CN117194763A - Method for recommending next POI based on user preference and space-time context information - Google Patents

Abstract

The invention discloses a next POI recommendation method based on user preference and space-time context information, which comprises the following steps: acquiring a user sign-in record for preprocessing, carrying out vector embedding on interest points and various auxiliary information by using a low-dimensional dense vector, constructing and training a next interest point recommendation model based on user preference and space-time context information, inputting a long-term sign-in sequence and a short-term sign-in sequence of a user, generating k interest points, and recommending the k interest points to the user; the next interest point recommendation model comprises a user preference feature extraction module, a long-term behavior feature extraction module, a short-term behavior feature extraction module and a prediction module. According to the invention, the interest point-to-interest point transfer sequence is fully mined, the sparsity of geographic space and time information is relieved, the transfer information of the GRU module learning sequence is introduced, the double-layer attention mechanism is introduced to mine the behavior characteristics of the user, the recommendation range is reduced through the interest point-geographic filter, and the recommendation accuracy is greatly improved.

Description

Method for recommending next POI based on user preference and space-time context information

Technical Field

The invention belongs to the technical field of data mining and recommending systems, and particularly relates to a next POI recommending method based on user preference and space-time context information.

Background

In recent years, with the popularization of mobile intelligent devices and the development of mobile internet technology, the popularity of Location-based social networking (LBSN) has rapidly increased, such as Foursquare, gowalla, microblog, etc. The user can interact with friends in the LBSN, and can also generate interaction relation with a certain place through information such as characters, images and the like with geographic marks, such as visiting a restaurant, a bus stop and the like, so that the digital footprint of the user is left. These sites with geographical information are called points of interest (Point of Interest, POIs), while a user visiting a POI constitutes a Check-in, and the large number of Check-in records generated by the user provides a possibility to know the behavior rules of the user. The space-time movement of the user tends to have periodicity, and by deeply mining a large number of check-in records existing in the LBSN, the user is helped to more accurately provide recommended products conforming to the preference of the user, wherein the next POI recommendation task has been paid attention in recent years, and has become an important research direction in the LBSN-based recommendation task.

The early next POI recommendation system mostly adopts a matrix decomposition and Markov chain method, utilizes the matrix decomposition to learn the overall preference of all users, utilizes the Markov chain to mine the transfer characteristic of the POI sequence, and finally carries out the recommendation of the POI. With the development of the deep learning field, a learner introduces a method in natural language processing into a sequence mining and POI prediction recommendation system similar to the task, captures the transfer characteristics of a POI sequence by using a cyclic neural network, and finally recommends the POI. However, the existing methods fail to solve the sparseness problem of the space-time information, and the historical information cannot be fully mined, which greatly influences the accuracy of POI recommendation.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a next interest point recommending method based on user preference and space-time context information, which fully excavates the historical sign-in information of the user, relieves the sparseness problem of the data space-time information and improves the recommending accuracy.

In order to solve the problems, the invention adopts the following technical scheme:

the invention discloses a next interest point recommending method based on user preference and space-time context information, which comprises the following steps:

step 1, acquiring all check-in records of a group of users from a social network based on a position;

step 2, data preprocessing: cleaning all the check-in data, deleting invalid check-in records of missing users, interest points, interest point longitude and latitude, interest point categories or check-in time, and deleting check-in records corresponding to the interest points with the accessed times less than 10 and the check-in records of users with the number of the check-in records less than 100; and then unifying formats of the rest sign-in data after cleaning, specifically: generating a week and a time in a unified format according to the time stamp of each sign-in record, and generating a corresponding geographic position code according to the longitude and latitude of each sign-in record; storing the check-in records according to the user classification, and reconstructing the check-in records of each user according to the time sequence; then, judging, if the number of the user check-in records exceeds 500, taking the latest 500 check-in records in the user check-in records to form a long-term check-in sequence, otherwise, taking all the check-in records of the user to form a long-term check-in sequence; dividing each s check-in records in each user check-in record into a short-term check-in sequence, and dividing the last remaining s1 check-in records and the first s-s1 check-in records into a short-term check-in sequence; traversing all users to obtain a long-term sign-in sequence and a short-term sign-in sequence corresponding to each user;

