A spatiotemporal perception-based point of interest recommendation method
By constructing a trajectory flow graph and a spatiotemporal awareness module, the system extracts users' short-term and long-term travel preferences, solving the problems of data sparsity and nonlinear spatiotemporal dependencies in point-of-interest (POI) recommendation, and achieving flexible and accurate POI recommendation.
Patent Information
- Application Number
- CN202311312887.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-11
- Publication Date
- 2025-11-04
- Estimated Expiration
- 2043-10-11
AI Technical Summary
Existing point-of-interest recommendation methods suffer from the cold start problem caused by data sparsity, cannot effectively capture the nonlinear spatiotemporal dependencies between non-adjacent check-ins, and struggle to handle the complexity and dynamism of user travel preferences.
We employ a spatiotemporally aware point-of-interest recommendation method. By constructing a trajectory flow graph, a spatiotemporally aware attention module, and a spatiotemporally aware gated recurrent unit, we extract users' long-term and short-term travel preferences. Combined with graph convolutional networks and deep learning models, we capture users' personalized travel patterns.
Providing accurate point-of-interest recommendations when historical data is insufficient improves the model's performance under sparse data, captures the transition rules between non-adjacent check-ins, and solves the problems of complexity and dynamism of user travel preferences.
Smart Images
Figure CN117271899B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the technical field of data processing, and particularly relates to a point of interest recommendation method based on space-time perception. BACKGROUND
[0002] The point of interest recommendation technology is a personalized recommendation task, which discovers potential patterns based on historical registration and other multi-modal information to predict the next set of interest points suitable for a specific user. In recent years, more and more mobile devices have brought about significant development of mobile cloud technology, including mobile storage, computing and artificial intelligence services based on location-based social networks. Many mobile cloud service providers allow users to share various forms of check-in data, including location (latitude, longitude), timestamp and rich semantic information, by taking advantage of mobile cloud technology. The massive check-in data generated in the mobile cloud preserves rich spatio-temporal context information, which is of great significance for understanding user travel preferences and predicting the next location. At the same time, the point of interest recommendation task is also of great significance for helping service providers analyze user behavior patterns and optimize personalized recommendation systems.
[0003] Early point of interest recommendation research mainly focused on content-based recommendation algorithms or collaborative filtering-based recommendation algorithms. The content-based recommendation algorithm extracts item features from the user's historical favorite items to represent the user's preferences with these features, and then calculates the similarity between items to recommend the most matching items to the user. The core concept of the collaborative filtering algorithm is to recommend the content of interest to the target user to the similar users of the target user. This method does not involve the relevance of item content and user description, so it has the characteristics of novelty to the recommended user. Then, due to the large number of users and items, the sparsity problem of existing data sets is quite serious, which reduces the recommendation effect of the model. At the same time, most collaborative filtering algorithms focus on the calculation of similar users, while ignoring how to distinguish the items to the maximum extent, thereby resulting in poor diversity of the recommendation results.
[0004] The existing technology applies deep learning to point of interest recommendation, but the point of interest recommendation method based on deep learning also has many challenges, for example:
[0005] 1. How to solve the cold start problem caused by the sparsity of historical data. Existing point of interest recommendation models usually need to learn the user's potential travel patterns based on the user's historical trajectory sequence, and cannot provide effective recommendations for new users or users with few historical trajectories.
[0006] 2. How to capture the nonlinear spatio-temporal dependence between non-adjacent check-ins. Most models only focus on the transition patterns between adjacent check-ins in the sequence, while ignoring the spatio-temporal dependence between non-adjacent check-ins, and cannot capture periodic features at different time scales.
