A POI recommendation method, device and equipment based on a multi-element relationship space-time network

Abstract

The application discloses a POI recommendation method based on a multi-element relationship space-time network, which comprises the following steps: S11, acquiring user information in an LBSN, and a historical track of a POI and a user; S12, constructing a plurality of explicit POI relationship graphs according to the POI and the historical track, and learning an implicit relationship of the explicit POI relationship based on a hypergraph architecture to obtain a candidate POI embedding vector set corresponding to all POIs; S13, obtaining a check-in set of a same space-time position from the historical track according to a space-time position of a last check-in of a user, and adding the check-in set into a current track to obtain an enhanced track; S14, constructing a corresponding space-time correlation matrix according to the enhanced track to obtain a track embedding vector; and S15, performing prediction according to the track embedding vector and the candidate POI embedding vector set to obtain a POI recommendation result. The application further provides a POI device and equipment. The method can effectively improve the accuracy of POI recommendation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of Internet, and is a POI recommendation method, device and equipment based on a multi-element relationship spatio-temporal network. BACKGROUND

[0002] With the increasing popularity of location-based social networks (LBSN), such as Foursquare, Facebook, etc., the check-in data of users at points of interest (POI) gradually increases, generating massive POI trajectory information containing user behavior and preferences. These information can support various personalized location-based services, and POI recommendation is one of them. The purpose is to predict the POI to be visited according to the preference information in the user's historical trajectory. POI recommendation can not only help users plan their travel, but also benefit POI holders to attract target groups by placing advertisements.

[0003] In POI recommendation, not only the sequential patterns and spatio-temporal information in the trajectory need to be considered, but also the multi-element relationship between different POIs. Therefore, it is an urgent need in the academic and industrial circles to design a POI recommendation method based on a multi-element relationship spatio-temporal network.

[0004] In the process of implementing the present application, the inventors found that at least the following problems exist in the prior art:

[0005] Firstly, they failed to fully utilize the rich information available in LBSN, including the multi-element relationship between POIs and the individualized movement patterns in user trajectories, resulting in insufficient information contained in the learned POI and trajectory representations, and making the final recommendation less accurate. Secondly, due to limited computing resources, existing methods often only select the most recent sub-trajectory for calculation, causing the omission of long-term historical records, thereby ignoring information about the user's long-term movement patterns.

[0006] Patent document CN115795182A discloses a next POI recommendation method based on a graph convolution network. First, an initialization embedding layer of user embedding, POI embedding, time embedding, category embedding, and relative position embedding is generated. Then, a convolution layer is built based on a graph convolution network, an attention mechanism, and a feedforward layer. Finally, the user's preference for POIs is obtained by performing an inner product operation on the obtained user features and POI features. This method analyzes user information by constructing a graph convolution network, but the entire history trajectory needs to be processed throughout the process, which requires high computing power.

[0007] Patent document CN115130018A discloses a POI recommendation method and device for distinguishing user exploratory mode and spatiotemporal revisiting mode. The method first models user preferences in exploratory mode and user preferences in spatiotemporal revisiting mode, calculates spatiotemporal correlation coefficients using time intervals and spatial intervals, then calculates scores of all POIs in the POI candidate set in the two modes respectively. Then the conversion probability of exploratory mode and spatiotemporal revisiting mode is calculated. Finally, the final POI recommendation score is calculated by combining the POI scores in the two modes and the conversion probability of the two modes. The top N POIs with the highest recommendation scores are selected by sorting the final POI recommendation scores. This method needs to analyze the historical data of all users in the region, but does not consider the omission of long-term historical records, thereby ignoring information about the long-term movement mode of the user. SUMMARY

[0008] In view of the deficiencies of the prior art, the purpose of the embodiments of the present application is to provide a POI recommendation method, device and equipment based on a multi-element relationship spatiotemporal network, which fully exploits spatiotemporal information in LBSN and efficiently utilizes long-term historical records to improve the accuracy of POI recommendation.

[0009] According to a first aspect of the purpose of the embodiments of the present application, a POI recommendation method based on a multi-element relationship spatiotemporal network is provided, comprising:

[0010] S11, obtaining user information in LBSN, POI and user historical trajectory, the historical trajectory being composed of a plurality of check-ins, wherein each check-in is a set of three tuples containing spatiotemporal information.

[0011] The POI includes the name, category, coordinate and classification of the point of interest.

