CN105893585A - Label data-based bipartite graph model academic paper recommendation method - Google Patents

Abstract

The invention relates to a label data-based bipartite graph model academic paper recommendation method. The theme content of a paper is summarized by a label of the paper through brief meaning, so that the label data-based bipartite graph model academic paper recommendation method naturally adds information about the label in the paper into content information about the paper; a bipartite relational graph is established by combining a reference relationship of the paper and a relationship of paper collection relation of a user; a graph model method capable of efficiently solving the problem of academic paper recommendation cold boot is proposed by applying random walk algorithm of reboot. The model only adds a small part of similarity relationship, thus reducing resources consumed in a parameter optimization process. Furthermore, the model fully utilizes various information in data, especially, the information about the label in the paper, thus guaranteeing the accuracy of paper recommendation. According to the method, related experiments are carried out on a real data set, so that a better experimental result is obtained.

Description

The bigraph (bipartite graph) model scientific paper of a kind of combination tag data recommends method

Technical field

The present invention relates to data mining, information retrieval and commending system field, the two of a kind of combination tag data Portion's graph model scientific paper recommends implementation method.

Background technology

The form of the big multiplex scientific paper of the scientific payoffs of research worker carries out record in early days, permissible by consulting such paper Offer reference to researcher；In addition, scientific paper have recorded again the scientific achievement of contemporary, can be that researcher provides Reference.By consulting relevant scientific paper, researcher can avoid the duplication of labour of other researcheres, improves the speed of scientific research And benefit.Through the ages all successful researcheres, are all on the basis of extensively absorbing other researcher knowledge, are subject to Good inspiration and successful.Therefore, anyone is when being engaged in the academic activities of specific subject, or newly starts one During Task, will devote a tremendous amount of time, such subject correlative theses is comprehensively investigated, understand both at home and abroad This problem research conditions, if someone did or had people to do, and had been obtained for which achievement, the most not yet What the problem solved is, accomplishes to know what's what.Only so, just it is avoided that the duplication of labour, has innovated, advanced.

Along with fast development and the expansion of subject of digitalized network, scientific paper information is explosive growth.This makes to grind The person of studying carefully is increasingly difficult to find oneself paper interested wherein.Under such circumstances, paper proposed algorithm is arisen at the historic moment, can Well to recommend relevant scientific paper for researcher.But, most paper proposed algorithm does not use the mark of paper The data signed.Label summarises the purport of paper to a certain extent, reasonably adds label data information and can strengthen in paper Relation between appearance, can effectively solve to recommend in the middle of the cold start-up problem that faces, and then newly deliver for researcher recommendation Paper, so can make researcher be better understood by the technology of subject forefront.

Scientific paper is recommended in user oriented proposed algorithm, there is a kind of random walk restarted based on graph model Algorithm.Figure is a kind of form of expression presenting data, has some fixing positionalitys.In the drawings, under same principle, The information of multiple data can be shown easily.By with the node in figure and the weighting limit between them represent object and The relation existed between it.Additionally, the weights on weighting limit also may indicate that the intensity of the relation between object.Utilize graph model, The various information in data set can be utilized easily.Tian and Jing proposed a kind of based on bigraph (bipartite graph) model in 2013 Scientific paper recommend method.The similarity relation of user-paper relation, the similarity of user and paper is attached to by the method Together, paper interested is recommended for user.In the same year, Meng and Gao etc. proposes a kind of science opinion based on multilayer graph model Literary composition recommendation method.The method is passed through LDA (Latent Dirichlet Allocation) and is found out the topic model that paper is potential, and will It combines with author information, citation information and lexical information, and the various information in data that make full use of realize the opinion of personalization Literary composition is recommended.

Summary of the invention

Based on above-mentioned background technology, the present invention proposes the bigraph (bipartite graph) model of a kind of combination tag data, makes full use of data In various information, the label information in data is added in the content of paper, is ensureing on the basis of precision, the most quickly For researcher recommend scientific paper.Traditional scientific paper recommends method often to have ignored the label information of paper, but, learn The label information of art paper summarises the purport of paper with brief semanteme, finds oneself paper process interested helping user In play very important effect.By adding label information, the present invention can effectively strengthen the content association between paper, Can preferably represent under this relation, the mutual relation between scientific paper, effectively solve face in the middle of recommendation cold Starting problem, and then optimize whole algorithm, improve the precision recommended.The recommendation method using the present invention to provide, may apply to In the search system of paper, it is provided that the recommendation service newly published thesis or the precision improving recommendation.

