CN107133277B - A Tourist Attraction Recommendation Method Based on Dynamic Topic Model and Matrix Factorization - Google Patents

Abstract

The invention discloses a tourist attraction recommendation method based on a dynamic theme model and matrix decomposition, which obtains the change of the user's travel preference by analyzing the user's travel history information in different time intervals, and provides fine-grained travel recommendation services for the user. The method first obtains the information of the dataset from the social network. Second, a dynamic topic model is used to mine the implicit feature information of users and attractions from user history information. Thirdly, through the analysis of the data set, the explicit feature information of users and attractions is mined, and the user-user and attraction-attraction similarity information is obtained by combining the implicit feature information of users and attractions. Finally, the user-user and attraction-attraction similarity information is fused using a matrix factorization method with a joint regularization term. This method can obtain the change of the user's travel preference and recommend suitable tourist attractions for the user.

Description

A Tourist Attraction Recommendation Method Based on Dynamic Topic Model and Matrix Factorization

technical field

The invention relates to the technical field of information recommendation, in particular to a tourist attraction recommendation method based on a dynamic theme model and matrix decomposition.

Background technique

In recent years, with the rapid development of mobile Internet, smart phones, and photo-sharing websites (such as: Flickr, Panoramio, and Instagram, etc.), a large amount of photo data with geographic location information has appeared on the Internet. Moreover, the number of photos with geographical location information contributed by these groups shows a trend of rapid growth. Based on these photos with geographic location information, it is possible to mine tourist attractions in the city, obtain popular tourist routes in the city, analyze tourists' travel preferences, and further provide users with personalized tourist attractions or tourist route recommendation services.

The current tourist attraction recommendation method based on photo mining with geographic location information usually directly uses the number of times users visit the scenic spot to obtain the similarity information between users, and combines the user-based collaborative filtering method to recommend scenic spots for users. However, due to the limitation of travel time or money, users usually only visit a small number of scenic spots in tourist cities, which leads to the problem of data sparsity in the modeling process of the recommendation system based on user-attraction matrix.

In order to solve the data sparsity problem in the above-mentioned scenic spot recommendation, there are currently scenic spot recommendation methods based on data dimensionality reduction, such as static topic models. This model is a relatively popular method in the field of text mining to obtain hidden topics in documents. In this model, the user's travel history is regarded as documents, and tourist attractions are regarded as words. Through this model, topic probability distributions of users and attractions can be obtained. However, when the static topic model obtains the topic probability distribution of the user's travel preference based on the user's travel history, the user's travel history information in all time periods (for example: year) is regarded as a document, and the user's travel history in different time periods is ignored. The problem of changing preferences.

Dynamic topic model is an extended form to obtain document topic changes based on static topic model. It divides the document set into sub-document sets according to the time period, assuming that the theme of each sub-document set has a front-to-back dependency, and the theme evolves with time. Through this model, the topic probability distribution of documents and words in different time periods can be obtained, and these topics in different time periods reflect the topic evolution process of documents and words. This model provides the possibility to solve the problem that the static topic model cannot obtain the change of user's travel preference in travel recommendation.

Contents of the invention

The problem to be solved by the present invention is how to obtain the user's travel preference change information through the dynamic theme model, and provide the user with fine-grained tourist attraction recommendation service.

A tourist attraction recommendation method based on dynamic topic model and matrix decomposition, including:

Obtaining the dataset information stage: Obtain the photo dataset D _photo from the social network and perform denoising processing on it to obtain the travel photo dataset D _photo-travel , then check-in the data D _check-in from the social network, and extract the check-in data The category information of the check-in place in D _check-in , and the category data set D _category of the check-in place is obtained;

Obtaining the user's travel preference stage: using the dynamic topic model to extract the implicit features of users and attractions from the tourist photo dataset D _photo-travel ; then, through the analysis of the tourist photo dataset D _photo-travel , the statistics of the users and attractions are obtained. The explicit features are combined with the implicit features of users and attractions to obtain user-user and attraction-attraction similarity information; finally, the user-user and attraction-attraction similarity information is fused using a matrix decomposition method with a joint regularization term, Complete the sparse user-attraction matrix Y to obtain a matrix Y' containing user travel preference information;

Recommend tourist attractions stage: use matrix Y' to score the attractions in the candidate set, and recommend the top N tourist attractions to users.

