A next point of interest recommendation method based on adaptive graph and future preference

Abstract

The application discloses a next interest point recommendation method based on adaptive graph and future preference, comprising the following steps: preparing an interest point historical check-in data set, the data set comprising user historical check-in places, check-in time, check-in place categories and the like information, and performing standardization processing on the data; initializing POI node embedding information, constructing a global adjacency matrix by using an adaptive graph, and obtaining final interest point embedding expression by using a similarity function and a convolution network; fusing interest point embedding expression with other side information of space-time node coding information, introducing a self-attention mechanism, and coding to obtain long-term preference of a user; then, long-term memory recurrent neural network is used to fuse side information of the interest point, introduce an attention mechanism, and obtain short-term preference of the user; subsequently, the long-term preference and the short-term preference are added to the attention mechanism in a self-assembly manner, the next interest point is captured by using next time moment coding as a query vector. The application selects an adaptive graph and a graph neural network, can capture the correlation of real interest points, and has good recommendation performance.

Description

Technical Field

[0001] This invention belongs to the field of recommender systems and applies adaptive graph learning and future preference to implement a method for recommending the next point of interest. Background Technology

[0002] In today's rapidly modernizing world, the demand for personalized services is increasing, and various leisure and entertainment destinations are emerging in endless streams. The rapid development of location-based social networks (LBSNs) has spurred the integration of recommendation systems with these networks, resulting in Points of Interest (POI) recommendations. The explosive growth of online geographic resources has led to a surge in demand for POI recommendations, propelling their rapid development and providing valuable geographic information for service providers and users. Given the growing demand for real-time and personalized services in business scenarios, the next phase of POI recommendation has attracted widespread attention from academia and industry. With the maturity of technologies such as spatiotemporal big data, the Internet of Things, and deep learning, Points of Interest (POI) recommendations have also developed rapidly. In POI recommendations, there are strong temporal and spatial correlations among the check-in data. Common recommendation methods face significant challenges in capturing user preferences due to the personalization and complexity of user preferences. Furthermore, how to balance the geographical dependencies between global POI nodes is also a direction that urgently needs research.

[0003] POI (Point of Interest) recommendation refers to recommending nearby or interesting locations to users based on their location information and historical behavior. This recommendation system can help users discover new locations, plan travel routes, and find nearby services and facilities. There are two main problems in the recommendation process: First, using GNN-based methods to construct a global POI map. However, GNN-based methods are limited by the inherent limitations of fixed graphs: the construction of the input graph is constrained by prior knowledge, and the selection of statistical indicators leads to a certain degree of incomplete graph structure and noise, restricting the performance of representation learning. Second, existing future preference methods ignore the interaction of attribute features, especially spatiotemporal information. Individual visits are influenced not only by users' travel plans but also by the spatiotemporal relationships between different POIs. Ignoring spatiotemporal weighted information treats all historical check-in behaviors as a uniform contribution, thus weakening the ability to construct other aspects of information features.

[0004] In recent years, deep learning has attracted attention due to its ability to perform high-dimensional nonlinear modeling of data and has achieved good results in recommender systems. However, most models rely on predefined adjacency matrices based on prior knowledge as the foundation for global graph construction. Processing graph structures solely with graph neural networks and predefined adjacency matrices yields poor results. In handling users' future preferences, how to effectively model long-term time-series data and perform related spatiotemporal encoding is also a problem that needs to be solved. Existing models do not fully consider how to use adaptive graph learning and learning users' future preferences to model users' final preferences and obtain recommendation results. Summary of the Invention