[0012] S12, constructing a plurality of explicit POI relationship graphs according to the POI and the historical trajectory, learning the implicit relationship of the explicit POI relationship based on a hypergraph architecture to obtain a candidate POI embedding vector set corresponding to all POIs.

[0013] S13, according to the spatiotemporal position of the last check-in of the user, obtaining a check-in set of the same spatiotemporal position from the historical trajectory and adding the check-in set to the current trajectory to obtain an enhanced trajectory.

[0014] S14, constructing a corresponding spatiotemporal correlation matrix according to the enhanced trajectory to obtain a trajectory embedding vector.

[0015] S15, predicting according to the trajectory embedding vector and the candidate POI embedding vector set to obtain a POI recommendation result.

[0016] Specifically, in S12, the construction process of the candidate POI embedding vector set is as follows:

[0017] S21, initialize a learnable embedding vector for each POI;

[0018] S22, judge the distance of the geographical position in each two POIs, if less than the geographical distance threshold, as a neighbor node to construct a distance relationship graph;

[0019] S23, based on the historical trajectory, two consecutive check-ins are taken as neighbor nodes to construct a relationship graph, and the nodes are replaced by the corresponding POIs according to the spatio-temporal information of the check-ins, to obtain a conversion relationship graph;

[0020] S24, weighted fusion based on the distance relationship graph and the conversion relationship graph to obtain a candidate POI embedding vector set.

[0021] Specifically, in S13, the enhanced trajectory is obtained as follows:

[0022] S31, create a corresponding spatio-temporal record table according to the historical trajectory;

[0023] S32, according to the spatio-temporal position of the last check-in of the user, the corresponding time and geographical information are mapped to high-dimensional embedding representation respectively;

[0024] S33, according to the obtained embedding representation, KNN clustering method is used to search in the spatio-temporal record table to obtain similar historical check-ins;

[0025] Step 3-4, combine the historical check-ins with the current trajectory to obtain an enhanced trajectory.

[0026] Specifically, in S14, the spatio-temporal correlation matrix is obtained based on the time and space interval between each check-in in the enhanced trajectory, and is constructed by using linear interpolation method.

[0027] Specifically, in S14, the trajectory embedding vector is calculated by using the self-attention mechanism expanded by adding spatio-temporal information.

[0028] Specifically, in S15, the POI recommendation result is obtained as follows:

[0029] S51, using a score calculation function, the obtained trajectory embedding vector and all candidate POI embedding vectors are calculated to obtain the probability score corresponding to the candidate POI embedding vector;

[0030] S52, according to the probability score, the candidate POI embedding vector is sorted to obtain the top-K POIs as the POI recommendation result output.

[0031] According to a second aspect of the object of the embodiments of the present application, a POI recommendation device is provided, which is implemented by the POI recommendation method based on the multi-element relationship spatio-temporal network.

[0032] The acquisition module is configured to acquire the user, the POI and the historical trajectory of the user in the LBSN.

[0033] The construction module is configured to generate a corresponding multi-element explicit POI relationship graph according to the POI and the historical trajectory.

[0034] The POI embedding module is configured to obtain a candidate POI embedding vector set by fusion based on the generated multi-element explicit POI relationship graph.

[0035] The spatio-temporal enhancement module is configured to generate an enhanced trajectory and a corresponding spatio-temporal correlation matrix.

[0036] The trajectory embedding module is configured to learn a trajectory representation by using the spatio-temporal correlation matrix and the enhanced trajectory, and obtain a trajectory embedding vector.

[0037] The recommendation module is configured to perform probability score calculation and ranking by using the trajectory embedding vector and all candidate POI embedding vectors, and obtain a final POI recommendation result.

[0038] According to a third aspect of the object of the embodiments of the present application, an electronic device is provided, which comprises:

[0039] One or more processors.

[0040] A memory configured to store one or more programs.

[0041] When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the POI recommendation method based on the multi-element relationship spatio-temporal network.

[0042] Compared with the prior art, the present application has the following beneficial effects:

[0043] (1) A data-driven hypergraph structure is constructed to capture the implicit relationship between POIs, and more accurate representation and prior knowledge are provided for sequential learning.

[0044] (2) A motion-aware spatio-temporal trajectory embedding module is designed to fully extract the movement patterns and preferences of users contained in the trajectory.