[0007] 3. How to solve the complexity of user travel preferences. User check-in sequences have real-time and dynamic characteristics, and even different behaviors on different mobile cloud platforms. The interest points that users will visit can change dynamically over time, location, and user characteristics. In addition, even after visiting the same interest point, different users may choose different interest points next time. Therefore, it is still challenging to model user travel preferences from multiple scales and extract personalized travel preferences. SUMMARY
[0008] In view of the deficiencies of the prior art, the present application proposes a POI recommendation method based on spatiotemporal perception, which comprises:
[0009] S1: Obtain user check-in data, which includes user number, location number, POI category number and timestamp data;
[0010] S2: Input the user check-in data into the embedding module for processing to obtain the user trajectory embedding matrix and the spatiotemporal embedding matrix;
[0011] S3: Construct a trajectory flow graph according to the user check-in data and calculate the user's general behavior pattern according to the node features of the trajectory flow graph;
[0012] S4: Process the trajectory embedding matrix and the spatiotemporal embedding matrix of the user's historical trajectory sequence using the spatiotemporal perception attention module to obtain the user's long-term travel preference representation;
[0013] S5: Process the trajectory embedding matrix and the spatiotemporal embedding matrix of the user's current trajectory sequence using the spatiotemporal perception gated recurrent unit to obtain the user's short-term travel preference representation;
[0014] S6: Process the user's general behavior pattern, long-term travel preference representation and short-term travel preference representation using the interest point matching module to obtain the user's POI recommendation result; recommend the corresponding user with the POI according to the POI recommendation result.
[0015] Preferably, the process of obtaining the trajectory embedding matrix comprises: converting the user number, location number, POI category number and timestamp data corresponding to the interest point into embedding representation respectively; merging the four kinds of embedding representation as trajectory embedding representation; obtaining the trajectory embedding matrix according to the trajectory embedding representation of all interest points.
[0016] Preferably, the process of obtaining the time-space interval embedding matrix comprises: calculating the time interval and the location interval between all interest points of the user; converting the time interval and the location interval between the interest points into embedding representations and merging them to obtain a time-space interval embedding representation between the interest points; and obtaining the time-space interval embedding matrix of the user according to the time-space interval embedding representations between all the interest points.
[0017] Preferably, the process of obtaining the general behavior pattern of the user comprises: merging the location embedding representation corresponding to the interest point and the POI category embedding representation as the initial feature of the corresponding graph node; processing the initial feature of the graph node by using a stacked GCN network to obtain an output node feature; and obtaining the general behavior pattern of the user by averaging the output node feature of the last layer of the network.
[0018] Preferably, the process of obtaining the long-term travel preference representation of the user comprises: performing linear projection processing on the trajectory embedding matrix to obtain Q, K and V vectors of the trajectory embedding matrix, performing self-attention operation on the Q, K and V vectors and the time-space interval embedding matrix to obtain a global travel feature of the user; and processing the trajectory embedding matrix and the global travel feature of the user by using a feedforward neural network layer to obtain the long-term travel preference representation of the user.
[0019] Preferably, the formula for obtaining the long-term travel preference representation of the user is:
[0020] L=LayerNorm(E t +Attention(e z ))
[0021] L′=max(0,LW1+b1)
[0022] wherein L represents a residual connection result, e t represents the trajectory embedding matrix of the user u, Attention(E z ) represents the global travel feature of the user extracted by the attention layer, LayerNorm() represents a layer normalization operation, w1 represents a weight matrix of the space-time perception self-attention layer feedforward neural network, b1 represents a bias term of the space-time perception self-attention layer feedforward neural network, and L' represents the long-term travel preference of the user.
[0023] Preferably, the process of obtaining the short-term travel preference representation of the user comprises: introducing time-space interval information into the update gate and the reset gate of the gated recurrent unit to control the influence degree of the hidden feature of the historical state in the calculation; calculating the candidate hidden state and the hidden state of the current time step by using the reset gate and the update gate containing the time-space interval information, respectively; and obtaining the short-term travel preference representation of the user according to the hidden state of the current time step.
[0024] Further, the reset gate and the update gate containing the time-space interval information are represented as:
[0025] z t = σ(W z x t + U z h t-1 + O z m ij )
[0026] r t = σ(W r x t + U r h t-1 + O r m ij )
[0027] wherein, z t represents an update gate, r t represents a reset gate, x t represents an embedding representation of a user's check-in point at time t, h t-1 represents a hidden state at the last time, m ij represents a spatio-temporal separation embedding representation between the user's check-in at the i-th interest point and the check-in at the j-th interest point, W z , U z , O z respectively represent the first, second and third weight matrices of the update gate, W r , U e , O r respectively represent the first, second and third weight matrices of the reset gate, and σ represents an activation function.