step 3, embedding sign-in record information vectors: vector embedding is carried out on the interest points and various accessory information by using the low-dimensional dense vector, and the vector embedding of the accessory information is respectively carried out on the user, the geographic position code, the interest point category, "time" and "week" and category preference;

and 4, constructing and training a next interest point recommendation model based on user preference and space-time context information, wherein the method specifically comprises the following steps of: (1) constructing a user preference feature extraction module; the module adopts a double-layer attention mechanism to mine user preference characteristics in a long-term sign-in sequence of a user, and adopts a transform encoder based on a multi-head self-attention mechanism to capture correlations among sign-in records in the whole long-term sign-in sequence in parallel; the second-layer attention mechanism adopts a bilinear attention mechanism to capture the contribution degree of each sign-in record to the user category preference; (2) constructing a long-term behavior feature extraction module: the input of the module is a long-term sign-in sequence of a user, interest points, geographical position codes, interest point categories and time and week attribute embedded vectors of each sign-in record are spliced, then the spliced vectors of the whole long-term sign-in sequence are input into a plurality of layers of GRUs, the output vectors of the plurality of layers of GRUs are spliced with the user embedded vectors, and then the spliced vectors are added with the output vectors of the user preference feature extraction module through a full connection layer; (3) constructing a short-term behavior feature extraction module: the input of the module is a short-term sign-in sequence of a user, the interest point embedded vector, the geographic position code embedded vector and the interest point category embedded vector of each sign-in record are respectively spliced with the time embedded vector of the interest point embedded vector to obtain three spliced vectors, the three spliced vectors are respectively input into GRUs with the three parameters not shared, the output vectors of the three GRUs are weighted and added, then the three GRUs are spliced with the user embedded vector, and the three GRUs are added with the output vector of the user preference feature extraction module through a full connection layer; (4) constructing a prediction module: the method comprises the steps that the output of a long-term behavior feature extraction module and a short-term behavior feature extraction module passes through a nonlinear layer to output two interest point probability distribution vectors and two geographic position coding probability distribution vectors, the two interest point probability distribution vectors are combined into an interest point probability distribution vector through an interest point fusion layer, and the two geographic position coding probability distribution vectors are combined into a geographic position coding probability distribution vector through the geographic position fusion layer; then, ordering the probability distribution vectors of the interest points and the probability distribution vectors of the geographic position codes from big to small according to probability values, taking the first 10 geographic position codes with the largest probability values of the geographic position codes, taking the first k interest points with the largest probability values in the geographic position codes to form a recommendation list for outputting, wherein the k value is 1, 5 or 10, and obtaining the next interest point recommendation model based on user preference and space-time context information; (5) Training a next point of interest recommendation model based on user preferences and spatiotemporal context information: constructing a training set according to the long-term sign-in sequence and the short-term sign-in sequence of each user, calculating the interest point prediction loss and the geographic position coding prediction loss of the next interest point recommendation model by adopting a cross entropy loss function, and updating the parameters of the next interest point recommendation model by using an Adam optimizer so as to realize the training of the next interest point recommendation model;

and 5, inputting the long-term sign-in sequence and the short-term sign-in sequence of the user into a trained next interest point recommendation model based on user preference and space-time context information, and generating k interest points to be recommended to the user.

Preferably, the geographic position coding adopts a DBSCAN clustering algorithm.

Preferably, the specific process of embedding the interest points is as follows:

(1) establishing an isomorphic graph according to the sign-in sequences of all users, taking each interest point as a node in the isomorphic graph network, taking the transfer relation from the interest point to the interest point as an edge in the isomorphic graph network, wherein the weight of the edge is the transfer times between two interest points, and marking the direction from the interest point to the interest point according to the time sequence to form a directed graph;

(2) setting sampling length, sampling based on the weight of the edges in the isomorphic graph, learning the sequence obtained by the sampling by adopting a structured deep network embedding model, and embedding all interest points on the directed graph into vectors.

Preferably, the user, the geographic position code, the interest point category, the time and the week are embedded, and the specific processes are as follows: and respectively encoding the user, the geographic position code, the interest point category, the time and the week on the sign-in data into independent heat vectors, and inputting the independent heat vectors into a common embedded model to obtain embedded vectors of the user, the geographic position code, the interest point category, the time and the week.