[0005] Existing POI recommendation methods mostly rely on GNN graph convolutional networks and statically constructed global POI graphs. However, GNN-based methods are limited by the inherent limitations of fixed graphs: the construction of the input graph is constrained by prior knowledge, and the selection of statistical indicators leads to a certain degree of incomplete graph structure and noise, which restricts the performance of representation learning. Therefore, we propose using adaptive graphs to address this issue. By learning the adjacency matrix of the adaptive graph through similarity metrics, we can automatically infer the potential geographical relationships between POIs. Meanwhile, existing future preference-based recommendation methods mainly do not consider spatiotemporal encoding problems. To solve these problems, the purpose of this invention is to propose a new POI recommendation method based on adaptive graph learning and future preferences.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A next point of interest recommendation method based on adaptive graphs and future preferences includes the following steps:

[0008] Step 1: Prepare the user's historical check-in dataset. The historical dataset includes data values ​​such as the latitude and longitude of the check-in location, the category of the check-in location, and the time of the check-in location. Establish a preprocessing method, construct the user's historical check-in trajectory, and the initial interest point adjacency matrix to obtain the initial input data.

[0009] In step 1, we used three real datasets from New York City, Phoenix, and Singapore. Each dataset contains information such as check-in time, check-in location category, and check-in user ID. Taking Singapore as an example, this dataset contains 8,648 users, 33,172 points of interest, and 355,337 check-in records. The sparsity of this dataset is 0.12%. We selected the time information (including hour and whether it is a weekday), category information, latitude and longitude information, and user information from this data as the main feature values ​​to predict the recommendation sequence of the next point of interest.

[0010] Step 2: Construct the final embedding representation of the points of interest (POIs). The input is the initial embedding vector of the POIs. Using the initial embedding representation of the POIs, cosine similarity is used as the metric function. An adaptive adjacency matrix is ​​learned through the similarity metric to automatically infer the potential geographical relationships between POIs. A threshold is introduced to sparsify the graph. Furthermore, neighborhood information is convolved and aggregated to obtain the final embedding representation of the POIs. And serve as subsequent input.

[0011] In step 2, the overall structure of adaptive graph learning is as follows:

[0012]

[0013] Here we initialize the initial embedding representation of the global POI. d id The feature dimension is represented, and then the dependency between each pair of POIs is inferred by calculating the embedding similarity between two POIs. Cosine similarity is used as the metric function, and finally, each entry A in the adjacency matrix A is obtained. i,j This is a fully connected graph, A i,j ∈(-1,1), the adaptive graph-based terms in this matrix reflect an accurate estimate of the geographic dependencies between points of interest. However, this result leads to excessive fragmented edges, resulting in a noisy structure. Therefore, a threshold is introduced for graph sparsification to improve the dynamic features related to spatial dependencies, thereby obtaining a non-negative adjacency matrix A. i,j ∈(0,1), and to enhance the expressiveness of interest points, neighborhood information is aggregated, and the initial features of the nodes in the graph are convolved, as shown below:

[0014]

[0015] Obtaining higher-order representations of points of interest use As the final embedded expression of points of interest.

[0016] Step 3, Build long-term user preferences Its input consists of the user's historical trajectory data over n-1 days, obtained by concatenating the final embedding representation of interest points with the initial embeddings of other side information. in This represents a user's single-day trajectory formed by multiple embedded data points. Then, a spatial interval matrix and a time interval matrix of the user's embedded trajectory are constructed. The spatiotemporal weight ω is obtained by encoding the spatiotemporal interval matrix. T ,ω D Secondly, by utilizing a self-attention mechanism and taking the user's trajectory data over a specific number of days as input, the association weights between nodes are obtained, thus revealing the user's long-term preferences.

[0017] In step 3, the overall structure of the long-term preference encoder is as follows:

[0018]

[0019] User history trajectory The data is obtained from the user's check-in points on that day. Each user's check-in record in the sequence has three characteristics: user ID (u), POI (Point of Interest) representation, etc. Category ID c. Furthermore, since POIs are closely related to time information, we added two new time information items T to capture user activity at specific times, such as people liking to go to parks on weekends and bars in the evening. Therefore, these two time information items are the weekly ID feature t containing two categories. d and hourly ID features containing 24 categories t h The embedding layer transforms each id feature into a d. id The feature dimension is represented by a dense vector. Then, the pooling layer merges all five embeddings to generate the check-in data x in the sequence, as shown in the following formula:

[0020]

[0021] The dimension of x is d h =5d id The trajectory is represented as Where m is the total number of check-ins in the entire trajectory.