[0045] (3) The related historical search method is used to search the relevant historical records to enrich the movement patterns of the user, reduce the noise interference and minimize the computational overhead, and improve the accuracy of the recommendation result. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1is a flow chart of a POI recommendation method based on a multi-element relationship spatio-temporal network according to an exemplary embodiment;

[0047] Figure 2 is a block diagram of a POI recommendation device based on a multi-element spatio-temporal network according to an exemplary embodiment;

[0048] Figure 3 is a schematic diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0049] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, the same numbers are used to indicate the same or similar components. The embodiments described in the following exemplary embodiments are not meant to represent all embodiments consistent with the present application.

[0050] The terminology used in the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used in the description of the application and the appended claims, the singular forms "a", "an" and "the" are intended to include plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0051] It should be understood that although the terms first, second, third, etc. can be used herein to describe various information, these terms are not intended to denote a particular order or hierarchy. These terms are used only to distinguish one sign-in from another. For example, a first sign-in can be termed a second sign-in, and similarly, a second sign-in can also be termed a first sign-in, without departing from the scope of the present application. Depending on the context, the word "if' as used herein can be interpreted to mean "when" or "in response to determining" or "in response to a determination".

[0052] As shown in FIG. 1, the method provided by the present embodiment includes the following steps: Figure 1

[0053] S11, obtaining users, POIs and historical trajectories of each user in the LBSN, wherein the historical trajectory of the user is composed of a plurality of sign-ins, and each sign-in r is a set of triples containing spatio-temporal information;

[0054] S12, constructing a plurality of explicit POI relationship graphs according to the POIs and the historical trajectories, learning implicit relationships of the explicit POI relationships based on a hypergraph architecture to obtain a candidate POI embedding vector set corresponding to all POIs;

[0055] ​​S13, obtaining a check-in set of the same space-time position from a historical trajectory according to a space-time position where a user last checked in, and adding the check-in set to a current trajectory to obtain an enhanced trajectory;

[0056] S14, constructing a corresponding space-time correlation matrix according to the enhanced trajectory to obtain a trajectory embedding vector;

[0057] S15, performing prediction according to the trajectory embedding vector and the candidate POI embedding vector set to obtain a POI recommendation result.

[0058] From the above embodiments, it can be seen that the application proposes a multi-element relationship space-time attention network for a POI recommendation task. On the one hand, a data-driven hypergraph structure is constructed to capture the implicit relationship between POIs, to provide more accurate representation and prior knowledge for sequential learning. On the other hand, a motion-aware space-time trajectory embedding module is designed to fully extract the movement patterns and preferences of users contained in the trajectory. Finally, the related historical search method is used to enrich the movement patterns of users by searching the relevant historical records, to reduce the noise interference and minimize the computational overhead, and to improve the accuracy of the recommendation result.

[0059] In the implementation of step S11: obtaining users in an LBSN, POIs, and historical trajectories of each user, wherein the user data is represented as a user set U = {u1, u2, …, u |U|}, the POIs are represented as a POI set L = {l1, l2, …, l |L|}, and the POIs usually refer to the locations of various entities, such as a certain restaurant or a certain store. The historical trajectory of the user is composed of a plurality of check-ins, and represents the historical trajectory of the user u before time t i , which is a check-in sequence arranged in chronological order. Each check-in r is a triple containing space-time information, and each triple r = (u, l i , t i ) is composed of a user u, a POI l i with unique geographical longitude and latitude coordinates, and a discrete time frame t i . In a popular way, the triple represents that the user u visited the POI l i at time t i . After defining the representation of the POI, the user, and the user trajectory, the historical trajectory of the user to be predicted , a candidate POI set L = {l1, l2, …, l |L|}, and a current time t i are input. The POI recommendation task is to recommend POIs for the user at t i+1A list of POIs is recommended, each POI is associated with a score, the higher the score, the higher the probability that the user wants to visit the corresponding POI.

[0060] In the implementation of S12:

[0061] S21, initialize a learnable embedding vector for each POI.

[0062] S22, judge the distance of the geographical position of each two POIs, if less than the geographical distance threshold, it is a neighbor node to construct a distance relationship graph.

[0063] Specifically, initialize a graph G d =(V d ,E d ), where V d represents all POI nodes, E d represents edges with geographical neighbor relationship, and the adjacency matrix of G d is represented as A d ∈R N×N , set a threshold δ d =2km, if the geographical distance between POI l i and POI l j is less than δ d , then A d [i,j]=1, otherwise A d [i,j]=0.

[0064] S23, based on the two consecutive check-ins in the historical trajectory as the neighbor node to construct the relationship graph, and replace the node with the corresponding POI according to the spatiotemporal information of the check-in, to obtain the conversion relationship graph.