Preferably, the saidThe specific process of embedding category preference is as follows: initializing a Y-dimensional vector v for user U e U ₀ ，…v _Y-1 ]Recording the times of accessing each interest point category by the user u to the position corresponding to the vector, and recording the non-accessed interest point category as 0;

obtaining a vector v ₀ ，…v _Y-1 ]The category preference vector for user u is then defined as follows:

traversing all users in the U to obtain the class preference vector corresponding to each user.

Preferably, the specific calculation process of the long-term behavior feature extraction module is as follows:

wherein O is ^long The output of the long-term behavior feature extraction module is represented, FC (·) represents a fully connected layer, mulGRU (·) represents a multi-layer GRU; the GRU calculation process comprises the following steps:

z _i ＝σ(W _xz x _i +W _hz h _i-1 +b _z )

r _i ＝σ(W _xr x _i +W _hr h _i-1 +b _r )

c _i ＝tanh(W _xc x _i +r _i ⊙(W _hc h _i-1 )+b _c )

h _i ＝(1-z _i )⊙c _i +z _i ⊙h _i-1

wherein σ represents a Sigmoid function, x _i For input at instant i, h _i For output at time i, h _i-1 For output at instant i-1, z _i 、r _i 、c _i As intermediate variable, W _xz 、W _hz 、W _xr 、W _hr 、W _xc 、W _hc B as trainable parameters _z 、b _r 、b _c For a trainable bias term, tanh () is a tanh activation function, and by-element multiplication is indicated.

Preferably, the specific calculation process of the short-term behavior feature extraction module is as follows:

wherein O is ^short The output of the short-term behavior feature extraction module is represented, alpha, beta and gamma are represented by weights, GRU(s) are represented by a single-layer GRU module, and V ^l 、V ^g 、V ^c And V ^t All the interest point embedded vectors, all the geographic position code embedded vectors, all the interest point category embedded vectors and all the time embedded vectors in the short-term sign-in sequence are respectively represented.

Preferably, the cross entropy loss function is defined as:

Loss＝Loss ^poi +Loss ^geo

wherein Loss is ^poi 、Loss ^geo And Loss is the predicted Loss of interest, the predicted Loss of geo-location coding and the total Loss, l _i At time t for user u _i Point of interest accessed, g _i For the interest point l _i Is used for the geographic location encoding of (c) the code,derived for prediction module>At time t _i Probability value of>Derived for prediction module>At time t _i Probability values of (a) are provided.

The invention has the following beneficial effects:

according to the invention, three vector embedding (vector embedding of interest points, vector embedding of users, geographic position codes, interest point categories, "time" and "week", vector embedding of category preference) modes are introduced, the interest point-to-interest point transfer sequence with obvious graph structural features is fully mined, sparsity of geographic space information is relieved on the basis of introducing a spatial clustering algorithm, sparsity of time information is relieved on the basis of extracting time sub-attributes in a time stamp, transfer information of a GRU module learning sequence with simpler structural parameters and better effects is introduced, behavior features of users are better mined by introducing a double-layer attention mechanism, and finally recommendation range is reduced by an interest point-geographic filter, so that recommendation accuracy is greatly improved.

Drawings

FIG. 1 is a flow chart of the method of the present invention.

FIG. 2 is a flow chart of the sign-in record information vector embedding in the present invention.

FIG. 3 is a block diagram of a next point of interest recommendation model based on user preferences and spatio-temporal context information in accordance with the present invention.

FIG. 4 is a schematic diagram of a point of interest-geographic filter of the present invention.

Detailed Description

The invention is further described below with reference to the accompanying drawings.

As shown in fig. 1, the next point of interest recommendation method based on user preference and space-time context information includes the steps of:

step 1, all check-in records of a group of users are obtained from a location-based social network (LBSN), and each check-in record comprises five check-in record information of the user, the interest point category, the longitude and latitude of the interest point and the check-in timestamp.

Step 2, all the check-in records obtained in the step 1 are processed as follows: deleting the invalid check-in records of the missing check-in record information, deleting the check-in records corresponding to the interest points with the accessed times less than 10 and the check-in records of the users with the check-in records less than 100, and obtaining the valid check-in records. And arranging the effective check-in records in time sequence to obtain check-in data.