The concrete steps of the described stage of obtaining data set information are:

(1-1) Utilize the public API of photo sharing website to download the photo data with geographic location information in the tourist city, form photo data set D _photo ;

(1-2) utilize entropy-based mobility method to filter the non-tourist photos in the photo data set D _photo , remove the noise photos in the photo collection, and obtain the tourist photo data set D _photo-travel ;

(1-3) Utilize the public API of the location-based social media website to download the check-in data D _check-in of the user in the tourist city;

(1-4) Extract the category information of the check-in places in the check-in data D _check-in , and count the category data set D _category that constitutes the check-in places.

The photo data in the photo data set D _photo includes: photo identification information, photo shooting time, latitude and longitude information of the photo shooting location, text description information added by the user for the photo, and photo uploader identification information.

The check-in data in the check-in data D _check-in includes: identification information of the user's check-in behavior, time information of the user's check-in, latitude and longitude information of the check-in location, category information of the check-in location, and identification information of the check-in user.

The specific steps in the stage of obtaining the user's travel preference are as follows:

(2-1) Utilize density-based clustering method to carry out spatial clustering to the photos in the tourist photo data set D _photo-travel , obtain tourist attraction collection L;

(2-2) according to the number of times that the user visits the scenic spot obtained from the tourist photo data set D _photo-travel statistics, the tourist attraction in the tourist attraction collection L, construct the user-attraction matrix Y;

(2-3) The document represents the user's travel history, and the word represents the tourist attraction. The dynamic topic model is used to infer the potential topic probability distribution of the user and the scenic spot in different time periods, and the topic probability distribution of all time periods is combined to obtain all the latent topics of the user. formula features and the full implicit features of the sights

(2-4) Extract explicit features of users from the travel photo dataset D _photo-trave and the category dataset D _categorye of check-in locations and explicit features of attractions Then combine all the implicit characteristics of the user and the full implicit features of the sights Establish user portrait F _user and scenic spot portrait F _location , and use cosine function to construct m×m user-user similarity matrix A and n×n scenic spot-sightseeing similarity matrix B;

(2-5) According to the established Y, A, and B, utilize the user-user and attraction-attraction similarity relationship to construct a matrix decomposition model with joint regularization items, and complete the decomposition of Y;

(2-6) Complement the sparse user-sightseeing matrix Y according to the result of matrix decomposition with joint regularization items, and obtain a matrix Y' containing user travel preference information.

The concrete steps of described step (2-2) are:

(2-2-1) Extract all the visit information v=(l, u, t) from the tourist photo data set D _photo-trave , wherein, v represents that user u visits tourist attraction l at time t;

(2-2-2) count the number of times each user visits each scenic spot and the total number m of users in the tourist photo data set D _photo-trave according to all visit information v=(l, u, t);

(2-2-3) According to the number of times each user visits each scenic spot, the total number of users m, and a total of n tourist attractions in the tourist attraction set L, construct a user-attraction matrix Y, where Y∈R ^{m× n} , the value at position (i, j) in the matrix Y is the number of times the i-th user visits the j-th scenic spot.

The concrete steps of described step (2-3) are:

(2-3-1) Slice all the access information v=(l, u, t) in the travel photo data set D _photo-trave according to the same time length, and obtain sub-data sets corresponding to each time period, with a total of M indivual;

(2-3-2) The sub-dataset is used as the input of the dynamic topic model, and the topic probability distribution of users and scenic spots in different time periods is obtained through training,

in, is the topic probability distribution of the user in the Tth time period, The theme probability distribution of the scenic spot in the Tth time period, k is the number of themes;

(2-3-3) Concatenate the user's topic probability distributions in all time periods together in time to form all the implicit features of the user Concatenate the theme probability distributions of attractions in all time periods together in time to form all the implicit features of attractions