[0022] Subsequently, considering the influence of spatiotemporal environment on long-term user preference modeling, we incorporated the impact of spatiotemporal weights when designing the long-term preference encoder. We then considered the two registration points p... i ,p j The time-space intervals between them are denoted as ΔT. i,j ,ΔD i,j The time-related influence weight and the distance-related influence weight are denoted as W respectively. T (ΔT i,j ),W D ΔD i,j Specifically, we use the Havercosine periodic function to capture user time-period data and set a time decay weight, as shown in the following formula:

[0023] W period (ΔT i,j )=hvc(2πΔT i,j )

[0024]

[0025] Where α is the time decay rate, and the control weight varies with time interval ΔT. i,j The rate of descent.

[0026] From a spatial perspective, we believe that the closer a POI is to the current location, the more helpful it is for location prediction. Therefore, the distance-exponential decay weighting method represents:

[0027]

[0028] Where β is the time decay rate, controlling the weight as a function of distance interval ΔD. i,j The rate of descent.

[0029] Finally, we employ a self-attention mechanism to encode long-term preferences, which are obtained by fusing multiple lateral and spatiotemporal information, as shown in the following formula:

[0030]

[0031] in Both belong to the transformation weight matrix.

[0032] Step 4, construct user long-term preference H short The input is the user's historical trajectory for that day. Using a Long Short-Term Memory (LSTM) recurrent neural network, a user's short-term preference H is constructed. short .

[0033] In step 4, short-term preference H short The overall structure is as follows:

[0034]

[0035] The input is the user's historical trajectory for the day. The short-term preference encoder captures the user's time-perceived order dependency from the user's current check-in behavior, and obtains the user's short-term preference as H. short .

[0036] The calculation of the short-term preference encoder can be defined as follows:

[0037] We utilize the current trajectory S short =(x1,x1,...x m As input data, it is fed into the LSTM network, and the specific formula is as follows:

[0038]

[0039] in, and They are timestamps t i With t i-1 The hidden state.

[0040] Step 5: Construct user future preferences. Considering the user's self-planned activities, an attention mechanism is introduced. Future time-embedded expressions are used as query vectors, and the hidden state matrix of long-term preferences is used as content vectors. Potential future behaviors are extracted from the user's daily cyclical behaviors to obtain future time preference weights. This implicitly simulates the user's activity schedule. The obtained time preference weights are compared with short-term preference H. short By combining these methods and using user embeddings as query vectors, we can obtain the user's future preferences H. f We learn the probability distribution of user preferences to obtain the final prediction results.

[0041] In step 5, the future preference H f The overall structure is as follows:

[0042]

[0043] Where input X is the future time preference weight obtained by fusing future behavior and long-term encoding. Because users may have their own activity plans when they are out and about, time-aware attention can be used to identify behaviors in the long-term past trajectory that are most relevant to the future time context. Specifically, the user's long-term trajectory Encoded as The hidden state matrix extracts future time preference weights by considering latent future behaviors derived from users' daily cyclical actions. Future time embeddings are then used. k+1 As the query vector, the formula is as follows:

[0044]

[0045] Then, the weights are combined with short-term preferences using a join method to generate hidden states that represent user preferences. The user's ultimate future preference is determined by The user's final future preference is thus summed up from each hidden state in the dataset:

[0046]

[0047] Where u represents the user embedding representation. Next, we learn the probability distribution of user preferences and decode the points of interest:

[0048]

[0049] in It is a transformation matrix. Attached Figure Description

[0050] Figure 1 This is a diagram illustrating the overall structure of adaptive graph learning in this invention.

