CN105913296A - Customized recommendation method based on graphs - Google Patents

Abstract

The invention provides a customized recommendation method based on graphs and can effectively reduce influence of sparsity on the recommendation effect. According to the method, a first step, a hidden meaning model is utilized to calculate historical scoring records of users to acquire hidden relationships among users and among objects; a second step, similarity among the users is calculated by utilizing the hidden relationships acquired in the first step, similarity among the objects is calculated, and a user graph and an object graph are constructed for the similar users and the similar objects; a third step, a user-object graph model is constructed by utilizing a user graph model and an object graph model acquired in the second step and a bipartite graph of the users and the objects acquired through utilizing the historical scoring records of the users; and a fourth step, the access probability of objects without scoring record of each user is ordered in a descending mode by utilizing a random walk personalrank algorithm, and front N objects are acquired to form a recommendation list for recommendation to the users.

Description

A kind of personalized recommendation method based on figure

Technical field

The invention belongs to recommend method and technology field, relate to a kind of personalized recommendation method based on figure.

Background technology

Commending system (Recommender System, RS) be a kind of preference that can utilize user be actively user Recommendations or the system of project, it utilize the historical data of user to excavate user interest preference, thus will User may be pushed to specific user by article interested, and good commending system can bring considerable for businessman Economic benefit.

The composition of one complete commending system must include three key elements: user model, recommended model, Proposed algorithm.Wherein proposed algorithm is the core of commending system.At present, more ripe proposed algorithm mainly has: Recommendation based on collaborative filtering, implicit semantic model, recommendation based on graph model, combined recommendation etc..

The invention similar to the present invention: the personalized recommendation method based on figure under community network and system thereof, The invention of the present invention and Ta sums up to get up to have 3 differences:

(1) calculating of similarity.He is the similarity utilizing label for labelling information to calculate between user, this Bright is the user's class vector utilizing implicit semantic model to obtain user's history scoring record, utilizes this to ask The vector obtained, calculates the similarity between user.

(2) foundation of graph model.What he built is the non-directed graph model of cum rights, and does not accounts for article Between similarity, the present invention build be not set up between Weighted Coefficients, and similar article directly to connect Connect.

(3) speed of service problem.His patent does not accounts for carrying out the speed of service of random walk on figure Problem.This patent utilizes CUDA that personalrank algorithm has been carried out parallelization.

Wherein, implicit semantic model: utilize implicit semantic analytical technology, implicit theme can be found out or divide Class, sets up the contact between feature by implicit theme or classification.Common implicit semantic analytical technology Mainly there are LFM and LSI, LDA, TopicModel etc..These technology are the most all to lead in text mining Putting forward in territory, they were also constantly applied in other field in recent years, and had obtained good application Effect.

Proposed algorithm based on graph model: user behavior can represent with bipartite graph, recommends article to user u Task just can be converted into measure user vertex v_uAnd with v_uThere is no the article node that limit is joined directly together on figure Dependency, dependency the highest article weight in recommendation list is the biggest.

The opposite vertexes that general dependency is high has the feature that

1) the connected number of path between two summits is more；

2) path being connected between two summits is the most comparatively short；

3) between two summits, the out-degree on the summit of phase access path process is bigger.

Research worker devises the method for summit dependency in a lot of calculating figure at present.Wherein based on random walk Personalrank algorithm be one therein.

CUDA be 2006 by NVIDIA company release based on graphics processing unit (GPU) on the basis of The framework that the support universal parallel of exploitation calculates, thought is to make full use of CPU and GPU in the application Respective advantage.Play GPU and solve at a high speed the advantage of complicated calculations problem.And similar C language can be used High-level language writes concurrent program, is widely used in the numerical computations of every field.

CUDA programming is divided into host side (Host) and equipment end (Device) two parts, in a system A main frame and multiple equipment can be there is.In this programming model, CPU and GPU is mutually to work in coordination with work Make.CPU is responsible for carrying out the strong transaction of logicality and serial computing, and GPU is then absorbed in and performs height Threading parallel processing task.The serial in CPU of the program of host side runs, when running to core letter During number time (Kernel), then call GPU and perform, carry out the parallel processing of multithreading.

Present commending system is primarily present openness problem, cold start-up problem, scalability problem etc., Wherein Sparse Problems is owing to present data scale is huge, and the overlap selected between two users is less, scoring Sparse, causes the recommendation quality degradation of algorithm.Scalability problem is the fortune due to proposed algorithm Evaluation time is as user and the increase of number of items and what sharp increase caused.

Summary of the invention

The present invention provides a kind of personalized recommendation method based on figure, by utilizing implicit semantic model to obtain Result sets up the graph model between the graph model between user and article, set up as much as possible between user with And the implication relation between article, thus it is effectively reduced the openness impact on recommendation effect.