The concrete steps of described step (2-4) are:

(2-4-1) Make statistics on all the visit information v=(l, u, t) in the tourist photo data set D _photo-trave and the category data set D _category of the check-in location, and obtain the explicit characteristics of the user and explicit features of attractions Among them, r is the total number of explicit features of the user, and s is the total number of explicit features of the scenic spot;

(2-4-2) Combine the user's explicit features and implicit features to build a user portrait Combine the explicit and implicit features of the scenic spot to construct a portrait of the scenic spot

(2-4-3) Combine the cosine formula to obtain the user-user similarity matrix A:

Among them, f _pi and f _qi represent the i-th explicit feature of users p and q, respectively;

Also use the cosine formula to obtain the attraction-attraction similarity matrix B. At this time, the cosine formula is:

Among them, f _x ′ _i and f _y ′ _i denote the ith explicit features of scenic spots x and y, respectively.

In step (2-5), in the decomposition process of matrix Y, the similarity information of A and B is used as an additional regular term to restrict the decomposition of Y. The specific objective function is:

Among them, R _ij is the value at the position (i, j) in the matrix Y, I _ij indicates whether the user i visits the scenic spot j or not, if it visits, its value is 1, otherwise it is 0, Sim _ig indicates user i and The similarity information between users g, Sim _jq represents the similarity information between scenic spot j and scenic spot q; U _i represents the latent feature vector of user i, U _g represents the latent feature vector of user g, L _j represents is the latent feature vector of scenic spot j, L _g is the latent feature vector of scenic spot g, Indicates the distance between the potential feature vector of user i and the latent feature vector of user g, G(i) represents the similarity user group of user i, Q(j) represents the similarity scenic spot group of scenic spot j, and U is the decomposed The potential feature vector of the user is d×m-dimensional; L is the potential feature vector of the scenic spot after Y decomposition, which is d×n-dimensional; where m, n, and d represent the number of users, the number of scenic spots, and and the number of latent feature vectors;

The specific steps of decomposing Y based on the matrix factorization model with joint regularization items are as follows:

(a) randomly initialize parameter U and parameter L, and set learning rate α, error threshold δ, parameters λ ₁ and λ ₂ ;

(b) For each non-zero value R _ij in Y, according to Calculate the estimated value X _ij of R _ij , and according to Calculate the error between X _ij and the real value R _ij , and finally according to Statistics of the total error θ for all non-zero values, where w is the number of non-zero values;

(c) Judging whether the total error θ is greater than the error threshold δ, if so, perform step (d), if not, the iteration ends, U and L at this time are optimal values, and the decomposition of the matrix Y is completed;

(d) Use the gradient descent method to update the values of U and L, and then jump to step (b). The formula of the gradient descent method is:

In step (2-6), according to the formula Fill in the missing values in Y, and the filled values represent the user's travel preference information.

The specific steps in the stage of recommending tourist attractions are as follows:

(3-1) According to the information input by the user, the collection of scenic spots of the user in the tourist city is obtained, and these scenic spots are used as candidate sets for recommendation;

(3-2) According to the matrix Y', obtain the user's scoring to the recommended candidate concentrated scenic spots, and obtain the user's preferred scenic spots;

(3-3) Arrange the scores of the preferred attractions in descending order, and select the top N tourist attractions with the highest scores to recommend to the user.

The present invention aims at the theme evolution problem that the traditional static theme model is not enough to obtain the user's travel preference, and proposes a tourist attraction recommendation method based on the dynamic theme model and matrix decomposition. Compared with the existing method, its advantages are:

(1) Use the dynamic topic model to obtain the topic change information (implicit feature information) of the user's travel preference.

(2) Through the analysis of the data set information, a large amount of explicit feature information of users and attractions is obtained, which together with implicit feature information can fully describe the characteristics of users and attractions.

(3) Fusion of user-user and attraction-attraction similarity information through matrix decomposition method with joint regularization, this method can limit the potential feature vectors of users and attractions at the same time in the process of user-attraction matrix decomposition, accurate Complete the user-attraction matrix.