The paper recommendation method that the present invention proposes is the mixed method of combination tag data bigraph (bipartite graph) model.By the label of paper Information organically combines with other information of paper, improves the precision recommended, and adds the most again the part similarity relation between paper, To ensure the efficiency recommended.Organizing the Heterogeneous Information of paper for convenience, the present invention needs to do some initializations, concrete steps It is:

1. vectorization label information

First label data is carried out denoising, removes the occurrence number label less than 5 in all papers.Secondly, statistics Remaining label data, the vector of one label of composition, the numbering of the line number correspondence paper of each of which row, each in vector Item represents that this label occurs the most in that article, occurs that then value is 1, and otherwise value is 0.As shown in formula (1):

$A_{i_t}=\{t_1,t_2,...,t_l\}---\left(1\right)$

The sum of label during l represents data in above formula.

2. vectorization papers contents information

Extract title and the summary info of paper, remove stop words, form paper text vector, the line number of each of which row The numbering of corresponding paper, in vector, each represents whether this vocabulary occurs in paper, if occurring, the value of correspondence position is 1, it is otherwise 0.As shown in formula (2):

$A_{i_w}=\{w_1,w_2,...,w_v\}---\left(2\right)$

3. integrate paper vector information and content information

Owing to label summarises the purport of paper with brief semanteme, such that it is able to help user preferably to find to grind with oneself Study carefully the paper that neighborhood is relevant.Its effect is similar to the keyword message of paper, and therefore the present invention integrates paper label vector sum paper Content vector mode, the label information of paper is added in the content information of paper, finally gives the characteristic vector of paper. As shown in formula (3):

$A_i=\{A_{i_w},W_T\·A_{i_t}\}---\left(3\right)$

W in its Chinese style_TRepresent label information weight in text vector.

4. calculate the similarity of scientific paper

According to the characteristic vector of paper, use the algorithm of cosine similarity, calculate the similarity between paper.Such as formula (4) Shown in:

$sim(A_i,A_j)=\frac{\left|A_i\·A_j\right|}{\left|A_i\right|\left|A_j\right|}---\left(4\right)$

5. build the bigraph (bipartite graph) model of combination tag data

1) the summit all users in data set and paper being seen as in figure, each user or paper are when in bigraph (bipartite graph) A summit.

2) according to the relation of consulting of user-paper, between structure user's vertex set and paper vertex set, limit contacts.If used Family U has collected paper A, then there is limit, the most not between user U summit and paper A summit corresponding in bigraph (bipartite graph) There is limit.As shown in formula (5):

3) according to the adduction relationship between paper, build the adduction relationship of paper-paper in two graphs of a relation, be used for adding hadron Internal relation between paper in figure.If paper A quotes paper B, then corresponding paper A summit and opinion in paper subgraph There is limit between B summit in literary composition, the most there is not limit.As shown in formula (6)

4) according to paper similarity calculated after combination tag information, using k nearest neighbor algorithm, before finding paper, K is individual Closest paper, then gives in bigraph (bipartite graph) model and adds a limit between corresponding summit.

6 present invention use the similarity in the Random Walk Algorithm calculating bigraph (bipartite graph) model restarted between summit, according to result Scientific paper is recommended for user.

1) represent the bigraph (bipartite graph) model of combination tag data with symbol G, M represents its adjacency matrix, and to adjacency matrix M carries out column criterion and obtains the probability transfer matrix of its regularization

2) on bigraph (bipartite graph) G, the Random Walk Algorithm restarted is used, as shown in formula (7):

${\stackrel{\→}{r}}^{(t+1)}=(1-c)\·\tilde{M}\·{\stackrel{\→}{r}}^t+c\·\stackrel{\→}{q}---\left(7\right)$

Wherein c is the probability restarted, and returns to the probability of starting point in iterative process i.e. every time.It is to restart vector, table Show original state.Restart vectorIn to take seed vertex value be 1, remaining is 0.Represent probability distribution in t step figure,Represent that t step is transferred to the probability of summit i by kind of son vertex.

3) in order to reduce time and the memory consumption of formula (7), BEAR (Block Elimination Approach is used For Random Walk with Restart on Large Graphs) algorithm, bigraph (bipartite graph) model vertices is rearranged, piecemeal meter Calculate each inverse of a matrix, obtain final result.

4) to final probability distributionSequence, finds out the Top N number of summit similar to planting son vertex.

Accompanying drawing explanation

Fig. 1 is the bigraph (bipartite graph) model of the combination tag data of the present invention；

Fig. 2 is that the present invention is on data set and other model test results times and the contrast of memory efficient；

Fig. 3 is that the present invention is on data set and the contrast of other model test results recall rates；

Fig. 4 is that the present invention is on data set and the contrast of other model test results success rates.