The present invention is achieved through the following technical solutions:

A kind of personalized recommendation method based on figure, comprises the following steps:

Step one, record of marking the history of user utilize implicit semantic model to calculate and respectively obtain between user And the implication relation between article；

Step 2, the implication relation utilizing step one to obtain calculate the similarity between user, and thing respectively Similarity between product, and for similar user between, and between similar article build figure, user With article as node, if the similarity degree between consumer articles is higher than predetermined threshold value, then set up a limit, Until constructing the graph model between user and between article, and by history scoring record construct user and Connection figure between article；

Step 3, utilize user's graph model and article graph model that step 2 obtains, and by the history of user User and the bipartite graph of article that scoring record obtains build user-article graph model；

Step 4, utilize personalrank algorithm based on random walk each user is not marked record The access probability of article carries out descending, takes top n article, forms recommendation list and recommends a user.

Further, the method using the method for solving from matrix angle to combine with CUDA parallelization carries The speed of service of high personalrank algorithm.

Detailed description of the invention

The invention will be described further below.

(1) implicit semantic analysis

The present invention uses implicit semantic model (LFM), and main thought is that the product utilizing two low-dimensional matrixes comes Represent user's rating matrix to article.Firstly the need of gathering user's history scoring record to article, then make With LFM, it is modeled, can obtain model as shown below:

R matrix is user*item matrix, and what matrix value Rij represented is the useri interest-degree to itemj, and this is just intended to The value asked.If LFM algorithm can extract Ganlei, as user and item from user to article scoring record Between connect bridge, R matrix table is shown as P matrix and Q matrix multiple.

$R_{UI}=P_U{Q}_I=\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{P}_{U,k}{Q}_{k,I}$

Wherein P matrix is user-class matrix, and what matrix value Pij represented is the useri interest-degree to classj； Q matrix form class-item matrix, what matrix value Qij represented is itemj weight in classi, and weight is more Gao Yueneng is as such representative.So LFM calculates the user U interest to article I according to equation below Degree:

For calculating parameter value in matrix P and matrix Q.Stochastic gradient descent method typically can be used to minimize Loss function seeks parameter.

$C(p,q)=\underset{(u,i)\∈Train}{\mathrm{\Σ}}(r_{ui}-\underset{f=1}{\overset{F}{\mathrm{\Σ}}}{p}_{uf}{q}_{if})+\mathrm{\λ}\left(\right|\left|p_u\right|{|}^2+\left|\right|q_i\left|{|}^2\right)$

Wherein λ (| | p_u||²+||q_i||²) it is the regularization term preventing over-fitting.Hidden feature number F and regularization parameter λ need to be obtained by experiment.

This step utilizes implicit semantic model, completely from the angle of data, is taken based on user behavior system The automatic cluster of meter.Implicit semantic analytical technology has a following four advantage:

1) classification of implicit semantic analytical technology is from the statistics to user behavior, represents user to taxonomy of goods View.

2) can be with the granularity of control tactics, the final classification number of setting is the biggest, and the granularity of classification will be the thinnest, Otherwise the granularity of classification is the thickest.

3) the counting user behavior decision article weight in each class can be passed through, so each article are not hard It is assigned to some apoplexy due to endogenous wind to property.

4) can provide each classification is different dimensions, is to be calculated by the historical data of user completely 's.

(2) Similarity Measure

This step is that the matrix P utilizing step (1) to obtain and matrix Q calculates the similarity between user respectively, And the similarity between article, and for similar user between, and between similar article build figure, User and article, as node, if the similarity degree between user's (article) is higher than a certain threshold value, are then built A vertical limit.

Wherein Similarity Measure use Euclidean distance computational methods, for two n-dimensional vector a (x11, x12 ... x1n) With b (x21, x22 ... x2n) between Euclidean distance.

(3) recommendation based on user-article graph model

This step is the user's graph model and article graph model utilizing step (2) to obtain, and going through by user User and the bipartite graph of article that commentary on historical events or historical records member record obtains build user-article graph model.Recycling The article that user is interested are predicted by personalrank algorithm.

The thought of this algorithm is that user u is from start node v_uProceed by random walk.When migration to a certain with During machine node, first determine that being to continue with migration still terminates migration and from start node v according to probability α_uStart Again migration.To continue, then the node equal probability pointed to from this node randomly choose a node conduct Next node of migration.So migration is to last, and the probability that each article node is accessed to can converge to one Above number, thus as the final access probability of article node.It is expressed as follows with formula:

$PR\left(i\right)=(1-\mathrm{\α})r_i+\mathrm{\α}\underset{j\∈\∈in\left(i\right)}{\mathrm{\Σ}}\frac{PR\left(j\right)}{\left|out\left(i\right)\right|}$

$r_i=\left\{\begin{array}{cc}1& i=u\\ 0& i\≠u\end{array}\right.$

Wherein, d is to continue with the probability of migration, and | out (i) | is the degree of node i, and PR (i) is that the access of node i is general Rate.