Description of drawings

Fig. 1 is the flow chart of the tourist attraction recommendation method based on dynamic theme model and matrix decomposition of the present invention;

Fig. 2 is a flowchart of the stage of obtaining data set information;

Fig. 3 is a flow chart of the stage of obtaining user's travel preference;

Fig. 4 is a schematic diagram of document generation based on a dynamic topic model;

Fig. 5 is a flowchart of the stage of recommending tourist attractions.

Detailed ways

In order to describe the present invention more specifically, the technical solutions of the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

As shown in Figure 1, the tourist attraction recommendation method based on the dynamic theme model and matrix decomposition of the present invention is divided into three stages: obtaining data set information, obtaining user travel preferences, and recommending tourist attractions:

Get dataset information stage

The flow chart of data processing is shown in Figure 2, and the steps are as follows:

S1-1, using the public API of the photo sharing website to download the photo dataset D _photo with geographic location information in the tourist city.

The specific steps for obtaining photo data include:

S1-1-1, through the public API provided by the photo sharing website (such as: Flickr), download the photos taken in the city and the corresponding metadata information according to the city. A photo p with geographic location information can be expressed as: p=(p _id , p _t , p _g , p _x , p _u ). Among them, p _id , p _t , p _g , p _x , p _u represent the unique identification number of the photo, the time when the user took the photo, the latitude and longitude information of the photo, the text description information added by the user for the photo, and the photo uploader's ID unique identification number;

S1-1-2, according to the unique identification information of the collected photos, obtain all photo information taken by each user, H _i ={p ₁ ,p ₂ ,…,p _e }, where e is taken by user i All photo counts for .

S1-2, using the entropy-based mobility method to filter the non-travel photos in the photo dataset D _photo , remove the noise photos in the photo set, and obtain the tourist photo dataset D _photo-travel .

The specific steps to remove non-tourism photos are as follows:

S1-2-1. Manually mark a small number of tourist photos and non-tourist photos in the user's photo collection according to the photo content and experience knowledge;

S1-2-2, divide the city into x′×y′ small grids (each small grid is represented as ( _xi ,y _j ), i=1,2,…,x; j=1,2,…, y), count the number of photos of the user's photos in these small grids, and calculate the ratio a of the photos in each small grid to the total number of photos, calculate the mobility entropy H _mobility of the user's photo set according to the principle of information entropy, and calculate The process is as shown in the following formula

According to H _mobility > ε, adjust the value of ε (from 0 to 1, every 0.1), and count the classification accuracy of the artificially marked photo set, select ε with high accuracy to classify the photos in the entire data set, and remove non-tourism photo.

S1-3, using the public API of the location-based social media website to download the check-in data D _check-in of the user in the tourist city.

The specific steps to obtain sign-in data include:

S1-3-1, through the public API provided by the location-based social media website (such as: SinaWeibo), download the data of the user's check-in in the city according to the city. A user's check-in behavior q can be expressed as: q=(q _id ,q _t ,q _g ,q _c ,q _u ); where, q _id ,q _t ,q _g ,q _c ,q _u represent the sign-in behavior The unique identification number, the time when the user checks in, the latitude and longitude information of the point of interest, the category information of the point of interest, and the unique identification number of the user who checks in;

S1-3-2, extract all the check-in data of each user according to the user identification number of the user check-in data; Q _i ={q ₁ ,q ₂ ,…,q _o }, where o is all the Check-ins.

S1-4, extracting the category information of the check-in places in the check-in data D _check-in , and counting the category data set D _category that constitutes the check-in places.

The user’s check-in behavior information includes some category information added when the user visits the POI. Statistics on this information can obtain all the category information of each POI. Specifically, it can be expressed as: C _POI = (c ₁ ,c ₂ ,…, c _z ). Among them, z is the number of categories contained in the interest point.

Get user travel preference stage

The flow chart of the stage of obtaining user travel preference is shown in Figure 3, and the steps are as follows:

S2-1, use the density-based clustering method to spatially cluster the photos with geographic location information, so as to obtain a set L of tourist attractions in the tourist city.