[0028] Preferably, the process of obtaining the interest point recommendation result of the user comprises: fusing the general behavior pattern of the user, the long-term travel preference representation of the user and the short-term travel preference representation of the user to obtain a user travel feature; calculating a candidate POI embedding matrix and performing linear projection processing on the candidate POI embedding matrix and the user travel feature to obtain a Q vector of the candidate POI embedding matrix and K and V vectors of the user travel feature; calculating a spatio-temporal separation embedding matrix between the candidate POI and the user trajectory; performing attention operation on the Q, K and V vectors and the spatio-temporal separation embedding matrix between the candidate POI and the user trajectory to obtain a user travel attention feature; and calculating the interest point recommendation result of the user according to the user travel attention feature.
[0029] Further, the formula for obtaining the user travel feature is:
[0030] F = max(0, concat[(L' + S), r]W + b)
[0031] Wherein, F represents a user travel feature, concat represents splicing from a feature dimension, L' represents a user long-term travel preference, S represents a user short-term travel preference, r represents a general behavior mode, W represents a feature fusion layer feedforward neural network weight matrix, and b represents a feature fusion layer feedforward neural network bias term.
[0032] The application has the advantages that: the application uses a sequence prediction model that fuses space-time attention, a space-time recurrent neural network and a graph convolution to explore user personalized travel preferences and general behavior modes, well captures the transfer rules between non-adjacent check-ins and provides recommendation services in the case of insufficient historical data, converts user check-in sequences collected by a mobile network into a trajectory flow graph and learns global general features thereon, well handles the insufficient features of users with less check-in records, thereby improving the performance of the model in the case of sparse data, proposes a heterogeneous space-time perception module (a space-time perception attention module and a space-time perception gated recurrent unit) to extract long-term stable preferences and short-term dynamic preferences, the preferences capture the travel modes of users in different time periods and provide a multi-scale feature basis for personalized destination prediction, and the application can flexibly and accurately recommend interest points to users and has good application prospects. BRIEF DESCRIPTION OF DRAWINGS
[0033] Figure 1 FIG. 1 is a structure schematic diagram of an interest point recommendation prediction model based on space-time perception in the application;
[0034] Figure 2 FIG. 2 is a structure schematic diagram of a space-time perception attention module in the application;
[0035] Figure 3 FIG. 3 is a structure schematic diagram of a space-time perception gated recurrent unit in the application. DETAILED DESCRIPTION
[0036] The technical solutions in the embodiments of the application will be apparently and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.
[0037] The application proposes an interest point recommendation method based on space-time perception, and the method comprises the following contents:
[0038] The application designs an interest point recommendation prediction model based on space-time perception, inputs user check-in data into the trained model, and can output interest point recommendation results of the user; as shown in FIG. 1, the process of interest point recommendation of the model comprises: Figure 1
[0039] S1: Obtain user check-in data, the user check-in data including user number, location number, POI category number and timestamp data.
[0040] The user check-in data can be obtained from an Internet platform, and the user check-in data is represented as a check-in tuple c = (u, p, s, t), u, p, s, t respectively represent the user number, the location number, the POI category number and the timestamp data. The location number is the location number of the interest point where the user is located, the POI category number is the number of the POI category where the user is located, and the timestamp data is the relative time number of the user arriving at the interest point (the relative time number is calculated according to the offset of the current time relative to the zero point of Monday of the current week). The user check-in data is divided into historical trajectory data and current trajectory data.
[0041] S2: Input the user check-in data into an embedding module for processing to obtain a trajectory embedding matrix and a time-space interval embedding matrix of the user.
