CN103886067B - Method for recommending books through label implied topic - Google Patents

Abstract

The invention discloses a method for recommending books through a label implied topic. The method comprises the steps that the books are used as documents, book labels are used as words in the documents, topic modeling is conducted on the book labels through the LDA-Gibbs algorithm, a label-topic model is obtained, the corresponding relation of users and the labels is obtained according to book reading records of the users, a user-topic model is obtained through the LDA-inference algorithm, the users with the similar interests are found according to similarity of the users on topic distribution, and the books are collaboratively filtered and recommended. The method for recommending the books through the label implied topic sufficiently mines the semantic information of the book labels, reduces dimensionality representing the requirements of the users by using the topic, reduces calculated amount, is beneficial to improving the quality of recommended results, and has the certain practical value.

Description

Imply the method that theme carries out book recommendation using label

Technical field

The present invention relates to Artificial is intelligent, more particularly, to a kind of use label implies the side that theme carries out book recommendation Method.

Background technology

Enter the web2.0 epoch with the Internet, everybody can become the supplier of content, and this makes the Internet be carried Quantity of information more and more huger.Far beyond the treatable amount of user institute, user will therefrom look for the information of these magnanimity It is extremely difficult to the information useful to oneself, often will take a substantial amount of time and energy.So, how to make user accurate And rapidly find oneself required resource, become internet information supplier problem in the urgent need to address.In this back of the body Under scape, personalized recommendation technology is arisen at the historic moment.By using personalized recommendation technology, website can effectively improve and take quality and effect Rate.It is not only does this facilitate and solve the problems, such as that the network information is spread unchecked moreover it is possible to avoid the unification of website service, therefore can be both permissible Prevent the loss of user, more users can also be attracted.

In ongoing research area, personalized recommendation algorithm can be largely classified into two kinds: content-based recommendation method And collaborative filtering (collaborative filtering) (content-based).Content-based recommendation algorithm firstly the need of Calculate the similarity between the chosen or used project of user and project to be recommended, then project to be recommended according to phase It is ranked up like degree size, the higher project of similarity, preferential recommendation is to user.The commending system being currently based on content can lead to Cross analysis user and the chosen or used content of user, user and project are set up respectively and retouches for its content characteristic State file.System directly can be recommended to describe literary composition with it to user by comparing the similarity of user and the description file of project The higher project of part similarity.Collaborative Filtering Recommendation Algorithm is the algorithm typically utilizing group wisdom.Entered using collaborative filtering When row is recommended, system finds with certain specific user at this to the selection of each project or usage record according to all users first Relatively more similar customer group in a little behaviors, that is, find the customer group similar with targeted customer's interest, then used according to this The selected or used project of family group is recommended.

Content of the invention

The purpose of the present invention is for this application of digital library, compensate for traditional collaborative filtering shortage right The deficiency of the utilization of this valuable source of book labels, provides a kind of use label to imply the method that theme carries out book recommendation.

Comprised the following steps using the method that the implicit theme of label carries out book recommendation:

1) obtain book labels data set from data base, the label that is, each books have；Obtain from server log Obtain the books reading record of user, the books that is, each user was read；

2) label-topic model is set up using lda-gibbs algorithm according to book labels data set；

3) the books reading record according to user and books and the corresponding relation of label, obtain user-label data collection；

4) according to user-label data collection and label-topic model, set up user-master using lda-inference method Topic model；

5) when producing recommendation for certain specific user, found and this use according to the user having built up-topic model The similar user of the theme distribution at family, i.e. nearest neighbor；

6) obtain the candidate's books for recommending from the list of read book of nearest neighbor, find out nearest neighbor and read N this books of recommending user most like with the theme distribution of specified user in the book crossed.

Described step 2) include: first books are regarded as document, label regards word as it is assumed that having m piece document, corpus V word, all of word and corresponding theme is had to represent in the following way:

$\stackrel{\→}{w}=({\stackrel{\→}{w}}_1,...,{\stackrel{\→}{w}}_m)$

$\stackrel{\→}{z}=({\stackrel{\→}{z}}_1,...,{\stackrel{\→}{z}}_m)$

Wherein,Represent the word of m piece document,Represent that the corresponding theme of these words is numbered, such as w_1,2Represent the 1st The 2nd word in piece document, then z_1,2Represent the theme meaning that corresponding to this word, using lda-gibbs algorithm, w is entered The implicit Topics Crawling of row, θ andIt is the matrix being exported by algorithm as a result, θ is m × k dimension matrix, every a line represents certain this figure Distribution on k theme for the book, i.e. p (topic | doc),It is k × v dimension matrix, every a line represents in certain theme k occur respectively The probability of individual label, in lda-gibbs algorithm, the more new regulation of gibbs iteration sampling is

Wherein,Represent the word being designated as i under removing,Represent and belong to after removing i-th word in m piece document The word number of theme k,Represent the number being designated as word t after the word of i under removing in k-th topic,Represent in the case that the theme of other all words determines, under be designated as the word of i and belong to the bar of theme k Part probability, α_kAnd β_tFor the predefined parameter in lda model.