Users usually take photos at places they are more interested in, and if a large number of users take photos at one place, then the place can be considered as a tourist attraction. A density-based clustering method (for example: P-DBSCAN) is used to spatially cluster a large number of photos with geographic location information. Each cluster obtained represents a tourist attraction, and the cluster center is the location of the tourist attraction. . Through this process, a set of tourist attractions L={l ₁ ,l ₂ ,…,l _n } is excavated, where l={P _l ,g _l }, P _l is a collection of all photos belonging to a scenic spot, and g _l is a The latitude and longitude information of the attractions.

S2-2, according to the tourist attractions mined from the photos and the user's visit history at the attractions, count the number of times the user visits the attractions, and construct the user-attraction matrix Y.

The specific steps to establish the user-attraction matrix include:

S2-2-1, extracting the user's historical visit information to scenic spots from the photo collection, v=(l,u,t). Among them, l, u, t are respectively the tourist attraction visited by the user, the user's label, and the time when the user visits the scenic spot.

S2-2-2, counting the number of times the user visits the attractions. The constructed user-attraction matrix can be expressed as Y: Y∈R ^m×n , where m and n represent the number of users and attractions respectively, and the values in the matrix are the number of times users visit the attractions.

S2-3, the document represents the user's travel history, and the word represents the tourist attraction, using the dynamic topic model to infer the potential topic probability distribution of the user and the scenic spot in different time periods, and combine the topic probability distributions of all time periods to obtain all the user's implicit feature and the full implicit features of the sights

Dynamic topic models are applications that have been successfully used by researchers to obtain latent topic variations of documents. For a document, its specific generation process is shown in Figure 4. In this figure, the document set is divided into several sub-document sets according to time. The document sets and word themes in different time periods have a front-to-back dependency relationship, and the theme distribution of the latter time period is evolved from the theme of the previous time period. . α is the parameter that controls the topic distribution of the sub-document set, θ is the parameter that controls the topic distribution of a single document, β is the parameter that controls the probability distribution of the topic on the word, z represents the topic, k represents the number of topics, w represents the word, N Represents a document, and A represents a sub-document set. Through this model, the topic probability distribution of documents and words in different time periods can be obtained.

In the travel recommendation based on photo mining with geographic location information, the user's travel history information can be regarded as a combination of multiple topics, and each topic is the probability distribution of multiple scenic spots. Simply put, when using a dynamic topic model, documents represent user travel history and words represent tourist attractions.

S2-3-1. The specific steps to obtain the implicit characteristics of users and attractions include:

Divide the user's travel history information into different sub-datasets according to time slices (for example: year).

S2-3-2, the sub-dataset is used as the input of the dynamic topic model, and the topic probability distribution of users and scenic spots in different time periods is obtained through training,

in, is the topic probability distribution of the user in the Tth time period, The theme probability distribution of the scenic spot in the Tth time period, k is the number of themes;

S2-3-3, concatenating the topic probability distributions of the user in all time periods by time to form all the implicit features of the user Concatenate the theme probability distributions of attractions in all time periods together in time to form all the implicit features of attractions

S2-4, extract user's explicit features from user's travel history and explicit features of attractions Then combine all the implicit characteristics of the user and the full implicit features of the sights Establish user portrait F _user and scenic spot portrait F _location , and use cosine function to construct m×m user-user similarity matrix A and n×n scenic spot-sightseeing similarity matrix B.

The specific steps to establish A and B include:

S2-4-1, make statistics on the historical information of users visiting scenic spots (for example: the total number of users visiting scenic spots, the total number of users visiting a scenic spot, scenic spot category information, etc.), and obtain a large amount of display information to describe users and scenic spots The characteristics of are expressed as: and Among them, r and s respectively represent the total number of user and attraction display features, and each specific feature is shown in Table 1 and Table 2. The acquisition of some of these features requires the use of third-party network services, as detailed below:

Gender and age information: This information of tourists is obtained by analyzing the content of photos through third-party network services. For example: www.alchemyapi.com, when the web service obtains an uploaded photo, it will call its API function (ie, Alchemy Vision Face Detection and Recognition API) to analyze the photo, and then return the photo to the uploader The gender and age information of the faces that appear. The gender and age information of faces in all the photos taken at a scenic spot are counted to obtain the distribution of the gender and age information of the scenic spot. Similarly, by counting the gender and age information of faces in all the photos taken by a user, the distribution of the gender and age information of the faces in the photos taken by the user is obtained.