[0042] The embedding layer converts the user number, the location number, the POI category number and the timestamp data corresponding to the interest point into embedding representations e u , e p , e s , and e t respectively; the four kinds of embedding representations are combined as a trajectory embedding representation e; the trajectory embedding matrix is obtained according to the trajectory embedding representations of all interest points, that is, for a user check-in sequence represents the nth check-in tuple of the user, and the trajectory embedding matrix obtained after the embedding layer processing is represents the trajectory embedding representation of the nth interest point of the user.
[0043] The time interval and the location interval between all interest points of the user are calculated, and the calculation formula is:
[0044]
[0045]
[0046] wherein, represents the normalized time interval of the user checking in at the interest point i and checking in at the interest point j, represents the normalized location interval of the user checking in at the interest point i and checking in at the interest point j, t i and t j respectively represent the check-in time of the user at the interest point i and the interest point j, s i and s j respectively represent the location of the user checking in at the interest point i and checking in at the interest point j; t max -t minMaximal time difference value of user interest point check-in sequence, s max -s min Maximal spatial distance of user interest point check-in sequence.
[0047] The time interval and location interval between interest points are input into the embedding layer to convert into embedding representations and merge to obtain a time-space interval embedding representation m between interest points ij A time-space interval embedding matrix of the user is obtained according to the time-space interval embedding representations between all interest points.
[0048] The historical trajectory data and the current trajectory data are processed by the above method respectively, and the trajectory embedding matrix and the time-space interval embedding matrix of the user historical trajectory sequence, and the trajectory embedding matrix and the time-space interval embedding matrix of the user current trajectory sequence can be obtained.
[0049] S3: Constructing a trajectory flow graph according to the user check-in data and calculating a general behavior pattern of the user according to the node features of the trajectory flow graph.
[0050] The general pattern learning module is used to construct a trajectory flow graph, and all interest points appearing in the historical trajectory of the same batch of users are added to the trajectory flow graph as nodes, and a transition from an interest point v i to v j is regarded as a directed edge in the graph.
[0051] The location embedding representation e p and the category embedding representation e s corresponding to the interest points are combined as the initial features of the corresponding graph nodes.
[0052] The information of adjacent nodes is aggregated by a graph convolution network, specifically: a stacked GCN network (graph convolutional neural network) is used to process the initial features of the graph nodes to obtain output node features:
[0053]
[0054] Wherein, H (l+1) represents the node features of the l+1th layer of the network, W (l) represents the weight matrix of the lth layer, and σ represents the activation function, represents the degree matrix of , and A represents the adjacency matrix with self-loop added.
[0055] By stacking multiple GCN modules, the performance of the model can be enhanced, and the average value of the output node features in the node dimension of the last layer of the network is obtained to obtain a general behavior pattern at the graph level:
[0056]
[0057] wherein r represents a general behavior pattern, x n represents the output node feature of the nth node, and N represents the number of nodes of the graph.
[0058] S4: processing the trajectory embedding matrix of the user historical trajectory sequence and the space-time interval embedding matrix by using the space-time perception attention module to obtain a user long-term travel preference representation.
[0059] As shown in Figure 2 , the long-term stable travel preference in the user historical trajectory is extracted by using the space-time perception attention module, and specifically:
[0060] The trajectory embedding matrix of the user historical trajectory sequence is linearly projected to obtain Q, K and V vectors of the trajectory embedding matrix:
[0061] Q = E u W Q , K = E u W K , and V = E u W V
[0062] wherein Q, K and V respectively represent a query matrix, a key matrix and a value matrix of the trajectory embedding matrix, W Q , W K and W V respectively represent weight matrices of the query matrix, the key matrix and the value matrix.
[0063] Self-attention operation is performed on the Q, K and V vectors and the space-time interval embedding matrix to obtain a user global travel feature:
[0064]
[0065] wherein M represents the space-time interval embedding matrix of the user, and Attention(E u ) represents the user global travel feature.