[0051] Figure 2 This is an overall structural diagram of the long-term preference encoder in this invention;

[0052] Figure 3 This is a structural diagram of the short-term preference encoder constructed in this invention;

[0053] Figure 4 This is an overall structural diagram of user future preferences in this invention;

[0054] Figure 5 This is a diagram showing the overall structure of the model in this invention; Detailed Implementation

[0055] The present invention will be further explained below with reference to the accompanying drawings in the embodiments of the present invention.

[0056] The present invention provides a next point of interest recommendation method based on adaptive graphs and future preferences, comprising the following steps:

[0057] Step 1: Prepare the user's historical check-in dataset. The historical dataset includes data values ​​such as the latitude and longitude of the check-in location, the category of the check-in location, and the time of the check-in location. Establish a preprocessing method, construct the user's historical check-in trajectory, and the initial interest point adjacency matrix to obtain the initial input data.

[0058] Step 2: Construct the final embedding representation of the points of interest (POIs). The input is the initial embedding vector of the POIs. Using the initial embedding representation of the POIs, cosine similarity is used as the metric function. An adaptive adjacency matrix is ​​learned through the similarity metric to automatically infer the potential geographical relationships between POIs. A threshold is introduced to sparsify the graph. Furthermore, neighborhood information is convolved and aggregated to obtain the final embedding representation of the POIs. And serve as subsequent input.

[0059] Step 3, Build long-term user preferences Its input consists of the user's historical trajectory data over n-1 days, obtained by concatenating the final embedding representation of interest points with the initial embeddings of other side information. in This represents a user's single-day trajectory formed by multiple embedded data points. Then, a spatial interval matrix and a time interval matrix of the user's embedded trajectory are constructed. The spatiotemporal weight ω is obtained by encoding the spatiotemporal interval matrix. T ,ω D Secondly, by utilizing a self-attention mechanism and taking the user's trajectory data over a specific number of days as input, the association weights between nodes are obtained, thus revealing the user's long-term preferences.

[0060] Step 4, construct user long-term preference H short The input is the user's historical trajectory for that day. Using a Long Short-Term Memory (LSTM) recurrent neural network, a user's short-term preference H is constructed. short .

[0061] Step 5: Construct user future preferences. Considering the user's self-planned activities, an attention mechanism is introduced. Future time-embedded expressions are used as query vectors, and the hidden state matrix of long-term preferences is used as content vectors. Potential future behaviors are extracted from the user's daily cyclical behaviors to obtain future time preference weights. This implicitly simulates the user's activity schedule. The obtained time preference weights are compared with short-term preference H. short By combining these methods and using user embeddings as query vectors, we can obtain the user's future preferences H. f We learn the probability distribution of user preferences to obtain the final prediction results.

[0062] In step 1, we used three real datasets from New York City, Phoenix, and Singapore. Each dataset contains information such as check-in time, check-in location category, and check-in user ID. Taking Singapore as an example, this dataset contains 8,648 users, 33,172 points of interest, and 355,337 check-in records. The sparsity of this dataset is 0.12%. We selected the time information (including hour and whether it is a weekday), category information, latitude and longitude information, and user information from this data as the main feature values ​​to predict the recommendation sequence of the next point of interest.

[0063] In step 2, please refer to the overall structure of adaptive graph learning. Figure 1 The overall structure of adaptive graph learning is as follows:

[0064]

[0065] Here we initialize the initial embedding representation of the global POI. d id The feature dimension is represented, and then the dependency between each pair of POIs is inferred by calculating the embedding similarity between two POIs. Cosine similarity is used as the metric function, and finally, each entry A in the adjacency matrix A is obtained. i,j This is a fully connected graph, A i,j ∈(-1,1), the adaptive graph-based terms in this matrix reflect an accurate estimate of the geographic dependencies between points of interest. However, this result leads to excessive fragmented edges, resulting in a noisy structure. Therefore, a threshold is introduced for graph sparsification to improve the dynamic features related to spatial dependencies, thereby obtaining a non-negative adjacency matrix A. i,j∈(0,1), and to enhance the expressiveness of interest points, neighborhood information is aggregated, and the initial features of the nodes in the graph are convolved, as shown below:

[0066]

[0067] Obtaining higher-order representations of points of interest use As the final embedded expression of points of interest.