[0065] Specifically, initialize a graph G t =(V t ,E t ), where V t represents all POI nodes, E t represents edges with geographical neighbor relationship, and the adjacency matrix of G t is represented as A t ∈R N×N , where A t [i,j]=freq(l i ,l j ), freq(l i ,l j ) represents the conversion frequency between POI l i and POI l j in all trajectories.

[0066] S24, performing weighted fusion based on the distance relation graph and the conversion relation graph to obtain a candidate POI embedding vector set.

[0067] Specifically, first, the two generated explicit relation graphs G t and G d are mapped to dense vector representations of POI nodes by using the initialization embedding vectors in S21. Meanwhile, the M-dimensional space is divided into K parts, denoted as representing the vector representation of the Kth part. Next, a graph convolution method based on decoupled learning is used to perform graph embedding learning on G t and G d , specifically, the embedding process can be represented as where D is the degree matrix, I is the identity matrix, represents the normalized adjacency matrix, and k represents the weight matrix for calculating the Kth part of the latitude in graph convolution. Finally, two different POI embeddings are obtained, the embedding H d based on geographic distance and the embedding H t based on conversion relation. To obtain the implicit multi-element relation graph, we first maintain a low-latitude conversion matrix Then, using the two POI embeddings H d and H t , the adjacency matrix A h of the hypergraph is generated as A h = (H·W T ) h where W (*) represents the weight matrix in hypergraph learning.

[0068] The information propagation process on the hypergraph is divided into two steps, aggregation and propagation. First, the aggregation function is used to aggregate the attributes of different POIs according to their influence weights in each hyperedge to generate the representation of the hyperedge, so as to capture the internal association in each implicit relation, which can be represented as the following function: Next, H (*) represents one of H d and H t . To obtain the embedding update of POI under the influence of different multi-element relations, the influence of different relations is fed back to the participating POI nodes to obtain the POI embedding Finally, to obtain the POI embedding under the influence of different relations, the POI embeddings obtained under different relation graphs are fused by weighted addition to obtain the final POI embedding X, specifically, the process is represented as X = w3[w1H d +(1-w1)H t]+(1-w3)[w2H ′ d +(1-w2)H t ′ Here, w1, w2, and w3 represent three weighted parameters.

[0069] In the specific implementation of S13:

[0070] S31: Create a spatiotemporal record table based on the user's complete historical trajectory.

[0071] Specifically, to fully record the spatiotemporal information of all trajectories, a spatiotemporal table (STT) of user historical trajectories was created. This table is used to enrich the user's movement patterns by filling in historical check-in records. Each record is stored as a 5-tuple: The quintuple represents user u i A visit to POI l at location d occurs at time t, where l is the POI's ID, d = (lon, lat) represents a geographic latitude and longitude location, j represents the j-th visit in the entire trajectory, and t is a timestamp. To prevent information loss from similar but different check-in records, we do not save the check-in records in their original form. Instead, we use the aforementioned POI and trajectory embedding to project and transform l, t, and d into the embedding space, obtaining... Finally, a spacetime table is generated using the mapped records. Here yes The abbreviation represents user u i All the sign-in records after projection transformation.

[0072] S32: Based on the user's current spatiotemporal location, map the time and location information therein to a high-dimensional embedded representation respectively;

[0073] Specifically, we approximate a user's current state using their most recent check-in record. In this way, STT allows us to search for users' historical check-in records that are similar to their current state.

[0074] S33: Based on the obtained spatiotemporal embedding representation, the KNN clustering method is used to perform an aggregation search in the spatiotemporal record table to obtain relevant historical check-in records;

[0075] Specifically, given a user's recent historical trajectory We will check in the most recent Mapping to a vector space to initialize a query q n =(E n E t (t n ),E d(d n And use this query to search in STT to get user u i Mem(u) is a collection of similar historical check-ins. i ) = Search(STT(u i ),q n The Search function can be replaced with different clustering algorithms, such as DBSCAN and KNN. Here we use multi-condition KNN, where E n and E d (d n As two conditions, a threshold β is set. If the similarity between a historical record and a recent record is higher than β, the historical record is added to a temporary set. After clustering, the top-k similar records are selected from the temporary set as the final result.

[0076] S34: Combine the relevant historical check-in records with the current trajectory to obtain the memory-enhanced trajectory.

[0077] Specifically, we combine the obtained relevant historical check-ins into the original sequence to compensate for the loss of long-cycle patterns. The process is represented as follows: Ultimately, user u i The trajectory H after memory enhancement ′ (u i ).