To better incorporate the example description, based on the check-in data described above, the following definitions are made:

definition u= { U ₀ ，u ₁ ，…，u _M-1 -representing a set of users in check-in data; definition l= { L ₀ ，l ₁ ，…，l _N-1 -representing a set of points of interest in check-in data; the invention adopts DBSCAN clustering algorithm to carry out spatial clustering according to longitude and latitude of all interest points in the sign-in data, gives each interest point a geographic position code according to clustering results, and defines G= { G ₀ ，g ₁ ，…，g _X1-1 -representing a set of geo-location encodings of points of interest in check-in data; definition c= { C ₀ ，c ₁ ，…，c _Y-1 -representing a set of point of interest categories (e.g., restaurants, schools, etc.) in check-in data; the time stamps of the original check-in records are consecutive by extracting the time sub-attributes in the check-in time stamps: "time" (several minutes) and "week" (several weeks) all time stamps are spaced within intervals of the same length. M, N, X1 and Y are the total number of users, the total number of points of interest, the total number of geo-location codes, and the total number of categories of points of interest in the check-in data, respectively.

Define the sign-in sequence of user U E U as H ^u ＝{h ₀ ，h ₁ ，…，h _n-1 And the sign-in record consists of a sign-in sequence, wherein n is the length of the sign-in sequence, and the sign-in sequence is a time sequence. Each check-in record h _i ∈H ^u Comprises four attributes (l _i ，g _i ，c _i ，t _i ) Indicating that user u is at time t _i Visit the interest point l _i ，g _i For the interest point l _i Geographic position coding of c _i For the interest point l _i Wherein t is _i Is a binary group consisting of time and week

Defining the long-term sign-in sequence of the user U epsilon U as L ^u ＝{h ₀ ，h ₁ ，…，h _l-1 If the total number of check-in records of the user u exceeds 500, the last 500 check-in records in the check-in records of the user u are taken to form a long-term check-in sequence, i.e. l=500, otherwise all the check-in records of the user u are taken to form the long-term check-in sequence. The user generates more check-in records in a long-term range to form a long-term check-in sequence, which contains long-term activity characteristics of the user, namely long-term behavior characteristics.

Splitting all check-in sequences of the user U epsilon U into a group of short-term check-in sequences with length S to obtain S ^u ＝{h ₀ ，h ₁ ，…，h _s-1 }。

And traversing all users in the U to obtain a long-term check-in sequence and a short-term check-in sequence corresponding to each user. The last short-term sign-in sequence of each user is taken to form a test data set, and the rest is a training data set.

And step 3, embedding the sign-in record information vector. As shown in fig. 2, the method specifically includes:

(1) Embedding interest points: map points of interest into vectors using graph embedding techniques. The sign-in sequence has obvious graph structural characteristics, and in a large number of sign-in sequences, the connection lines from the interest points to the interest points form isomorphic graphs, and the structural information and the potential characteristics of the sign-in sequence cannot be well represented by using a traditional word embedding mode. The Graph Embedding (Graph Embedding) technology can map Graph data into low-dimensional dense vectors, and each interest point on the isomorphic Graph is embedded to obtain a vector expression, so that the spatial characteristics and interaction relations among the interest points are well reserved.

The specific process is as follows:

(1) and establishing an isomorphic graph according to the sign-in sequences of all users, taking each interest point as a node in the isomorphic graph network, taking the transfer relation from the interest point to the interest point as an edge in the isomorphic graph network, wherein the weight of the edge is the transfer times between two interest points, and marking the direction from the interest point to the interest point according to the time sequence to form a directed graph G.

(2) Setting a sampling length, sampling based on the weight of the edge in the isomorphic graph, learning the sequence obtained by sampling by adopting a Structured Deep Network Embedding (SDNE) model, embedding all interest points on the directed graph G into vectors, wherein the dimension of the interest point embedded vector is D ^l 。

(2) User, geo-location coding, point of interest category, "time" and "week" embedding: and respectively encoding the user, the geographic position code, the interest point category, the time and the week on the sign-in data into independent heat vectors, and inputting the independent heat vectors into a common embedding model (such as an nn. Embedding model of pyrach) to obtain the embedding vectors of the user, the geographic position code, the interest point category, the time and the week. By D ^u 、D ^g 、D ^c 、D ^h And D ^w The user embedding vector dimension, the geographic position code embedding vector dimension, the interest point category embedding vector dimension, the "time" embedding vector dimension and the "week" embedding vector dimension are respectively represented. For convenience of explanation, the embedded vectors of the time and the week are spliced, and the dimension is D ^t ＝D ^h +D ^w 。

(3) Category preference embedding: initializing a Y-dimensional vector v for user U e U ₀ ，…v _Y-1 ]And recording the times of accessing each interest point category by the user u to the position corresponding to the vector, and recording the non-accessed interest point category as 0.