[0068] In step 3, please refer to the overall structure of the long-term preference encoder. Figure 2 Long-term preference for encoders

[0069] The body structure is as follows:

[0070]

[0071] User history trajectory The data is obtained from the user's check-in points on that day. Each user's check-in record in the sequence has three characteristics: user ID (u), POI (Point of Interest) representation, etc. Category ID c. Furthermore, since POIs are closely related to time information, we added two new time information items T to capture user activity at specific times, such as people liking to go to parks on weekends and bars in the evening. Therefore, these two time information items are the weekly ID feature t containing two categories. d and hourly ID features containing 24 categories t h The embedding layer transforms each id feature into a d. id The feature dimension is represented by a dense vector. Then, the pooling layer merges all five embeddings to generate the check-in data x in the sequence, as shown in the following formula:

[0072]

[0073] The dimension of x is d h =5d id The trajectory is represented as Where m is the total number of check-ins in the entire trajectory.

[0074] Subsequently, considering the influence of spatiotemporal environment on long-term user preference modeling, we incorporated the impact of spatiotemporal weights when designing the long-term preference encoder. We then considered the two registration points p... i ,p j The time-space intervals between them are denoted as ΔT. i,j ,ΔD i,j The time-related influence weight and the distance-related influence weight are denoted as W respectively. T (ΔT i,j ),W D ΔD i,jSpecifically, we use the Havercosine periodic function to capture user time-period data and set a time decay weight, as shown in the following formula:

[0075] W period (ΔT i,j )=hvc(2πΔT i,j )

[0076]

[0077] Where α is the time decay rate, and the control weight varies with time interval ΔT. i,j The rate of descent.

[0078] From a spatial perspective, we believe that the closer the POI is to the current location, the more helpful it is for location prediction.

[0079] The larger the distance, the more helpful it is for location prediction. Therefore, the exponential decay weighting of distance represents:

[0080]

[0081] Where β is the time decay rate, controlling the weight as a function of distance interval ΔD. i,j The rate of descent.

[0082] Finally, we employ a self-attention mechanism to encode long-term preferences, which incorporate multiple side signals.

[0083] The information and spatiotemporal information are obtained, and the formula is as follows:

[0084]

[0085] in Both belong to the transformation weight matrix.

[0086] In step 4, please refer to the overall structure of the short-term preference encoder. Figure 3 Short-term preference H short Overall structure

[0087] The structure is as follows:

[0088]

[0089] The input is the user's historical trajectory for the day. The short-term preference encoder captures the user's time-perceived order dependency from the user's current check-in behavior, and obtains the user's short-term preference as H. short .

[0090] The calculation of the short-term preference encoder can be defined as follows:

[0091] We utilize the current trajectory S short =(x1,x1,...xm As input data, it is fed into the LSTM network, and the specific formula is as follows:

[0092]

[0093] in, and They are timestamps t i With t i-1 The hidden state.

[0094] In step 5, please refer to the overall structure of future preferences. Figure 4 Future preference H f The overall structure is as follows:

[0095]

[0096] Where input X is the future time preference weight obtained by fusing future behavior and long-term encoding. Because users may have their own activity plans when they are out and about, time-aware attention can be used to identify behaviors in the long-term past trajectory that are most relevant to the future time context. Specifically, the user's long-term trajectory Encoded as The hidden state matrix extracts future time preference weights by considering latent future behaviors derived from users' daily cyclical actions. Future time embeddings are then used. k+1 As the query vector, the formula is as follows:

[0097]

[0098] Then, the weights are combined with short-term preferences using a join method to generate hidden states that represent user preferences. The user's ultimate future preference is determined by The user's final future preference is thus summed up from each hidden state in the dataset:

[0099]

[0100] Where u represents the user embedding representation. Next, we learn the probability distribution of user preferences and decode the points of interest:

[0101]

[0102] in It is a transformation matrix.