[0078] In the specific implementation of S14:

[0079] S41. Calculate the time and space intervals between different check-ins based on the geographical and time information of each check-in in the enhanced trajectory.

[0080] Specifically, input the enhanced trajectory First, each check-in is mapped to a vector space using the final POI embedding X obtained in S13, thus obtaining... in Representative sign-in The mapped vector representation.

[0081] Next, the relative time interval for each check-in in the trajectory is calculated. This represents the relative time length between the i-th visit and the j-th visit in the trajectory, where Represents trajectory S u The minimum time interval (excluding 0).

[0082] On the other hand, the Haversine algorithm is used to calculate the relative spatial intervals in longitude and latitude between each check-in. denotes the relative distance of the spatial interval between li and lj in the longitude coordinate transformation direction; The calculation method is similar, which denotes the relative distance in the latitude coordinate transformation direction, where denotes S u The minimum spatial interval (0 excluded) in the latitude and longitude directions.

[0083] S42, using the time and space interval, constructing a space-time correlation matrix based on linear interpolation method.

[0084] Specifically, using the above generated t i,j and Three kinds of time-space interval matrices are constructed, and a time threshold δ t and two space thresholds The three threshold values are used to cut the three kinds of time-space interval matrices constructed to prevent excessive intervals from affecting convergence. Next, a learnable time embedding dictionary and two learnable space embedding dictionaries and The three kinds of time-space matrices that have been cut are mapped to convert the sparse and discrete interval values to dense and continuous low-dimensional space, alleviate the sparse representation problem, and add the space interval matrices in two coordinates to finally obtain a space interval matrix that can express any direction and a time interval matrix.

[0085] S43, input the space-time correlation matrix and enhanced trajectory sequence into the trajectory embedding module, and calculate the trajectory embedding by using the self-attention mechanism extended by adding space-time information.

[0086] Specifically, according to the obtained space interval matrix and time interval matrix and enhanced trajectory sequence, the self-attention mechanism is used to aggregate space-time information to update the representation of each POI in the trajectory where W V denotes the model parameter weight matrix, p j denotes the position embedding of the jth position, to capture the interaction between POI visits and obtain the space-time dependency between user preference visits considering different relationships between POIs and POIs, where each denotes the influence weight between different POIs, where and respectively denote the elements in the time and space interval matrices constructed above, M denotes the dimension of the element, W Q and W K are two model weight matrices.

[0087] In the implementation of S15, the input candidate POI embedding vector and the obtained trajectory vector are used to calculate the correlation between the trajectory vector and different candidate POI embedding vectors where x t ′ represents the vector representation X of the POI visited at time t l represents the embedding vector of the candidate POI, and finally ranked according to the obtained correlation, and the final POI recommendation list is obtained according to the correlation from large to small.

[0088] The embodiment provides a POI recommendation device, which is realized by the POI recommendation method provided in the above embodiment.

[0089] As shown in Figure 2 includes the following modules:

[0090] The acquisition module is configured to acquire user, POI data and historical trajectory of each user in the LBSN (Location Based Social Network), and the trajectory includes a plurality of check-ins, wherein each check-in is composed of a user, a POI and a time;

[0091] The construction module is configured to construct a plurality of explicit POI relation graphs according to the check-in information of the POI data and the historical trajectory, and design a hypergraph architecture to adaptively learn an implicit POI multivariate relation graph.

[0092] The POI embedding module is configured to aggregate information in different types of graphs, and obtain a final POI embedding by using two different graph message passing methods and weighted fusion.

[0093] The space-time enhancement module is configured to enhance the trajectory by using the existing space-time location search related historical records of the user, and construct a space-time correlation matrix.

[0094] The trajectory embedding module is configured to learn trajectory representation by using the space-time correlation matrix and the enhanced trajectory, and obtain a trajectory embedding vector.

[0095] The recommendation module is configured to perform probability score calculation and ranking by using the trajectory embedding vector and all candidate POI embedding vectors, and obtain a final POI recommendation result.

[0096] For the device embodiment, since it basically corresponds to the method embodiment, the relevant part can be seen from the part of the method embodiment. The device embodiment described above is only illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the application scheme according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0097] Correspondingly, the application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the POI recommendation device method based on the multi-element relationship space-time network as described above.