The formula of parameter calculating lda model is

${\mathrm{\θ}}_{m,k}=\frac{n_m^{\left(k\right)}+{\mathrm{\α}}_k}{\underset{k=1}{\overset{k}{\mathrm{\σ}}}{n}_m^{\left(k\right)}+{\mathrm{\α}}_k}$

Wherein, θ_m,kMean that document m belongs to the probability of theme k,Represent the probability that in theme k, word t occurs, Represent the number of the word belonging to k-th theme in m piece document,Represent the number of word t in k-th theme.

Described step 4) includes: by user tag set tu_iAs document w, the label in set is then as in document Word t, document is carried out with theme modeling, that is, distribution σ on theme for the document to be obtained, then needs to use lda- Inference algorithm estimates unknown parameter σ, and the sampling more new regulation of lda-inference algorithm is as follows

WhereinIt is the document being made up of user tag tu,Representing matrixRow k t row, represent theme k in The probability of word t occurs, by lda-inference, has obtained distribution σ on each theme for the user, every a line of σ represents Probability distribution on k theme for certain user, obtains user-topic model.

Described step 5) includes: matrix σ gives low-dimensional on k theme for the user and represents, each of σ is worth generation Probability on certain theme for certain user by table, and matrix is as follows

$\mathrm{\σ}=\left[\begin{array}{cccc}{p}_{\mathrm{1,1}}& p_{\mathrm{1,2}}& ...& p_{1,k}\\ p_{\mathrm{2,1}}& p_{\mathrm{2,2}}& ...& p_{2,k}\\ ...& ...& ...& ...\\ p_{n,1}& p_{n,2}& ...& p_{n,k}\end{array}\right],$

p_{N, k}Represent user u_nIn theme z_kOn probability, and had according to the property of probabilityBy Probability p_{N, k} It is interpreted as user u_nTo theme z_kFavorable rating or scoring, the similarity between user, cosine are weighed using cosine similarity Similarity method regards user as the vector on k dimension space to the scoring of each theme, ifThat Calculating formula of similarity between user i and user j is as follows

$\mathrm{sim}(u_i,u_j)=\cos (\stackrel{\→}{i},\stackrel{\→}{j})=\frac{\stackrel{\→}{i}\·\stackrel{\→}{j}}{\left|\right|\stackrel{\→}{i}\left|\right|\·\left|\right|\stackrel{\→}{j}\left|\right|}$

According to above-mentioned Similarity measures formula, calculate the acquaintance tolerance of all users and active user, then current User u_iArest neighbors collection be combined into and be designated as

neighbors(u_i)={ u_j|sim(u_i, u_j)≥threshold}

Wherein threshold is a threshold value set in advance, and the similarity of two users is just more than or equal to during this threshold value It is considered similar neighborhood.

The present invention compared with prior art has the advantages that

1. the present invention has fully excavated the semantic information of book labels；

2. the present invention can still provide high-quality recommendation during user's read books negligible amounts；

3. the present invention calculates similarity on theme distribution for the user, reduces operand.

Brief description

Fig. 1 is the schematic diagram setting up label-topic model according to book labels；

Fig. 2 is the schematic diagram that books reading record according to user and label-topic model set up user-topic model；

Fig. 3 (a) is the books reading record of user；

Fig. 3 (b) is the books to user for the system recommendation.

Specific embodiment

Comprised the following steps using the method that the implicit theme of label carries out book recommendation:

1) obtain book labels data set from data base, the label that is, each books have；Obtain from server log Obtain the books reading record of user, the books that is, each user was read；

2) label-topic model is set up using lda-gibbs algorithm according to book labels data set；

3) the books reading record according to user and books and the corresponding relation of label, obtain user-label data collection；

4) according to user-label data collection and label-topic model, set up user-master using lda-inference method Topic model；

5) when producing recommendation for certain specific user, found and this use according to the user having built up-topic model The similar user of the theme distribution at family, i.e. nearest neighbor；

6) obtain the candidate's books for recommending from the list of read book of nearest neighbor, find out nearest neighbor and read N this books of recommending user most like with the theme distribution of specified user in the book crossed.