[0103] In terms of dataset processing, unlike general traffic flow prediction algorithms, this invention uses real traffic flow datasets for training, enhancing the algorithm's real-world usability. Regarding network structure, this invention employs a multi-head attention-spreading graph convolutional network to dynamically simulate traffic flow patterns, improving the accuracy of traffic flow prediction. In terms of prediction objectives, this invention selects traffic flow prediction for the near future as its target, making the predictions more accurate and real-time. Experiments show that the model provided by this method can effectively predict traffic flow, and the final results are realistic.

[0104] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A next point of interest recommendation method based on adaptive graph and future preference, comprising the following steps: Step 1, preparing a user historical check-in dataset, the historical dataset containing check-in location latitude and longitude, check-in location category, check-in location time and other data values, establishing a preprocessing method, constructing the user's historical check-in trajectory, and an initial point of interest adjacency matrix to obtain the initial input data; Step 2, the final embedding expression of the point of interest is constructed, the input of which is the initial embedding vector of the point of interest, the initial embedding expression of the point of interest is used, the cosine similarity is taken as a measurement function, the adjacency matrix of the adaptive graph is learned through the similarity measurement, so as to automatically infer the potential geographical relationship between the points of interest and introduce a threshold to sparsify the graph, and on this basis, the neighborhood information is convoluted and aggregated to obtain the final embedding expression of the point of interest and as subsequent input; Step 3, constructing user long-term preference The input is the historical trajectory data of the user in the past n-1 days, which is composed of the input data x obtained by connecting the final embedding expression of the interest point and the initial embedding of other side information Wherein The user single-day user trajectory formed by a plurality of embedding data is represented, and then the spatial interval matrix and the time interval matrix of the user embedding trajectory are constructed, and the time-space weight ω is obtained by encoding the time-space interval matrix T ,ω D Secondly, using the self-attention mechanism, the correlation weight between nodes is obtained by taking the trajectory data of the user in a specific day as the input, and the long-term preference of the user is obtained Step 4, build user long-term preference H short with input of user's historical trajectory of the day Build user's short-term preference H with long short-term memory recurrent neural network LSTM shor; Step 5, construct user future preference, consider user's self-activity plan, introduce attention mechanism, express as query vector through future time embedding, take hidden state matrix of long-term preference as content vector, extract potential future behavior from user's daily periodic behavior, obtain future time preference weight, which implicitly simulates user's activity schedule Combine the obtained time preference weight with short-term preference H short , and take user embedding as query vector to obtain user future preference H f , learn the probability distribution of user preference, and obtain the final prediction result. 2.The next POI recommendation method based on adaptive graph and future preference according to claim 1, characterized in that: In step 1, we use three real datasets from New York City, Phoenix City and Singapore, each dataset containing check-in time, check-in location category, check-in user id and other information. Taking Singapore City as an example, the dataset contains 8648 users, 33172 points of interest and 355337 check-in data. The sparsity of the dataset is 0.12%. We select the time information in the data, including hour and whether it is a weekday, category information, latitude and longitude information, and user information as the main feature value to predict the recommended sequence of the next point of interest. 3.The next POI recommendation method based on adaptive graph and future preference of claim 1, wherein: In step 2, the overall structure of adaptive graph learning is as follows: input initial point of interest embedding → adaptive graph learning → final embedding representation of points of interest where we initialize the initial embedding representation of global POIs d id representing the feature dimension, then the dependency between each pair of POIs is inferred by computing the embedding similarity between two POIs; cosine similarity is used as the metric function, and finally each entry A i,j in the adjacency matrix A i,j ∈(-1,1), which is a fully connected graph, A i,j ∈(-1,1), the entries in this matrix reflect an accurate estimation of the geographical dependency between POIs; but this result will lead to too many trivial edges, resulting in a noisy structure; therefore, a threshold is introduced to sparsify the graph, the purpose is to refine the dynamic features related to spatial dependency, to obtain a non-negative adjacency matrix A i,j ∈(0,1), and in order to enhance the expressiveness of POIs, the neighborhood information is aggregated, and the initial features of the nodes in the graph are convolved, expressed as follows: Obtaining higher-order representations of points of interest using as a final embedding representation of the point of interest. 