A POI recommendation method fusing crowd moving trajectory and space-time-category
By integrating crowd movement trajectories and spatiotemporal-category POI recommendation methods, and utilizing graph convolutional neural networks and self-attention mechanisms, the problem of ineffective modeling of spatiotemporal features in existing technologies is solved, achieving more efficient POI recommendation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHONGBEI UNIV
- Filing Date
- 2024-03-12
- Publication Date
- 2026-07-24
AI Technical Summary
Existing POI recommendation methods fail to effectively model the spatiotemporal features of user check-in sequences, ignoring contextual semantic information and user social attributes, resulting in insufficient recommendation accuracy.
We employ a POI recommendation method that integrates crowd movement trajectories and spatiotemporal-category information. We learn regional trajectory map features through graph convolutional neural networks, combine a self-attention mechanism to perform multi-dimensional feature fusion and capture dynamic user preferences, use a Transformer encoder to extract user trajectory sequence features, and use a multilayer perceptron to predict the next POI, the time period of visit, and the category.
It improves the accuracy and precision of POI recommendations, especially in the case of new users' cold start, and can effectively utilize users' common trajectories and spatiotemporal features to improve the accuracy of recommendations and reduce model computation time.
Smart Images
Figure CN118171000B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of big data analysis technology, specifically relating to a POI recommendation method that integrates population movement trajectories and spatiotemporal-category data. Background Technology
[0002] With the rapid development of the internet and the continuous innovation of location technology, people are increasingly sharing their current life status through location, text, and images, generating a wealth of internet data while causing information overload. Against this backdrop, POI recommendation, as a hot research topic in location-based social networks, helps users accurately find locations that match their interests from massive amounts of information, and has attracted widespread attention.
[0003] Currently, the exploration and application of various deep learning models have greatly improved the performance of POI recommendation models. Sun et al. proposed a long-short-term preference modeling recommendation model, LSTPM, which combines nonlocal networks and geographically extended RNNs to combine long-short-term preferences for POI recommendation. Fang et al. proposed a recommendation method based on user relationships and preference information, URPI-GRU, which uses a GRU model and k-nearest neighbor sequences to learn users' current and long-term preferences, respectively, and finally calculates a total score to recommend POIs. Liu Shuyue et al. proposed a POI sequence recommendation model, SA-TDS-PRec, which integrates spatiotemporal and semantic information under self-attention and captures users' dynamic preferences using a self-attention mechanism. Zheng et al. proposed an attention-based dynamic preference model, ADPM, which introduces a feature-gated fusion module to achieve adaptive fusion of long-short-term preferences.
[0004] Existing research has yielded many results and progress, but there are still problems that require further research and solutions: Traditional POI recommendation methods do not consider the spatiotemporal features of user check-in sequences and cannot effectively model the sequences. POI recommendation methods based on RNNs and their variants ignore contextual semantic information and cannot effectively capture users' long-term preferences. Attention-based methods ignore users' social attributes and do not fully utilize the check-in information of other users to help improve the accuracy of POI recommendations. Summary of the Invention
[0005] To address the issues of ineffective modeling of shared user trajectories and insufficient mining of spatiotemporal features, this invention provides a POI recommendation method that integrates crowd movement trajectories and spatiotemporal-category features.
[0006] To achieve the above objectives, the present invention employs the following technical solutions:
[0007] In a first aspect, the present invention provides a POI recommendation method that integrates crowd movement trajectories and spatiotemporal-category information, comprising the following steps:
[0008] Step 1: Obtain user check-in sequence data and perform data preprocessing and regional trajectory map construction;
[0009] Step 2: Construct a POI recommendation model that integrates crowd movement trajectories and spatiotemporal-category data;
[0010] Step 3: Input the user check-in sequence data and regional trajectory map obtained in Step 1 into the model obtained in Step 2 to make the next POI recommendation.
[0011] Further, the region trajectory graph mentioned in step 1 is defined as: a directed graph consisting of a node set V, an edge set E, node attributes S, and weights W; the node set V consists of POIs historically visited by the user; the edge set E is a set of directed edges, where the user has visited two POIs consecutively within a certain period of time: and Then in and Add a directed edge e between them; the directed edge e is the user's... arrive The sub-trajectories have a weight W equal to the number of times the sub-trajectory appears in the region trajectory graph; the node attributes S contain the spatial coordinates, category, and access frequency of each POI.