Obtaining a vector v ₀ ，…v _Y-1 ]The category preference vector for user u is then defined as follows:

traversing all users in the U to obtain the class preference vector corresponding to each user.

Step 4, constructing and training a next interest point recommendation model based on user preference and space-time context information, as shown in fig. 3, specifically including:

(1) And constructing a user preference feature extraction module. The module adopts a double-layer attention mechanism to mine user preference characteristics in a long-term sign-in sequence of a user, and inputs a long-term sign-in sequence H of a user U epsilon U ^u ＝{h ₀ ，h ₁ ，…，h _l-1 To fully consider the context information of the long-term check-in sequence, each check-in record h _i ∈H ^u Splicing four attributes of (a):

wherein,representing vector concatenation->And->Representing each check-in record h _i ∈H ^u Four attributes (l) _i ，g _i ，c _i ，t _i ) Is used to determine the embedded vector of (a).

First layer attention mechanism: using a transducer encoder based on a multi-headed self-attention mechanism, correlations between check-in records in the entire long-term check-in sequence can be captured in parallel. The transducer is parallel computing, can not capture the position information in the sequence like a cyclic neural network, and needs to splice the vector V of the whole long-term sign-in sequence ^P Bits corresponding to the position vector PE are added and input to the transducer encoder in parallel.

The input to the transducer encoder is denoted as x=v ^P +PE, the output is expressed as:

O ^E ＝lNo(lNo(X+MulH(X))+FFN(lNo(X+MulH(X))))

where lNo (-) represents layer normalization, FFN (-) represents fully connected feed forward network, mulH (-) represents multi-headed self-attention mechanism.

The present invention refers to the following geolocation codes:

and->

Wherein,and->Values representing even and odd dimensions of the position vector, i representing the position of the check-in record in the check-in sequence, E _size Representing a splice vector +.>K is from 0 to +.>An integer value in between.

Second layer attention mechanism: by adopting a bilinear attention mechanism, the contribution degree of each sign-in record to the user category preference can be captured to form a normalized vector output O ^per ：

Wherein O is ^per The output vector of the module is extracted for the user preference feature,is O ^E Is the i-th dimension vector of (a), alpha _i For the weights, f (·) is a bilinear attention function, and W is a matrix of parameters that can be learned.

(2) And constructing a long-term behavior feature extraction module. The input of the module is a long-term sign-in sequence L of the user U epsilon U ^u ＝{h ₀ ，h ₁ ，…，h _l-1 Each check-in record h _i ∈L ^u Splicing the four attributes of the long-term sign-in sequence, and then splicing the splicing vector V of the whole long-term sign-in sequence ^P Input to the multi-layer GRU, and embed the output vector of the multi-layer GRU with the user-embedded vector V ^u Splicing, and adding the spliced output vectors with the output vectors of the user preference feature extraction module through a full connection layer, wherein the specific calculation process is as follows:

wherein O is ^long Representing the output of the long-term behavioral characteristics extraction module, FC (·) representing fully connected layers, mulGRU (·) representing multi-layer GRUs. The GRU calculation process comprises the following steps:

x _i ＝V _i ^P

z _i ＝σ(W _xz x _i +W _hz h _i-1 +b _z )

r _i ＝σ(W _xr x _i +W _hr h _i-1 +b _r )

c _i ＝tanh(W _xc x _i +r _i ⊙(W _hc h _i-1 )+b _c )

h _i ＝(1-z _i )⊙c _i +z _i ⊙h _i-1

wherein sigma represents Sigmoid functionNumber, x _i For input at instant i, h _i For output at time i, h _i-1 For output at instant i-1, z _i 、r _i 、c _i As intermediate variable, W _xz 、W _hz 、W _xr 、W _hr 、W _xc 、W _hc B as trainable parameters _z 、b _r 、b _c For a trainable bias term, tanh () is a tanh activation function, and by-element multiplication is indicated.