4.The next POI recommendation method based on adaptive graph and future preference according to claim 1, characterized in that: In step 3, the overall structure of the long-term preference encoder is as follows: User history trajectory → long-term preference encoder → where the user historical trajectory is obtained from the user's check-in point data, and each user's check-in record in the sequence has three features: user id u, POI represents category id c; In addition, since the POI is closely related to the time information, we add two new time information T to capture the user's activities at a specific time, for example, people like to go to the park on weekends and go to the bar at night; Therefore, the two time information are t d week id features containing 2 categories and t h; hour id features containing 24 categories. id The embedding layer converts each id feature into a d id dimensional dense vector, representing its feature dimension, and then the pooling layer combines all five embeddings to generate the check-in data x in the sequence, as shown in the following formula: The dimension of x is d h = 5d id The trajectory is represented as where m is the total number of check-ins in the entire trajectory. After that, considering that the user long-term preference modeling is affected by the spatio-temporal environment, we consider the influence of spatio-temporal weights when designing the long-term preference encoder. We denote the time interval and distance interval between two registration points p i ,p j as ΔT i,j ,ΔD i,j , respectively, and the time influence weight and distance influence weight as W T (ΔT i,j ),W D ΔD i,j , respectively. Specifically, we use the Havercosine periodic function to capture the user's time period data, and set the time decay weight, as follows: W period (ΔT i,j )=hvc(2πΔT i,j ) where a is a time decay rate, controlling the weight of the time interval DT i,j the speed of the descent; From a spatial point of view, we believe that the closer the POI is to the current location, the more helpful it is to the location prediction, and therefore the distance exponential decay weight is used to represent: where β is a temporal decay rate, controlling how the weight decays with distance interval ΔD i,j the speed of the descent; Finally, we use self-attention mechanism to encode long-term preference, and long-term preference is obtained by fusing multiple side information and spatio-temporal information, as follows: wherein all belong to the transform weight matrix. 5.The next POI recommendation method based on adaptive graph and future preference according to claim 1, characterized in that, In step 4, the short-term preference H short The overall structure is as follows: Day history track → short-term preference encoder → H short wherein the input is the historical trajectory of the user on the day, the short-term preference encoder captures the user's time-aware sequential dependency from the user's current check-in behavior, and obtains the user's short-term preference as H short ; The calculation of the short-term preference encoder can be defined as follows: We use the current trajectory S short = (x1, x1,... x m ) as input data, which is fed into the LSTM network, with the specific formula as follows: wherein, and are the timestamps t i and the hidden state of t i-1 , respectively. 6.The next POI recommendation method based on adaptive graph and future preference according to claim 1, characterized in that, The future preference H f The overall structure is as follows: input X → long-term preference encoding → where input X is the future time preference weight obtained by fusing future behavior and long-term encoding As the user's out-of-home activities are considered to be planned by the user, the time-aware attention can be used to identify the most relevant behaviors in the past long-term trajectory to the future temporal context. Specifically, the user's long-term trajectory is encoded as a hidden state matrix, the future time preference weight is obtained by extracting the latent future behavior from the user's daily periodic behavior, using the future time embedding t k+1 as the query vector, the formula is as follows: The weight is then combined with the short-term preference by concatenation to generate a hidden state representing the user's preference The user's final future preference is summarized from each hidden state in the set, so the user's final future preference is: Where u represents the user embedding representation, and then we learn the probability distribution of user preference to decode the point of interest: wherein is a transformation matrix.

Images