[0098] As shown in Figure 3 , a hardware structure diagram of any data processing capable device for a POI recommendation method based on a multi-element relationship space-time network provided by the embodiment, in addition to the processor, memory and network interface as shown in Figure 3 , the any data processing capable device where the device in the embodiment is usually based on the actual function of the any data processing capable device, and can also include other hardware, and this will not be described again.

[0099] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. The application is intended to cover any variations, uses or adaptations of the application following, in general, the principles of the application and including such departures from the present disclosure as come within known or customary practice in the art to which the application pertains.

[0100] It should be understood that the application is not limited to the precise construction that has been described above and illustrated in the accompanying drawings, and that various modifications and changes can be made without departing from the scope thereof.

Claims

1. A POI recommendation method based on a multi-element relationship spatio-temporal network, characterized in that, Comprise: S11, obtain user information in LBSN, POI and user's historical trajectory, the historical trajectory is composed of multiple check-ins, wherein each check-in is a set of three tuples containing space-time information; S12, construct multiple explicit POI relationship graphs according to POI and historical trajectory, learn implicit relationship of explicit POI relationship based on hypergraph architecture to obtain candidate POI embedding vector set corresponding to all POIs, and the specific steps are: S21, initialize learnable embedding vector for each POI; S22, judge the distance of the geographical position in each two POIs, if less than the geographical distance threshold, then as a neighbor node to construct a distance relationship graph; S23, based on two consecutive check-ins in historical trajectory as neighbor nodes to construct a relationship graph, and replace the nodes with corresponding POIs according to the space-time information of check-in, to obtain a conversion relationship graph; S24, based on the distance relationship graph and the conversion relationship graph, weighted fusion is carried out to obtain the candidate POI embedding vector set; S13, according to the space-time position of the last check-in of the user, obtain the check-in set of the same space-time position from the historical trajectory and add the check-in set to the current trajectory to obtain an enhanced trajectory; S14, construct a corresponding space-time association matrix according to the enhanced trajectory to obtain a trajectory embedding vector; S15, according to the trajectory embedding vector and the candidate POI embedding vector set, prediction is carried out to obtain a POI recommendation result. 2.The POI recommendation method based on the multi-element relationship spatio-temporal network according to claim 1, characterized in that, In S13, the enhanced trajectory is obtained as follows: S31, create a corresponding space-time record table according to the historical trajectory; S32, according to the space-time position of the last check-in of the user, the corresponding time and geographical information are mapped to high-dimensional embedding representation respectively; S33, according to the obtained embedding representation, KNN clustering method is used to search in the space-time record table to obtain similar historical check-in; S34, combine the historical check-in with the current trajectory to obtain an enhanced trajectory. 3.The POI recommendation method based on the multi-element relationship space-time network according to claim 1, characterized in that, In S14, the space-time association matrix is obtained by using linear interpolation method based on the time and space interval between each check-in in the enhanced trajectory. 4.The POI recommendation method based on the multi-element relationship space-time network according to claim 1, wherein, In S14, the trajectory embedding vector is obtained by using self-attention mechanism expanded by adding space-time information. 5.The POI recommendation method based on the multi-element relationship space-time network according to claim 1, wherein, In S15, the POI recommendation result is obtained as follows: S51, use score calculation function to calculate the obtained trajectory embedding vector and all candidate POI embedding vectors to obtain probability score corresponding to the candidate POI embedding vector; S52, sort the candidate POI embedding vector according to the probability score to obtain the top-K POI as the POI recommendation result output.

6. A POI recommendation apparatus characterized by comprising: The POI recommendation method based on multi-element relationship space-time network is realized by the method of any one of claims 1-5, comprising: An acquisition module is configured to acquire users in a LBSN, POIs and historical trajectories of the users; A construction module is configured to generate corresponding multiple explicit POI relationship graphs according to POIs and historical trajectories; A POI embedding module is configured to obtain a candidate POI embedding vector set by fusion based on the generated multiple explicit POI relationship graphs. a spatio-temporal enhancement module configured to generate an enhanced trajectory and a corresponding spatio-temporal correlation matrix; a trajectory embedding module configured to learn trajectory representations using the spatio-temporal correlation matrix and the enhanced trajectory to obtain a trajectory embedding vector; a recommendation module configured to perform probability score calculation and ranking using the trajectory embedding vector and embedding vectors of all candidate POIs to obtain a final POI recommendation result.

7. An electronic device, comprising: comprising: one or more processors; a memory configured to store one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the POI recommendation method based on the multi-element relationship spatio-temporal network according to any one of claims 1-5.

Images