[0066] The output of the self-attention layer, i.e., the user global travel feature, is processed by using an activation function and normalization to accelerate convergence. In order to introduce nonlinearity into the network, the trajectory embedding matrix and the user global travel feature are finally processed by using a feedforward neural network layer with ReLU as the activation function to obtain the user long-term travel preference representation, which is represented as:
[0067] L = LayerNorm(E u + Attention(E u ))
[0068] L′=max(0,LW1+b1)
[0069] Where L represents the residual connection result, E u Let E represent the trajectory embedding matrix of user u, and Attention(E) u ) represents the global travel features of the user extracted by the attention layer, LayerNorm() represents the layer normalization operation, W1 represents the weight matrix of the feedforward neural network of the spatiotemporal awareness attention module, b1 represents the bias term of the feedforward neural network of the spatiotemporal awareness attention module, and L′ represents the user's long-term travel preferences.
[0070] S5: The spatiotemporal-aware gated loop unit is used to process the trajectory embedding matrix and spatiotemporal interval embedding matrix of the user's current trajectory sequence to obtain the user's short-term travel preference representation.
[0071] like Figure 3 As shown, a spatiotemporally aware gated recurrent unit (GRU) is used to extract short-term dynamic travel preferences from the user's current trajectory. First, spatiotemporal interval information is introduced into the GRU's update and reset gates; that is, the trajectory embedding matrix and spatiotemporal interval embedding matrix of the user's current trajectory sequence are processed using update and reset gates to control the influence of hidden features from historical states in the calculation. This allows for better capture of the user's short-term dynamic travel preferences. The specific formula is as follows:
[0072] z t =σ(W z x t +U z h t-1 +O z m ih )
[0073] r t =σ(W r x t +U r h t-1 +O r m ij )
[0074] Among them, z t Indicates the update gate, r t Indicates resetting the door, x t h represents the embedded representation of a user's check-in point at time t. t-1 m represents the hidden state at the previous time step. ij W represents the spatiotemporal embedding representation between a user's check-in at point i and check-in at point j. z U z O z W represents the first, second, and third weight matrices of the update gate, which are the weight matrices of the features corresponding to the update gate.r r r respectively represent the first, second and third weight matrix of the reset gate, i.e., the weight matrix corresponding to the feature of the reset gate, and σ represents an activation function.
[0075] The candidate hidden state of the current time step is calculated using the reset gate and the update gate containing the time-space separation information and the hidden state h t The user short-term travel preference representation S is formed according to the hidden state of the current time step:
[0076]
[0077]
[0078] wherein W' represents the weight matrix of the input at the current time, and U' represents the hidden state weight matrix.
[0079] S6: The interest point matching module is used to process the user general behavior pattern, the user long-term travel preference representation and the user short-term travel preference representation, to obtain the interest point recommendation result of the user; and the interest point is recommended to the corresponding user according to the interest point recommendation result.
[0080] The user general behavior pattern, the user long-term travel preference representation and the user short-term travel preference representation are fused to obtain the user travel feature:
[0081] F = max(0, concat[(L' + S), r] W + b)
[0082] wherein F represents the user travel feature, concat represents concatenation, L' represents the user long-term travel preference representation, S represents the user short-term travel preference representation, r represents the general behavior pattern, W represents the weight matrix of the feature fusion layer feedforward neural network, and b represents the bias term of the feature fusion layer feedforward neural network.
[0083] The candidate POI is obtained, and the embedding layer is used to calculate the candidate POI embedding matrix, and the calculation process is similar to the process of obtaining the trajectory embedding matrix, which is not repeated here; the candidate POI embedding matrix and the user travel feature are linearly projected to obtain the Q vector of the candidate POI embedding matrix and the K and V vectors of the user travel feature:
[0084] Q = DW Q K = FW K V = FW V
[0085] wherein D represents the candidate POI embedding matrix.
[0086] Calculate the time-space interval embedding matrix N between the candidate POI and the user trajectory, which is similar to the process of calculating the time-space interval embedding matrix of the user.
[0087] Perform attention operation on the Q, K, V vectors and the time-space interval embedding matrix N to obtain the user travel attention feature:
[0088]
[0089] Calculate the POI recommendation result of the user according to the user travel attention feature:
[0090] Output=max(0,Match(Q,K,V,N)W m +b m )
[0091] Wherein, Output represents the probability distribution vector of the candidate POI as the next access location of the user, Match(Q,K,V,N) represents the user travel feature, W m represents the output layer feedforward neural network weight matrix, and b m represents the output layer feedforward neural network bias term.