Weather information: Based on the third-party weather network service, the latitude and longitude information of the photo, and the time when the photo was taken, the weather information when the photo was taken can be obtained. For example: wundgerground.com, through the API function of this network service, the weather information of each location at different time points can be obtained. By making statistics on photos taken by users under different weather conditions, it is possible to obtain the degree of popularity of the scenic spot by tourists under different weather conditions.

S2-4-2, combine the explicit features and implicit features of users and attractions to construct a portrait of users and attractions, namely: and

S2-4-3. Combine the cosine formula to obtain the user-user similarity matrix A(m×m):

Among them, f _pi and f _qi represent the i-th explicit feature of users p and q, respectively;

Also use the cosine formula to obtain the attraction-attraction similarity matrix B(n×n), at this time, the cosine formula is:

where f′ _xi and f′ _yi denote the ith explicit features of scenic spots x and y, respectively.

S2-5, according to the established Y, A, and B, use the user-user and attraction-attraction similarity relationship to construct a matrix decomposition model with joint regularization items, and complete the decomposition of Y.

In the process of Y decomposition, the similarity information of A and B is used as an additional regular term to restrict the decomposition of Y. The specific objective function is:

In this formula, Sim _ig represents the similarity information between user u _i and user u _g , and Sim _jq represents the similarity information between scenic spots l _j and l _q . U(d×m) and L(d×n) are latent vector representations of users and attractions after Y decomposition, respectively. Among them, m, n, and d represent the number of users, the number of scenic spots, and the number of latent feature vectors, respectively.

The specific steps of decomposing Y based on the matrix factorization model with joint regularization items are as follows:

(a) randomly initialize parameter U and parameter L, and set learning rate α, error threshold δ, parameters λ ₁ and λ ₂ ;

(b) For each non-zero value R _ij in Y, according to Calculate the estimated value X _ij of R _ij , and according to Calculate the error between X _ij and the real value R _ij , and finally according to Statistics of the total error θ for all non-zero values, where w is the number of non-zero values;

(c) Judging whether the total error θ is greater than the error threshold δ, if so, perform step (d), if not, the iteration ends, U and L at this time are optimal values, and the decomposition of the matrix Y is completed;

(d) Use the gradient descent method to update the values of U and L, and then jump to step (b). The formula of the gradient descent method is:

S2-6, according to the result of the matrix decomposition with the joint regularization term, complete the sparse Y to obtain the user's travel preference.

According to the formula Fill in the missing values in Y, and the filled values represent the user's travel preference information.

Table 1 User explicit feature information

Table 2 Explicit feature information of attractions

Recommended Tourist Attraction Stage

The process of recommending tourist attractions is shown in Figure 5, which mainly includes the following steps:

S3-1. According to the information input by the user, a collection of attractions in the tourist city (destination) of the user is obtained, and these attractions are used as candidate sets for recommendation.

According to the user's ID and the user's tourist city c (destination), search in the completed user-sightseeing matrix to obtain the scenic spot set L' in the tourist city c.

S3-2, according to the result of step 1 and the completed user-attraction matrix, obtain the user's score for these scenic spots, that is, obtain the user's preference for these scenic spots.

Each value in the completed user-attraction matrix reflects the user's preference score for different scenic spots. According to this matrix and L', the user's preference for different scenic spots in the tourist city can be obtained.

S3-3, sort the attractions according to S3-2, and select top N tourist attractions to recommend to the user.

According to the user's ratings of different scenic spots in the tourist city, they are arranged in descending order, and the top N scenic spots are recommended to the user.