(3) And constructing a short-term behavior feature extraction module. The input of this module is the short-term check-in sequence S of user U ε U ^u ＝{h ₀ ，h ₁ ，…，h _s-1 Each check-in record h _i ∈S ^u The interest point embedding vector, the geographic position coding embedding vector and the interest point category embedding vector are respectively spliced with the time embedding vector to obtain three splicing vectors, all the three splicing vectors are input into GRUs which are not shared by the three parameters, the output vectors of the three GRUs are weighted and added, and then the three splicing vectors are embedded with the user embedding vector V ^u Splicing, adding the spliced output vectors with the output vectors of the user preference feature extraction module through the full connection layer, wherein the specific calculation process is as follows:

wherein O is ^short The output of the short-term behavior feature extraction module is represented, alpha, beta and gamma are represented by weights, GRU(s) are represented by a single-layer GRU module, and V ^l 、V ^g 、V ^c And V ^t All the interest point embedded vectors, all the geographic position code embedded vectors, all the interest point category embedded vectors and all the time embedded vectors in the short-term sign-in sequence are respectively represented.

(4) And constructing a prediction module. The output of the long-term behavior feature extraction module is passed through a nonlinear layer NL _POI ¹ And NL _Geo ¹ Passing the output of the short-term behavior feature extraction module through a nonlinear layer NL _POI ² And NL _Geo ² The method comprises the steps of outputting two interest point probability distribution vectors and two geographic position coding probability distribution vectors, merging the two interest point probability distribution vectors into one interest point probability distribution vector through an interest point fusion layer, merging the two geographic position coding probability distribution vectors into one geographic position coding probability distribution vector through the geographic position fusion layer, and specifically calculating:

where Pre is the final recommended result and,and->Encoding probability distribution vectors for points of interest probability distribution vectors and geolocations, softmax () is a Softmax function, NL _POI ¹ 、NL _POI ² 、NL _Geo ¹ And NL _Geo ² Is a nonlinear layer, and lambda and mu are learnable parameters. The nonlinear layer includes Dropout () function, reLU () function and fully connected layer, which can improve generalization ability of the model, and convert an input vector into a desired dimension.

Then, the recommendation list Pre is output through the point-of-interest geographic filter GeoFilter (·), which operates as follows: as shown in fig. 4, the probability distribution vectors of the points of interest and the probability distribution vectors of the geographic position codes are respectively sequenced from big to small according to probability values, the first 10 geographic position codes with the largest probability values of the geographic position codes are taken, the first k points of interest with the largest probability values in the geographic position codes are taken to form a recommendation list to be output, and k takes the value of 1, 5 or 10 to obtain the next point of interest recommendation model based on user preference and space-time context information;

(5) The next point of interest recommendation model based on the user preferences and the spatiotemporal context information is trained. And constructing a training set according to the long-term sign-in sequence and the short-term sign-in sequence of each user, calculating the interest point prediction loss and the geographic position coding prediction loss of the next interest point recommendation model by adopting a cross entropy loss function, and updating the parameters of the next interest point recommendation model by using an Adam optimizer so as to realize the training of the next interest point recommendation model. The cross entropy loss function is defined as:

Loss＝Loss ^poi +Loss ^geo

wherein Loss is ^poi 、Loss ^geo And Loss is the predicted Loss of interest, the predicted Loss of geo-location coding and the total Loss, l _i At time t for user u _i Point of interest accessed, g _i For the interest point l _i Is used for the geographic location encoding of (c) the code,derived for prediction module>At time t _i Probability value of>Derived for prediction module>At time t _i Probability values of (a) are provided.

And 5, inputting the long-term sign-in sequence and the short-term sign-in sequence of the user into a trained next interest point recommendation model based on user preference and space-time context information, and generating k interest points to be recommended to the user.