Detailed description of the invention

With reference to the accompanying drawings, and combine concrete data set, embodiments of the invention are described in detail.Hereinafter retouch The embodiment stated is merely exemplary, is served only for preferably explaining the present invention, it is simple to the research worker in field of the present invention is more preferable Understanding, it is impossible to be interpreted as limitation of the present invention.

The present invention is that the bigraph (bipartite graph) model scientific paper of a kind of combination tag data recommends method, mainly enters scientific paper Row is recommended.As it is shown in figure 1, the present invention comprises the following steps:

S1. data set introduction

Specific embodiment of the present invention uses data set to gather from CiteULike, specifically includes in data set 16980 scientific paper information, 5551 user list information, the label information of 46391 papers and 44709 papers draw By relation, the content information that wherein scientific paper is main includes its title and summary info.

S2. data prediction

Data prediction includes that text information processing, label information process and paper Similarity Measure three part.

1) text information processing

For title and the summary of scientific paper, after removing stop words therein, calculate its TF-IDF (term frequency- Inverse document frequency) value, and in descending order to its arrange, select front 8000 differ word composition vocabulary, In order it is numbered.Then according to vocabulary, by the vocabulary vector representation of every paper.Such as " 50 3:8 10:5 980:1 ... " this represents that the form of text message of paper vectorization, " 50 " represent in this paper vocabulary number altogether, " 3:8 " Represent that the vocabulary of numbered " 3 " occurs in that in this paper " 8 " are secondary.

2) label information processes

The paper label data that access times are less than 5 times is removed by the present invention, obtains 7386 different labels.According to Whole label data, by every paper composition label vector, such as " 10 4 578 7385 ... ", this represents the mark of paper vectorization The form of label information, " 10 " represent that the sum of label in this paper, " 4 " tag number are that the label of " 4 " goes out in this paper Existing.

3) paper Similarity Measure

By vocabulary and the tag combination of paper of paper, obtain 15386 different vocabulary, form all feature vocabulary Table.By 1), 2) combination of the paper vector that obtains, build paper characteristic vector.Such as " 60 3:8 10:5 980:1 ... 8004:k 8578:k 15385:k ... " this represent paper characteristic vector form, " 60 " represent the sum of all features in this paper, " 3:8 " Representing that the vocabulary of numbered " 3 " occurs in that in this paper " 8 " are secondary, " 8004:k " tag number is " 4 " (8004-8000) Label occur in this paper, wherein " k " represent label weight in paper.Afterwards according to the feature of final paper to Amount calculates the cosine similarity between paper.

S3 model training

All papers in data set are equally divided into 5 groups, in turn using one of which as test set, other 4 groups of conducts Training set.For training set, being classified as 5 parts equally, choose 1 part and do test set, 4 parts carry out five folding intersections for training set Checking.Determined the parameter of model by cross validation, the parameter choosing combination property best is predicted on test set, obtains 5 groups predict the outcome, then seek its average, as the estimation to algorithm performance.

S4 evaluation index

Recall rate be normally used for evaluate and test proposed algorithm precision, recall rate is the biggest, it is recommended that outcome quality the highest, recall The computing formula of rate is:

Owing to user is not interested to this paper or user does not knows this paper, zero during prediction all may be caused, Therefore accuracy rate may not apply in paper recommendation.Here with success@N as another evaluation index.It is defined as The top n user recommended finds the probability of the user of a necessary being.When certain paper being recommended N number of user exists One correct user, then success@N=1, on the contrary it is 0.Success@N is defined as:

The last present invention adds up all of recall@N and success@N, calculates last pre-as model of meansigma methods respectively Survey result.

Although detailed description of the invention illustrative to the present invention is described above, in order to the technology people of the art Member understands the present invention, it should be apparent that the invention is not restricted to the scope of detailed description of the invention, and the ordinary skill to the art From the point of view of personnel, as long as the thought that limits in appended claim of various change and in the range of determining, all utilize structure of the present invention The innovation and creation thought are all at the row of protection.

Claims (6)

1. the bigraph (bipartite graph) model scientific paper of combination tag data recommends a method, including related data pretreatment, combines mark The structure of the bigraph (bipartite graph) model signed and scientific paper recommend the realization of method.Concrete operation step is as follows:

Step a. preprocessed data collection, removes noise data therein；

The label information of paper is carried out vectorization process by step b.；

Papers contents information and paper label are combined by step c., calculate similarity between paper；

Step d. collects the adduction relationship structure between paper similarity and paper after paper relation, introducing label according to user Bigraph (bipartite graph) model；

Step e., on the bigraph (bipartite graph) model of final combination tag data, uses the Random Walk Algorithm restarted, and calculates node Between structural dependence.