The above-mentioned specific embodiments have described the technical solutions and beneficial effects of the present invention in detail. It should be understood that the above-mentioned are only the most preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, supplements and equivalent replacements made within the scope shall be included in the protection scope of the present invention.

Claims (7)

1. A scenic spot recommendation method based on a dynamic theme model and matrix decomposition comprises the following steps:

and acquiring data set information: obtaining a photo dataset D from a social network_photoAnd de-noising the image to obtain a data set D of the tourist photo_photo-travelThen obtaining check-in data D from social network_check-inAnd extracting the check-in data D_check-inThe category data set D of the check-in place is obtained_category；

Obtaining user travel preference stage: the method comprises the following specific steps:

(2-1) Tourism photograph dataset D Using Density-based clustering method_photo-travelCarrying out spatial clustering on the photos to obtain a tourist attraction set L;

(2-2) based on the data set D from the tourist photo_photo-travelCounting the obtained times of visiting the scenic spots by the user and the scenic spots in the scenic spot set L, and constructing a user-scenic spot matrix Y;

(2-3) representing the document as the user travel history, representing the words as tourist attractions, deducing the potential theme probability distribution of the user and the attractions in different time periods by using a dynamic theme model, and combining the theme probability distribution of all time periods to obtain all implicit characteristics of the userAll implicit features of a harmony scene

(2-4) from the tourist photo data set D_photo-travelAnd category dataset D of check-in places_categoryExtracting explicit characteristics of a userExplicit characteristics of a attractionThen all implicit characteristics of the user are combinedAll implicit features of a harmony sceneCreating a user representation F_userScene image F_locationAnd constructing mxm user-user by cosine functionA similarity matrix A and an nxn scenery-scenery similarity matrix B;

(2-5) according to the Y, A and the B, constructing a matrix decomposition model with a joint regular term by utilizing similarity relations between the user and between the scenic spot and the scenic spot, and completing the decomposition of Y;

(2-6) completing a sparse user-scenery spot matrix Y according to a matrix decomposition result with a joint regular term to obtain a matrix Y' containing user travel preference information;

and recommending tourist attractions: and scoring the scenic spots in the candidate set by using the matrix Y', and recommending the N tourist spots with the top scores to the user.

2. The method for tourist attraction recommendation based on dynamic topic model and matrix decomposition of claim 1, wherein the specific steps of the stage of obtaining data set information are as follows:

(1-1) downloading photo data with geographical position information in a tourist city by using public API of a photo sharing website to form a photo data set D_photo；

(1-2) treating the photograph data set D using the entropy-based fluidity method_photoFiltering the non-tourist photos to remove the noise photos in the photo set to obtain a tourist photo data set D_photo-travel；

(1-3) downloading user check-in data D in a travel city using a public API of a location-based social media website_check-in；

(1-4) extracting check-in data D_check-inAnd counting category data sets D constituting the check-in place_category。

3. The method for recommending tourist attractions based on dynamic topic model and matrix decomposition of claim 1 wherein the specific steps of step (2-2) are as follows:

(2-2-1) from the tourist photograph data set D_photo-travelWherein v represents the user u at time tVisiting a tourist attraction l;

(2-2-2) counting the number of times each user visits each attraction and the travel photograph data set D according to all the access information v ═ (l, u, t)_photo-travelThe total number m of users in (1);

(2-2-3) constructing a user-sight spot matrix Y according to the number of times each user accesses each sight spot, the total number m of the users and n total sight spots in the sight spot set L, wherein Y belongs to R^m×nThe value at the (i, j) position in the matrix Y is the number of times the ith user accesses the jth attraction.