Claims (8)

1. The next POI recommendation method based on the user preference and the space-time context information is characterized by comprising the following steps of: the method comprises the following steps:

step 1, acquiring all check-in records of a group of users from a social network based on a position;

step 2, data preprocessing: cleaning all the check-in data, deleting invalid check-in records of missing users, interest points, interest point longitude and latitude, interest point categories or check-in time, and deleting check-in records corresponding to the interest points with the accessed times less than 10 and the check-in records of users with the number of the check-in records less than 100; and then unifying formats of the rest sign-in data after cleaning, specifically: generating a week and a time in a unified format according to the time stamp of each sign-in record, and generating a corresponding geographic position code according to the longitude and latitude of each sign-in record; storing the check-in records according to the user classification, and reconstructing the check-in records of each user according to the time sequence; then, judging, if the number of the user check-in records exceeds 500, taking the latest 500 check-in records in the user check-in records to form a long-term check-in sequence, otherwise, taking all the check-in records of the user to form a long-term check-in sequence; dividing each s check-in records in each user check-in record into a short-term check-in sequence, and dividing the last remaining s1 check-in records and the first s-s1 check-in records into a short-term check-in sequence; traversing all users to obtain a long-term sign-in sequence and a short-term sign-in sequence corresponding to each user;

step 3, embedding sign-in record information vectors: vector embedding is carried out on the interest points and various accessory information by using the low-dimensional dense vector, and the vector embedding of the accessory information is respectively carried out on the user, the geographic position code, the interest point category, "time" and "week" and category preference;

and 4, constructing and training a next interest point recommendation model based on user preference and space-time context information, wherein the method specifically comprises the following steps of: (1) constructing a user preference feature extraction module; the module adopts a double-layer attention mechanism to mine user preference characteristics in a long-term sign-in sequence of a user, and adopts a transform encoder based on a multi-head self-attention mechanism to capture correlations among sign-in records in the whole long-term sign-in sequence in parallel; the second-layer attention mechanism adopts a bilinear attention mechanism to capture the contribution degree of each sign-in record to the user category preference; (2) constructing a long-term behavior feature extraction module: the input of the module is a long-term sign-in sequence of a user, interest points, geographical position codes, interest point categories and time and week attribute embedded vectors of each sign-in record are spliced, then the spliced vectors of the whole long-term sign-in sequence are input into a plurality of layers of GRUs, the output vectors of the plurality of layers of GRUs are spliced with the user embedded vectors, and then the spliced vectors are added with the output vectors of the user preference feature extraction module through a full connection layer; (3) constructing a short-term behavior feature extraction module: the input of the module is a short-term sign-in sequence of a user, the interest point embedded vector, the geographic position code embedded vector and the interest point category embedded vector of each sign-in record are respectively spliced with the time embedded vector of the interest point embedded vector to obtain three spliced vectors, the three spliced vectors are respectively input into GRUs with the three parameters not shared, the output vectors of the three GRUs are weighted and added, then the three GRUs are spliced with the user embedded vector, and the three GRUs are added with the output vector of the user preference feature extraction module through a full connection layer; (4) constructing a prediction module: the method comprises the steps that the output of a long-term behavior feature extraction module and a short-term behavior feature extraction module passes through a nonlinear layer to output two interest point probability distribution vectors and two geographic position coding probability distribution vectors, the two interest point probability distribution vectors are combined into an interest point probability distribution vector through an interest point fusion layer, and the two geographic position coding probability distribution vectors are combined into a geographic position coding probability distribution vector through the geographic position fusion layer; then, ordering the probability distribution vectors of the interest points and the probability distribution vectors of the geographic position codes from big to small according to probability values, taking the first 10 geographic position codes with the largest probability values of the geographic position codes, taking the first k interest points with the largest probability values in the geographic position codes to form a recommendation list for outputting, wherein the k value is 1, 5 or 10, and obtaining the next interest point recommendation model based on user preference and space-time context information; (5) Training a next point of interest recommendation model based on user preferences and spatiotemporal context information: constructing a training set according to the long-term sign-in sequence and the short-term sign-in sequence of each user, calculating the interest point prediction loss and the geographic position coding prediction loss of the next interest point recommendation model by adopting a cross entropy loss function, and updating the parameters of the next interest point recommendation model by using an Adam optimizer so as to realize the training of the next interest point recommendation model;

and 5, inputting the long-term sign-in sequence and the short-term sign-in sequence of the user into a trained next interest point recommendation model based on user preference and space-time context information, and generating k interest points to be recommended to the user.