Paper the most according to claim 1 recommends method, it is characterized in that step a, specifically includes in all data sets The occurrence number label data less than 5 times removes, and reduces label noise data, and rebuilds the relation that paper-label is subordinate to.

Recommendation method the most according to claim 1, is characterized in that step b, specifically includes the label information in statistics paper, The text vector of one label of composition, the numbering of the line number correspondence paper of each of which row, each this label of expression in vector Whether occurring occurring that then value is 1 in paper, otherwise value is 0.As shown in formula (1):

$A_{i_t}=\{t_1,t_2,\mathrm{...},t_l\}---\left(1\right)$

Recommendation method the most according to claim 1, is characterized in that step c, specifically includes

(c1) extract the title in scientific paper and summary info, build the content text information of paper.

(c2) the paper text message obtained is removed stop words, calculate its TF-IDF (term frequency-inverse Document frequency) value, according to calculated value, N item composition vocabulary before obtaining, and each vocabulary is carried out Numbering.

(c3) according to the vocabulary obtained, add up the text message of every paper, form paper text vector, each of which row The numbering of line number correspondence paper, in vector, each represents that whether this vocabulary occur in paper, if there is then correspondence position Value is 1, is otherwise 0.As shown in formula (2):

$A_{i_w}=\{w_1,w_2,...,w_v\}---\left(2\right)$

(c4) label information of paper is added in the content information of paper.

Owing to label summarises the content information of paper with brief semanteme, such that it is able to help user preferably to find to grind with oneself Study carefully the paper that neighborhood is relevant.Its effect is similar to the keyword message of paper, and therefore the present invention integrates paper label vector sum paper Content vector mode, the label information of paper is added in the content information of paper, finally gives the characteristic vector of paper. As shown in formula (3):

$A_i=\{A_{i_w},W_T\·A_{i_t}\}---\left(3\right)$ W in its Chinese style_TRepresent label information weight in text vector.

(c5) according to the characteristic vector of paper, the present invention uses cosine similarity to the similarity calculating between paper.Such as formula (4) shown in:

$sim(A_i,A_j)=\frac{\left|A_i\·A_j\right|}{\left|A_i\right|\left|A_j\right|}---\left(4\right)$

Recommendation method the most according to claim 1, is characterized in that step d, specifically includes:

(d1) according to paper corresponding in user list, the summit that each user and every paper are seen as in figure by the present invention, For building two graphs of a relation of user-paper.

(d2) according to the relation of consulting of user-paper, between structure user's vertex set and paper vertex set, limit contacts.If used Family U has collected paper A, then there is limit, the most not between user U summit and paper A summit corresponding in bigraph (bipartite graph) There is limit.As shown in formula (5):

(d3) according to the adduction relationship between paper, build the adduction relationship of paper-paper in two graphs of a relation, be used for adding hadron Internal relation between paper in figure.If paper A quotes paper B, then corresponding paper A summit and opinion in paper subgraph There is limit between B summit in literary composition, the most there is not limit.As shown in formula (6)

(d4) according to paper similarity calculated after combination tag information, using k nearest neighbor algorithm, before finding paper, K is individual Closest paper, then gives in bigraph (bipartite graph) model and adds a limit between corresponding summit.

Recommendation method the most according to claim 1, is characterized in that step e, specifically includes

(e1) the bigraph (bipartite graph) model symbol G of the combination tag data obtained according to right 5 represents, M represents its adjacent square Battle array, and adjacency matrix M is carried out column criterion obtain the probability transfer matrix of its regularization

(e2) on bigraph (bipartite graph) G, the Random Walk Algorithm restarted is used, as shown in formula (7):

${\stackrel{\→}{r}}^{(t+1)}=(1-c)\·\tilde{M}\·{\stackrel{\→}{r}}^t+c\·\stackrel{\→}{q}---\left(7\right)$

Wherein c is the probability restarted, and returns to the probability of starting point in iterative process i.e. every time.It is to restart vector, represents initial State.Restart vectorIn to take seed vertex value be 1, remaining is 0.Represent probability distribution in t step figure,Represent T step is transferred to the probability of summit i by kind of son vertex.

(e3) in order to reduce time and the memory consumption of formula (7), BEAR (Block Elimination Approach is used For Random Walk with Restart on Large Graphs) algorithm, bigraph (bipartite graph) model vertices is rearranged combination, point Block calculates the inverse of each submatrix, iteration, untilConvergence, representative points is stable with other each summit in figure by the time Probability distribution.

(e4) to final probability distributionSequence, finds out the N number of summit of the Top similar to representative points.