Described step 2) include: first books are regarded as document, label regards word as it is assumed that having m piece document, corpus V word, all of word and corresponding theme is had to represent in the following way:

$\stackrel{\→}{w}=({\stackrel{\→}{w}}_1,...,{\stackrel{\→}{w}}_m)$

$\stackrel{\→}{z}=({\stackrel{\→}{z}}_1,...,{\stackrel{\→}{z}}_m)$

Wherein,Represent the word of m piece document,Represent that the corresponding theme of these words is numbered, such as w_1,2Represent the 1st The 2nd word in piece document, then z_1,2Represent the theme meaning that corresponding to this word, using lda-gibbs algorithm, w is entered The implicit Topics Crawling of row, θ andIt is the matrix being exported by algorithm as a result, θ is m × k dimension matrix, every a line represents certain this figure Distribution on k theme for the book, i.e. p (topic | doc),It is k × v dimension matrix, every a line represents in certain theme k occur respectively The probability of individual label, in lda-gibbs algorithm, the more new regulation of gibbs iteration sampling is

Wherein,Represent the word being designated as i under removing,Represent and belong to after removing i-th word in m piece document The word number of theme k,Represent the number being designated as word t after the word of i under removing in k-th topic,Represent in the case that the theme of other all words determines, under be designated as the word of i and belong to the bar of theme k Part probability, α_kAnd β_tFor the predefined parameter in lda model.

The formula of parameter calculating lda model is

${\mathrm{\θ}}_{m,k}=\frac{n_m^{\left(k\right)}+{\mathrm{\α}}_k}{\underset{k=1}{\overset{k}{\mathrm{\σ}}}{n}_m^{\left(k\right)}+{\mathrm{\α}}_k}$

Wherein, θ_m,kMean that document m belongs to the probability of theme k,Represent the probability that in theme k, word t occurs, Represent the number of the word belonging to k-th theme in m piece document,Represent the number of word t in k-th theme.

Described step 4) includes: by user tag set tu_iAs document w, the label in set is then as in document Word t, document is carried out with theme modeling, that is, distribution σ on theme for the document to be obtained, then needs to use lda- Inference algorithm estimates unknown parameter σ, and the sampling more new regulation of lda-inference algorithm is as follows

WhereinIt is the document being made up of user tag tu,Representing matrixRow k t row, represent theme k in The probability of word t occurs, by lda-inference, has obtained distribution σ on each theme for the user, every a line of σ represents Probability distribution on k theme for certain user, obtains user-topic model.

Described step 5) includes: matrix σ gives low-dimensional on k theme for the user and represents, each of σ is worth generation Probability on certain theme for certain user by table, and matrix is as follows

$\mathrm{\σ}=\left[\begin{array}{cccc}{p}_{\mathrm{1,1}}& p_{\mathrm{1,2}}& ...& p_{1,k}\\ p_{\mathrm{2,1}}& p_{\mathrm{2,2}}& ...& p_{2,k}\\ ...& ...& ...& ...\\ p_{n,1}& p_{n,2}& ...& p_{n,k}\end{array}\right],$

p_nkRepresent user u_nIn theme z_kOn probability, and had according to the property of probabilityBy Probability p_nk It is interpreted as user u_nTo theme z_kFavorable rating or scoring, the similarity between user, cosine are weighed using cosine similarity Similarity method regards user as the vector on k dimension space to the scoring of each theme, ifThat Calculating formula of similarity between user i and user j is as follows

$\mathrm{sim}(u_i,u_j)=\cos (\stackrel{\→}{i},\stackrel{\→}{j})=\frac{\stackrel{\→}{i}\·\stackrel{\→}{j}}{\left|\right|\stackrel{\→}{i}\left|\right|\·\left|\right|\stackrel{\→}{j}\left|\right|}$

According to above-mentioned Similarity measures formula, calculate the acquaintance tolerance of all users and active user, then current User u_iArest neighbors collection be combined into and be designated as

neighbors(u_i)={ u_j|sim(u_i, u_j)≥threshold}

Wherein threshold is a threshold value set in advance, and the similarity of two users is just more than or equal to during this threshold value It is considered similar neighborhood.

Embodiment

As shown in Figure 3, give and imply, using label, the application example that theme carries out book recommendation.Under Face describes, with reference to the method for this technology, the concrete steps that this example is implemented in detail, as follows:

(1) label-topic model is set up using lda-gibbs algorithm according to the book labels data in data base, obtain Each label belongs to the probability of each theme.

(2) obtain the books reading record of user according to server log, books reading record such as Fig. 3 of user in this example Shown in (a).

(3) the books reading record according to user and label-thematic relation set up user-topic model, obtain user and exist Probability distribution on each theme.

(4) find the user similar to the theme distribution of this user, constitute the nearest neighbor of this user.

(5) obtain the candidate's books for recommending in the list of read book of nearest neighbor, find out nearest neighbor and read 24 books recommending user most like with the theme distribution of specified user in the book crossed, Fig. 3 (b) illustrates this 24 books In first 6.