2. The method of claim 1, wherein the next POI recommendation based on user preferences and spatiotemporal context information comprises: the geographic position coding adopts a DBSCAN clustering algorithm.

3. The method of claim 1, wherein the next POI recommendation based on user preferences and spatiotemporal context information comprises: the specific process of embedding the interest points is as follows:

(1) establishing an isomorphic graph according to the sign-in sequences of all users, taking each interest point as a node in the isomorphic graph network, taking the transfer relation from the interest point to the interest point as an edge in the isomorphic graph network, wherein the weight of the edge is the transfer times between two interest points, and marking the direction from the interest point to the interest point according to the time sequence to form a directed graph;

(2) setting sampling length, sampling based on the weight of the edges in the isomorphic graph, learning the sequence obtained by the sampling by adopting a structured deep network embedding model, and embedding all interest points on the directed graph into vectors.

4. The method of claim 1, wherein the next POI recommendation based on user preferences and spatiotemporal context information comprises: the user, the geographic position code, the interest point category, the time and the week are embedded, and the specific process is as follows: and respectively encoding the user, the geographic position code, the interest point category, the time and the week on the sign-in data into independent heat vectors, and inputting the independent heat vectors into a common embedded model to obtain embedded vectors of the user, the geographic position code, the interest point category, the time and the week.

5. The method of claim 1, wherein the next POI recommendation based on user preferences and spatiotemporal context information comprises: the specific process of embedding the category preference comprises the following steps: initializing a Y-dimensional vector v for user U e U ₀ ，…v _Y-1 ]Recording the times of accessing each interest point category by the user u to the position corresponding to the vector, and recording the non-accessed interest point category as 0;

obtaining a vector v ₀ ，…v _Y-1 ]The category preference vector for user u is then defined as follows:

traversing all users in the U to obtain the class preference vector corresponding to each user.

6. The method of claim 1, wherein the next POI recommendation based on user preferences and spatiotemporal context information comprises: the specific calculation process of the long-term behavior characteristic extraction module is as follows:

wherein O is ^long The output of the long-term behavior feature extraction module is represented, FC (·) represents a fully connected layer, mulGRU (·) represents a multi-layer GRU; the GRU calculation process comprises the following steps:

z _i ＝σ(W _xz x _i +W _hz h _i-1 +b _z )

r _i ＝σ(W _xr x _i +W _hr h _i-1 +b _r )

c _i ＝tanh(W _xc x _i +r _i ⊙(W _hc h _i-1 )+b _c )

h _i ＝(1-z _i )⊙c _i +z _i ⊙h _i-1

wherein σ represents a Sigmoid function, x _i For input at instant i, h _i For output at time i, h _i-1 For output at instant i-1, z _i 、r _i 、c _i As intermediate variable, W _xz 、W _hz 、W _xr 、W _hr 、W _xc 、W _hc B as trainable parameters _z 、b _r 、b _c For a trainable bias term, tanh () is a tanh activation function, and by-element multiplication is indicated.

7. The method of claim 1, wherein the next POI recommendation based on user preferences and spatiotemporal context information comprises: the specific calculation process of the short-term behavior feature extraction module is as follows:

wherein O is ^short The output of the short-term behavior feature extraction module is represented, alpha, beta and gamma are represented by weights, GRU(s) are represented by a single-layer GRU module, and V ^l 、V ^g 、V ^c And V ^t All the interest point embedded vectors, all the geographic position code embedded vectors, all the interest point category embedded vectors and all the time embedded vectors in the short-term sign-in sequence are respectively represented.

8. The method of claim 1, wherein the next POI recommendation based on user preferences and spatiotemporal context information comprises: the cross entropy loss function is defined as:

Loss＝Loss ^poi +Loss ^geo

wherein Loss is ^poi 、Loss ^geo And Loss is the predicted Loss of interest, the predicted Loss of geo-location coding and the total Loss, l _i At time t for user u _i Point of interest accessed, g _i For the interest point l _i Is used for the geographic location encoding of (c) the code,derived for prediction module>At time t _i Probability value of>Derived for prediction module>At time t _i Probability values of (a) are provided.