Relation between user and article can be set up by the recommendation of model based on figure more preferably more intuitively, more certainly Right generation Top-N recommendation results collection, but due to be that algorithm needs to be iterated on bipartite graph, and needs Will generate the convergence of corresponding PR value in each summit in figure, therefore time complexity is the highest.

Therefore, Personalrank is converted into the form of matrix operations, substitutes alternative manner.M is bipartite graph Transition probability matrix, i.e.

$M(i,j)=\frac{1}{\left|out\left(i\right)\right|}$

So, iterative formula can be converted into:

R=(1-α) r₀+αM^Tr

Solve:

R=(1-α) (1-α M^T)r₀

Have only to calculate once (1-α M^T)^-1, but need to be to 1-α M^TSparse matrix fast inversion.Therefore, This patent uses the programming technique of this parallelization based on GPU of CUDA to combine Gauss Jordan algorithm Solve inverse matrix (1-α M^T)^-1, solve the speed of service that personalrank algorithm utilizes continuous alternative manner to calculate Problem.

The thought of Gauss Jordan algorithm is: solution matrix A's is inverse, it is only necessary to unit matrix I is put on the right, Entirety carries out matrixing and obtains the left side is unit matrix, and the right is required.

Detailed design plan of based on CUDA parallelization personalrank be presented herein below:

Main frame (host) end false code is as follows:

A) distribution thread:

dim3thread(threads)；

dim3rBlock((int)ceil(columns/threads)+1)；

dim3cBlock((int)ceil(size*columns/threads)+1)；

B) to (1-α M^T) matrix and unit matrix I allocation space in the global storage of GPU:

cudaMalloc((void**)&devMatrix,size*columns*sizeof(float))；

C) by (1-α M^T) matrix and unit matrix I from memory copying to video memory:

cudaMemcpy(d_A,L,ddsize,cudaMemcpyHostToDevice)；

cudaMemcpy(dI,I,ddsize,cudaMemcpyHostToDevice)；

D) (1-α M is sought^T) inverse of a matrix:

E) matrix and unit matrix are copied back internal memory from video memory:

cudaMemcpy(matrix,devMatrix,size*columns*sizeof(float),

cudaMemcpyDeviceToHost)；

F) release GPU global storage space:

cudaFree(devMatrix)；

Equipment (device) end false code is as follows:

A) calling the rowExchange () in kernel, it is not 0 that this row diagonal element is exercised in exchange

B) call fixRows (), by full line divided by diagonal element, make diagonal element become 1

C) calling fixColumns () makes other elements of this row become 0

(4) Top-N recommends

This step be according to step (3) obtain general for the do not mark access of article of record of each user Rate carries out descending, takes top n article, forms recommendation list and recommends a user.

Claims (3)

1. a personalized recommendation method based on figure, it is characterised in that comprise the following steps:

Step one, record of marking the history of user utilize implicit semantic model to calculate and respectively obtain between user And the implication relation between article；

Step 2, the implication relation utilizing step one to obtain calculate the similarity between user, and thing respectively Similarity between product, and for similar user between, and between similar article build figure, user With article as node, if the similarity degree between consumer articles is higher than predetermined threshold value, then set up a limit, Until constructing the graph model between user and between article, and by history scoring record construct user and Connection figure between article；

Step 3, utilize user's graph model and article graph model that step 2 obtains, and by the history of user User and the bipartite graph of article that scoring record obtains build user-article graph model；

The article that step 4, utilization personalrank algorithm based on random walk are interested in user Being predicted, the access probability of the article of record of not marking each user carries out descending, takes front N Individual article, form recommendation list and recommend a user.

A kind of personalized recommendation method based on figure, it is characterised in that enter one Step ground, the method using the method for solving from matrix angle to combine with CUDA parallelization improves The speed of service of personalrank algorithm.

A kind of personalized recommendation method based on figure, it is characterised in that Further, personalrank algorithm is that user u is from start node v_uProceed by random walk, work as migration During to a certain random node, first determine that being to continue with migration still terminates migration and from initial according to probability α Node v_uStart migration again；To continue, then randomly choosing of the node equal probability pointed to from this node One node as next node of migration, such migration to last, the probability that each article node is accessed to Can converge to above a number, thus as the final access probability of article node.