step 623: calculate mean absolute error MAE value by collaborative filtering based on user-project, The computing formula of MAE is as follows:

$MAE=\frac{{\mathrm{\Σ}}_{u,i\∈N}|r_{ui}-{\hat{r}}_{ui}|}{\left|N\right|}$

Wherein, N represents scoring total number of users mesh, r_uiRepresent the prediction scoring of algorithm,Represent the reality of user Mark in border；

Step 624: the value of parameter alpha and β keeps constant, changes the numerical value of nearest-neighbors, when MAE value Hour, obtain the nearest-neighbors number needed.

The impact on MAE value of the table 2 nearest-neighbors number

N	MAE
		10	0.8607
15	0.8248
		20	0.8016
25	0.8183
		30	0.8207
35	0.8281
		40	0.8184
45	0.8098
		50	0.8113

As shown in table 2, for nearest-neighbors number MAE value affected statistical data, in conjunction with Fig. 4, permissible Finding out, along with the increase of nearest-neighbors number, the value of MAE is totally gradually lowered, and takes at nearest-neighbors number The when of 20, the value of the MAE of system tends to be steady, and MAE value is minimum when N=20, the most at this moment, Recommendation results accuracy rate is the highest.

In sum, the present invention considers user's similarity and two aspects of item similarity simultaneously, introduces and controls Factor-alpha and β, utilize weighting method, to obtain based on user's similarity project forecast scoring and based on The project forecast scoring that item similarity obtains is combined, and is considered user's similarity and project phase simultaneously Like the project forecast scoring of degree, and recommend according to the sequence of scoring.The data set that confirmatory experiment uses is MovieLens.Test result indicate that, this algorithm can reduce the value of mean error MAE, improves and recommends to calculate The accuracy rate of method.

The preferred embodiment of the present invention described in detail above.Should be appreciated that the ordinary skill of this area Personnel just can make many modifications and variations according to the design of the present invention without creative work.Therefore, all Technical staff passes through logical analysis the most on the basis of existing technology, pushes away in the art Reason or the limited available technical scheme of experiment, all should be at the protection model being defined in the patent claims In enclosing.

Claims (3)

1. the collaborative filtering mixed based on user and project, it is characterised in that comprise the following steps:

Step 1: arrange user-project score data collection, sets up user-project rating matrix U；

Step 2: calculate Pearson's coefficient, calculates the similarity between article, and by similarity according to from greatly Being ranked up to little, Pearson's coefficient computing formula is:

$sim(i,j)=\frac{\underset{i,j\∈N}{\Σ}(R_{u,i}-{\stackrel{\‾}{R}}_u)(R_{u,j}-{\stackrel{\‾}{R}}_u)}{\sqrt{{\Σ}_{i,j\∈N}{(R_{u,i}-{\stackrel{\‾}{R}}_u)}^2}\sqrt{{\Σ}_{i,j\∈N}{(R_{u,j}-{\stackrel{\‾}{R}}_u)}^2}}$

Wherein N represents the project that user marks, and represents the user u scoring to project i, represents that user u is to project The meansigma methods of scoring；

Step 3: according to the sequencing of similarity between article, produce " nearest-neighbors N " of article, utilize The prediction user T scoring to other article of nearest-neighbors N；

Step 4: use the Pearson's coefficient that step 2 draws, calculates between targeted customer T and other users Similarity, and similarity is ranked up according to order from big to small；

Step 5: according to the sequencing of similarity between user, produce " nearest-neighbors K " of user, utilize The prediction user T scoring to other article of nearest-neighbors K；

Step 6: comprehensive utilization article similarity and the scoring utilizing user's similarity to obtain, if target is used Family to the score in predicting of article both in " nearest-neighbors N ", again in " nearest-neighbors K ", then in advance Test and appraisal point are weighted, and are obtained by article similarity and " nearest-neighbors X " of the combination of user's similarity With prediction scoring；

Step 7: according to the sequence of nearest-neighbors X, according to prediction scoring sequence output top-N from big to small Recommended project collection.

A kind of collaborative filtering mixed based on user and project, its feature Being, described step 6 introduces parameter alpha and β, sets up dictionary pred, is used for depositing final nearest Neighbor information and prediction mark, the step introducing parameter alpha and β numerical value is:

Step 611: select data set；

Step 612: determine nearest-neighbors number；

Step 613: calculate mean absolute error MAE value by collaborative filtering based on user-project, The computing formula of MAE is as follows:

$MAE=\frac{{\mathrm{\Σ}}_{u,i\∈N}|r_{ui}-{\hat{r}}_{ui}|}{\left|N\right|}$

Wherein, N represents scoring total number of users mesh, r_uiRepresent the prediction scoring of algorithm,Represent the reality of user Mark in border；

Step 614: nearest-neighbors number keeps constant, changes the numerical value of α and β, when MAE value is minimum, Obtain parameter alpha and the value of β needed.

A kind of collaborative filtering mixed based on user and project, its feature Being, " nearest-neighbors X " calculation procedure of described step step 6 is:

Step 621: select data set；

Step 622: determine the value of parameter alpha and β；

Step 623: calculate mean absolute error MAE value by collaborative filtering based on user-project, The computing formula of MAE is as follows:

$MAE=\frac{{\mathrm{\Σ}}_{u,i\∈N}|r_{ui}-{\hat{r}}_{ui}|}{\left|N\right|}$

Wherein, N represents scoring total number of users mesh, r_uiRepresent the prediction scoring of algorithm,Represent the reality of user Mark in border；

Step 624: the value of parameter alpha and β keeps constant, changes the numerical value of nearest-neighbors, when MAE value Hour, obtain the nearest-neighbors number needed.