[0092] In the process of training the model, cross entropy is used as the loss function of the model, and negative samples are randomly sampled at each calculation to solve the problem of unbalanced number of positive and negative samples. The formula for calculating the loss function is as follows:
[0093]
[0094] Wherein, Loss represents the total loss of the model, S u represents the check-in sequence of user u, S represents the POI check-in sequence of all users, n represents the number of iterations, σ represents the activation function, y i represents the predicted result given by the model, o i represents the true label of the i-th iteration, and K represents the number of negative samples randomly sampled.
[0095] Preferably, the present application uses hit rate (HR) as the evaluation index of the result:
[0096]
[0097] Wherein, N represents the number of samples; hits(i) represents whether the value visited by the i-th user is in the Topk options of the prediction list, and 1 if yes, and 0 if no.
[0098] In summary, the application solves the cold start problem caused by sparse historical data by constructing a trajectory flow graph according to the trajectories of the same batch of users, then using a message passing mechanism to aggregate and update node features, and finally obtaining the general behavior pattern representation of the current user set through global average pooling at the graph level; the application designs a spatio-temporal perception structure, constructs a spatio-temporal interval embedding matrix according to the spatio-temporal interval difference between each user trajectory point, inputs the spatio-temporal interval difference into the embedding layer to combine and convert it into a hidden vector form, then inputs it into two kinds of heterogeneous feature extraction modules respectively and fuses it with the hidden state to enhance the model's perception ability of non-adjacent check-in spatio-temporal information, so as to capture the non-linear spatio-temporal dependence relationship between non-adjacent check-ins; the application extracts the user's long-term stable preference and short-term dynamic preference at the same time, divides the check-in sequence into historical trajectory and current trajectory, and uses the spatio-temporal perception structure to capture the user's travel preference at different time scales through the self-attention mechanism and the gated recurrent unit respectively, so as to solve the complexity of the user's travel preference; the application can realize flexible and accurate recommendation of interest points to users, and has good application prospect.
[0099] The above examples further illustrate the purpose, technical solutions and advantages of the application. It should be understood that the above examples are only preferred embodiments of the application and are not intended to limit the application. Any modifications, equivalent replacements, improvements, etc. made to the application within the spirit and principle of the application shall be included in the protection scope of the application.
Claims
1. A spatiotemporal awareness-based point-of-interest recommendation method, characterized in that, include: S1: Obtain user check-in data, which includes user ID, location ID, POI category ID, and timestamp data; S2: Input the user check-in data into the embedding module for processing to obtain the user's trajectory embedding matrix and spatiotemporal interval embedding matrix; S3: Construct a trajectory flow graph based on user check-in data and calculate the user's general behavior pattern based on the node characteristics of the trajectory flow graph; The process of obtaining the user's general behavior pattern includes: merging the location embedding representation corresponding to the point of interest (POI) with the POI category embedding representation as the initial features of the corresponding graph node; processing the initial features of the graph node using a stacked GCN network to obtain the output node features; and averaging the output node features of the last layer of the network to obtain the user's general behavior pattern. S4: A spatiotemporal awareness attention module is used to process the trajectory embedding matrix and spatiotemporal interval embedding matrix of the user's historical trajectory sequence to obtain the user's long-term travel preference representation. The process of obtaining the user's long-term travel preference representation includes: performing linear projection processing on the trajectory embedding matrix to obtain the Q, K, and V vectors of the trajectory embedding matrix; performing self-attention operation on the Q, K, and V vectors and the spatiotemporal interval embedding matrix to obtain the user's global travel features; and using a feedforward neural network layer to process the trajectory embedding matrix and the user's global travel features to obtain the user's long-term travel preference representation. S5: A spatiotemporally aware gated recurrent unit is used to process the trajectory embedding matrix and spatiotemporal interval embedding matrix of the user's current trajectory sequence to obtain the user's short-term travel preference representation. The process of obtaining the user's short-term travel preference representation includes: introducing spatiotemporal interval information into the update gate and reset gate of the gated recurrent unit to control the influence of the hidden features of historical states in the calculation; using the reset gate and update gate containing spatiotemporal interval information to calculate the candidate hidden state and hidden state of the current time step respectively, and obtaining the user's short-term travel preference representation based on the hidden state of the current time step. S6: The interest point matching module processes the user's general behavior pattern, long-term travel preference representation, and short-term travel preference representation to obtain the user's interest point recommendation results; and recommends interest points to the corresponding users based on the interest point recommendation results.