[0012] Furthermore, the POI recommendation model in step 2 includes a regional trajectory map learning module, a feature learning fusion layer, and an encoder-decoder structure. The regional trajectory map learning module performs feature aggregation and noise filtering on the regional trajectory map through a graph convolutional neural network with fused gated recurrent units, learning the features of each POI. The feature learning fusion layer performs preliminary feature extraction and fusion on information from different dimensions in the user check-in sequence data through a POI-User fusion module, a time-space fusion module, and a time-category fusion module. The encoder extracts features from the user trajectory sequence. The decoder predicts the next POI, the access time period, and the POI category through three multilayer perceptrons.
[0013] Furthermore, the region trajectory map learning module learns the features of each POI in the region trajectory map through a graph convolutional neural network that integrates gated recurrent units. The specific learning process includes:
[0014] Step 2.1: Calculate the normalized Laplacian matrix of the adjacency matrix of the region trajectory map. ;
[0015] Step 2.2: Feature extraction using a single-layer graph convolutional neural network;
[0016] Step 2.3: Update the hidden state of each node using the gated loop unit;
[0017] Step 2.4: After repeating steps 2.2 and 2.3 multiple times, combine the obtained feature vector with the feature vector extracted in the first graph convolution. The features of each POI are obtained by concatenating and merging the data. The calculation process is expressed by the following formula:
[0018]
[0019] In the formula, This represents a single-layer graph convolution operation. This represents a gated loop unit with time step T. and These represent the weights and biases, respectively. Indicates splicing, For Laplace matrix, These are the eigenvectors.
[0020] Furthermore, the feature learning fusion layer learns the POI-User feature representation for each user's check-in activity through the POI-User fusion module; learns the time-space fusion feature representation for the user's check-in activity using the time-space fusion module; and learns the time-category fusion feature representation for the user's check-in activity using the time-category fusion module. The specific learning process includes:
[0021] Step 2.5: Embed user u's personal information into a user feature vector. Representation; the feature vector of each POI in the user u check-in sequence. With user feature vector The data is concatenated and then nonlinearly transformed in a fully connected layer to obtain the POI-User fusion feature representation for each user's check-in activity. The calculation process is expressed by the following formula:
[0022]
[0023] In the formula, and These represent the weights and biases, respectively. Indicates splicing, LeakyReLU represents the non-linear activation function.
[0024] Step 2.6: Obtain timestamp feature vectors using the encoding strategy of the neural network model CTLE (Context and Time Aware LocationEmbeddings). The coordinates of each POI are input into the fully connected layer to obtain spatial feature vectors. The timestamp feature vector and spatial feature vector are concatenated and input into a fully connected layer for feature fusion, constructing a time-space fusion feature representation of user check-in activities. The calculation process is expressed by the following formula:
[0025]
[0026] In the formula, Represents a learnable weight vector. Indicates deviation, LeakyReLU represents a non-linear activation function.
[0027] Step 2.7: Encode the absolute access time using the time2vector neural network model; embed each POI category into a feature vector. Representation; class feature vector and time feature vector The data is then concatenated and fed into a fully connected layer for feature fusion to construct a time-category fusion feature representation of user check-in activities. The calculation process is expressed by the following formula:
[0028]
[0029] In the formula, This represents a learnable weight vector. Indicates deviation, This represents the nonlinear activation function LeakyReLU.
[0030] Furthermore, the user trajectory sequence is defined as: composed of the user's continuous check-in records within 48 hours in chronological order; based on the POI-User fusion feature representation of the user's check-in activities. Temporal-spatial fusion feature representation and time-category fusion feature representation Then, the embedding representation of the user check-in sequence q is divided into two parts: and ; trajectory sequence of each user Both are embedded by two user check-in sequences. and express.
[0031] Furthermore, the encoder utilizes a Transformer to extract user trajectory sequence features, specifically including:
[0032] Step 3.1: Embedding of a given user trajectory sequence and ,Will and These are passed as input tensors to the Transformer encoder.
[0033] Step 3.2: For the m-th layer of the encoder, the input... and First, a transformation is performed using a multi-head self-attention layer, and the calculation process is expressed by the following formula:
[0034]
[0035]
[0036]
[0037] In the formula, , , These represent the query weight matrix, key weight matrix, and value weight matrix, respectively. This indicates the correlation between the i-th and j-th check-ins. This indicates the influence weight of other check-in activities on the current check-in, and uses the softmax function to ensure that the sum of attention weights is 1;
[0038] Step 3.3: Stack the outputs from different attention heads and use layer normalization and residual connection techniques. The calculation process is shown in the following formula:
[0039]
[0040] Step 3.4: After the multi-head self-attention layer, a feedforward neural network (FNN) and the ReLU activation function are used to give the model nonlinearity. The calculation process is shown in the following formula:
[0041]
[0042]
[0043] in , Indicates weight, , Indicates deviation, This represents the final output of the encoder at layer m.