Claims (2)

1. a kind of implicit theme of use label carries out the method for book recommendation it is characterised in that comprising the following steps:

1) obtain book labels data set from data base, the label that is, each books have；Obtain use from server log The books reading record at family, the books that is, each user was read；

2) label-topic model is set up using lda-gibbs algorithm according to book labels data set；

3) the books reading record according to user and books and the corresponding relation of label, obtain user-label data collection；

4) according to user-label data collection and label-topic model, set up user-theme mould using lda-inference method Type；

5) when producing recommendation for certain specific user, found with this user's according to the user having built up-topic model The similar user of theme distribution, i.e. nearest neighbor；

6) obtain the candidate's books for recommending from the list of read book of nearest neighbor, find out what nearest neighbor was read N this books of recommending user most like with the theme distribution of specified user in book；

Described step 2) include: first books are regarded as document, label regards word as it is assumed that there being m piece document, and corpus has v Individual word, all of word and corresponding theme represent in the following way:

$\stackrel{\→}{w}=({\stackrel{\→}{w}}_1,...,{\stackrel{\→}{w}}_m)$

$\stackrel{\→}{z}=({\stackrel{\→}{z}}_1,...,{\stackrel{\→}{z}}_m)$

Wherein,Represent the word of m piece document,Represent that the corresponding theme of these words is numbered, such as w_1,2Represent the 1st The 2nd word in document, then z_1,2Represent the theme meaning that corresponding to this word, using lda-gibbs algorithm, w is carried out Implicit Topics Crawling, θ andIt is the matrix being exported by algorithm as a result, θ is m × k dimension matrix, every a line represents certain this books Distribution on k theme, i.e. p (topic | doc),It is k × v dimension matrix, every a line represents in certain theme k occur each The probability of label, in lda-gibbs algorithm, the more new regulation of gibbs iteration sampling is

Wherein,Represent the word being designated as i under removing,Represent and belong to theme after removing i-th word in m piece document The word number of k,Represent the number being designated as word t after the word of i under removing in k-th topic,Represent in the case that the theme of other all words determines, under be designated as the word of i and belong to theme k Conditional probability, α_kAnd β_tFor the predefined parameter in lda model；

The formula of parameter calculating lda model is

${\mathrm{\θ}}_{m,k}=\frac{h_m^{\left(k\right)}+{\mathrm{\α}}_k}{\underset{k=1}{\overset{k}{\σ}}{n}_m^{\left(k\right)}+{\mathrm{\α}}_k}$

Wherein, θ_m,kMean that document m belongs to the probability of theme k,Represent the probability that in theme k, word t occurs,Represent The number of the word of k-th theme is belonged in m piece document,Represent the number of word t in k-th theme；

Described step 4) specifically include: by user tag set tu_iAs document w, the label in set is then as in document Word t, carries out theme modeling to document, that is, distribution σ on theme for the document to be obtained, and then needs to use lda- Inference algorithm estimates unknown parameter σ, and the sampling more new regulation of lda-inference algorithm is as follows:

WhereinIt is the document being made up of user tag tu,Representing matrixRow k t row, represent occur in theme k single The probability of word t, by lda-inference, has obtained distribution σ on each theme for the user, every a line of σ represents certain use Probability distribution on k theme for the family, obtains user-topic model.

2. a kind of use label according to claim 1 imply theme carry out book recommendation method it is characterised in that: institute The step 5 stated) include: matrix σ gives low-dimensional on k theme for the user and represents, each of σ value represents certain use Probability on certain theme for the family, matrix is as follows

$\mathrm{\σ}=\left[\begin{array}{cccc}{p}_{1,1}& p_{1,2}& \mathrm{...}& p_{1,k}\\ p_{2,1}& p_{2,2}& \mathrm{...}& p_{2,k}\\ \mathrm{...}& \mathrm{...}& \mathrm{...}& \mathrm{...}\\ p_{n,1}& p_{n,2}& \mathrm{...}& p_{n,k}\end{array}\right],$

p_{N, k}Represent user u_nIn theme z_kOn probability, and had according to the property of probabilityBy Probability p_{N, k}Understand For user u_nTo theme z_kFavorable rating or scoring, the similarity between user is weighed using cosine similarity, cosine is similar Property method regards user as the vector on k dimension space to the scoring of each theme, ifSo use Calculating formula of similarity between family i and user j is as follows

$sim(u_i,u_j)=cos(\stackrel{\→}{i},\stackrel{\→}{j})=\frac{\stackrel{\→}{i}\·\stackrel{\→}{j}}{\left|\right|\stackrel{\→}{i}\left|\right|\·\left|\right|\stackrel{\→}{j}\left|\right|}$

According to above-mentioned Similarity measures formula, calculate the acquaintance tolerance of all users and active user, then active user u_i Arest neighbors collection be combined into and be designated as

neighbors(u_i)={ u_j|sim(u_i, u_j)≥threshold}

Wherein threshold is a threshold value set in advance, and the similarity of two users is considered as when being more than or equal to this threshold value It is similar neighborhood.