2. The method for recommending points of interest based on spatiotemporal awareness according to claim 1, characterized in that, The process of obtaining the trajectory embedding matrix includes: converting the user ID, location ID, POI category ID, and timestamp data corresponding to the points of interest into embedding representations respectively; merging the four embedding representations into the trajectory embedding representation; and obtaining the trajectory embedding matrix based on the trajectory embedding representations of all points of interest.
3. The interest point recommendation method based on spatiotemporal awareness according to claim 1, characterized in that, The process of obtaining the spatiotemporal embedding matrix includes: calculating the time interval and location interval between all user interest points; converting the time interval and location interval between interest points into embedding representations and merging them to obtain the spatiotemporal embedding representation between interest points; and obtaining the user's spatiotemporal embedding matrix based on the spatiotemporal embedding representation between all interest points.
4. The interest point recommendation method based on spatiotemporal awareness according to claim 1, characterized in that, The formula for obtaining a user's long-term travel preferences is as follows: L=LayerNorm(E u +Attention(E u )) L′=max(0,LW1+b1) Where L represents the residual connection result, E u Let E represent the trajectory embedding matrix of user u, and Attention(E) u ) represents the global travel features of the user extracted by the attention layer, LayerNorm() represents the layer normalization operation, W1 represents the weight matrix of the spatiotemporal awareness self-attention layer feedforward neural network, b1 represents the bias term of the spatiotemporal awareness self-attention layer feedforward neural network, and L′ represents the user's long-term travel preferences.
5. The interest point recommendation method based on spatiotemporal awareness according to claim 1, characterized in that, The reset gate and update gate, which include spatiotemporal interval information, are represented as follows: z t =σ(W z x t +U z h t-1 +O z m ij ) r t =σ(W r x t +U r h t-1 +O r m ij ) Among them, z t Indicates the update gate, r t Indicates resetting the door, x t h represents the embedded representation of a user's check-in point at time t. t-1 m represents the hidden state at the previous time step. ij W represents the spatiotemporal embedding representation between a user's check-in at point i and check-in at point j. z U z O z W represents the first, second, and third weight matrices of the update gate, respectively. r U r O r Let represent the first, second, and third weight matrices of the reset gate, respectively, and σ represent the activation function.
6. The interest point recommendation method based on spatiotemporal awareness according to claim 1, characterized in that, The process of obtaining user point of interest (POI) recommendation results includes: fusing user's general behavioral patterns, long-term travel preference representations, and short-term travel preference representations to obtain user travel features; calculating the candidate POI embedding matrix and performing linear projection processing on the candidate POI embedding matrix and user travel features to obtain the Q vector of the candidate POI embedding matrix and the K and V vectors of the user travel features; calculating the spatiotemporal interval embedding matrix between the candidate POI and the user trajectory; performing attention operations on the Q, K, and V vectors and the spatiotemporal interval embedding matrix between the candidate POI and the user trajectory to obtain user travel attention features; and calculating the user's point of interest recommendation results based on the user travel attention features.
7. The method for recommending points of interest based on spatiotemporal awareness according to claim 6, characterized in that, The formula for obtaining user travel characteristics is: F=max(0,concat[(L′+S),r]W+b) Where F represents user travel characteristics, concat represents concatenation from the feature dimension, L′ represents user long-term travel preferences, S represents user short-term travel preferences, r represents general behavior patterns, W represents the weight matrix of the feedforward neural network of the feature fusion layer, and b represents the bias term of the feedforward neural network of the feature fusion layer.