4. The method for tourist attraction recommendation based on dynamic theme model and matrix decomposition as claimed in claim 1, wherein said steps (2-3) are specifically:

(2-3-1) Tourism photograph data set D_photo-travelAll the access information v ═ l, u, t in the sub data sets are sliced according to the same time length, and M sub data sets corresponding to each time period are obtained;

(2-3-2) using the subdata sets as the input of the dynamic theme model, obtaining the theme probability distribution of the user and the scenic spots in different time periods through training,

wherein,is the topic probability distribution of the user at the T time period,is the topic probability distribution of the scenic spot in the Tth time period, k is the number of topics,for the kth topic probability distribution for the user at the T time period,the k topic probability distribution of the scenic spot in the T time period;

(2-3-3) concatenating the topic probability distributions of the users in all time periods together in time to form all implicit characteristics of the usersThe theme probability distributions of the scenic spots in all time periods are connected in series according to time to form all implicit characteristics of the scenic spots

5. The method of claim 4, wherein the steps (2-4) comprise the following steps:

(2-4-1) for tourist photo data set D_photo-travel(l, u, t) and the category data set D of the check-in place_categoryMaking statistics to obtain user explicit characteristicsExplicit characteristics of a attractionWherein r is the total number of the explicit characteristics of the user, and s is the total number of the explicit characteristics of the scenic spots;

(2-4-2) combining the explicit characteristics and the implicit characteristics of the user to construct the user portraitCombining the explicit characteristics and the implicit characteristics of the scenic spot to construct the scenic spot portrait

(2-4-3) combining the cosine formula to obtain a user-user similarity matrix A:

wherein u is_p,u_qRepresenting user p and user q, f, respectively_piAnd f_qiRespectively represents u_pAnd u_qR is the total number of explicit features;

the sight-sight similarity matrix B is also obtained by using a cosine formula, and at this time, the cosine formula is:

wherein l_x,l_yRespectively representing sight x and sight y, f_x′_iAnd f_y′_iRespectively represent l_xAnd l_yS is the number of explicit features.

6. The method as claimed in claim 5, wherein in the step (2-5), in the decomposition of the matrix Y, the similarity information of A and B is used as an additional regular term to limit the decomposition of Y, and the specific objective function is:

wherein R is_ijIs a matrixValue at the (I, j) position in Y, I_ijThe identifier indicating whether the user i accesses the sight j has a value of 1 if the user i accesses the sight j, and 0 if the user i does not access the sight j_igInformation indicating the similarity between user i and user g, Sim_jqIndicating a point of sight_jSimilarity information between q scenic spots and the scenic spot; u shape_iPotential feature vector, U, denoted as user i_gPotential feature vector, L, represented as user g_jPotential feature vector, L, denoted as sight j_qRepresented as a potential feature vector of the sight q,representing the distance between the potential feature vector of the user i and the potential feature vector of the user g, G (i) representing a similarity user group of the user i, Q (j) representing a similarity scenery group of a scenery j, U being the potential feature vector of the user after Y decomposition, and being d multiplied by m dimension; l is a potential feature vector of the scenic spot after Y decomposition and is d multiplied by n dimension; wherein m, n and d respectively represent the number of users, the number of scenic spots and the number of potential feature vectors;

the specific steps of decomposing Y based on the matrix decomposition model with the joint regularization term are as follows:

(a) randomly initializing U and L, and setting learning rate α, error threshold delta, parameter lambda₁And λ₂；

(b) For each non-zero value R in Y_ijAccording toCalculation of R_ijEstimated value X of_ijAccording toCalculating X_ijWith the true value R_ijIs finally based onCounting the total error theta of all nonzero values, wherein w is the number of the nonzero values;

(c) judging whether the total error theta is larger than an error threshold value delta or not, if so, executing the step (d), otherwise, finishing iteration, and finishing the decomposition of the matrix Y, wherein the U and the L are optimal values;

(d) updating the values of U and L by adopting a gradient descent method, and then skipping to execute the step (b), wherein the formula of the gradient descent method is as follows:

7. the method of claim 1, wherein the scenic spot recommendation method based on the dynamic topic model and the matrix factorization comprises the following steps:

(3-1) acquiring a scenic spot set of the user in the tourist city according to the user input information, wherein the scenic spots are used as a recommended candidate set;

(3-2) obtaining the scores of the recommended candidate concentrated scenic spots by the user according to the matrix Y' to obtain the scenic spots preferred by the user;

and (3-3) sorting the scores of the preferred scenic spots in a descending order, and selecting N tourist attractions with top scores and ranks to recommend to the user.