[0044] Step 3.5: After stacking multiple multi-head self-attention blocks, the output feature Z of the Transformer encoder is transformed into temporal-space output features. Time-Category Output Features The weighted sum is defined as follows:
[0045]
[0046] in This represents the activation function tanh. and This represents two aggregation hyperparameters.
[0047] Furthermore, the decoder uses three multilayer perceptrons to predict the next POI: Access time period and POI categories The definition is as follows:
[0048]
[0049]
[0050]
[0051] In the formula, Z represents the output characteristic of the Transformer encoder. , , Indicates weight, , , Indicates deviation.
[0052] Furthermore, the overall loss function of the POI recommendation model is defined as follows:
[0053]
[0054] In the formula, This represents the overall loss function of the network model. and Let these represent the loss functions for predicting the next POI and POI category, respectively, using the common cross-entropy loss. The mean squared error (MSE) was used for performance evaluation of access time prediction.
[0055] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.
[0056] Thirdly, the present invention provides a computer device including a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, and the processor executes the computer program to implement the steps of the method described in the first aspect.
[0057] Compared with the prior art, the present invention has the following advantages:
[0058] The method of this invention is applicable to recommending the next POI that users are interested in. It models common trajectories by using a proposed gated graph convolutional neural network, and performs multi-dimensional feature fusion of spatiotemporal and category information in the user check-in sequence. It also uses a self-attention mechanism to capture the user's dynamic preferences. Attached Figure Description
[0059] Figure 1 This is a schematic diagram of the POI recommendation method that integrates crowd movement trajectories and spatiotemporal-category information according to the present invention;
[0060] Figure 2 This is a schematic diagram of the regional trajectory map of the present invention;
[0061] Figure 3 This is a schematic diagram of the gated graph convolutional neural network module of the present invention;
[0062] Figure 4 This is a schematic diagram illustrating the cold start test for new users according to the present invention. Detailed Implementation
[0063] To facilitate understanding of the present invention, a more comprehensive description will be given below. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the present invention.
[0064] A method for recommending points of interest (POIs) that integrates crowd movement trajectories and spatiotemporal-category information, such as... Figure 1 As shown, it includes the following steps:
[0065] Step 1: Obtain user check-in sequence data and perform data preprocessing and regional trajectory map construction;
[0066] Specifically, the FourSquare dataset is used as the dataset for the POI recommendation task that integrates crowd movement trajectories and spatiotemporal-category data in this invention. The FourSquare-NYC dataset was collected in New York from April 2012 to February 2013, and the FourSquare-TKY dataset was collected in Tokyo during the same period. Each check-in record contains information such as user ID, POI ID, POI category, GPS coordinates, and timestamp. The first 80% is used as the training set, and the remaining 20% is used as the test set. POIs and users with fewer than 10 check-in records are deleted; the check-in sequence data of each user is divided into trajectories with 24-hour intervals. If a trajectory contains only one check-in record, the trajectory data is deleted.
[0067] Step 2: Construct a POI recommendation model that integrates crowd movement trajectories and spatiotemporal-category information (GGCN-STC).
[0068] The POI recommendation model in step 2 includes a regional trajectory map learning module, a feature learning fusion layer, and an encoder-decoder structure. The regional trajectory map learning module performs feature aggregation and noise filtering on the regional trajectory map through a graph convolutional neural network with gated recurrent units, learning the features of each POI. The feature learning fusion layer performs preliminary feature extraction and fusion on information from different dimensions in the user check-in sequence data through a POI-User fusion module, a time-space fusion module, and a time-category fusion module. The encoder extracts features from the user trajectory sequence. The decoder predicts the next POI, the access time period, and the POI category through three multilayer perceptrons.
[0069] The region trajectory graph is defined as: a directed graph consisting of a node set V, an edge set E, node attributes S, and weights W; the node set V consists of POIs visited historically by the user; the edge set E is a set of directed edges, where edges are formed when a user visits two POIs consecutively within a certain period of time. and Then in and Add a directed edge e between them; the weight W of each edge is equal to the number of times the sub-trajectory appears in the region trajectory map; the node attributes S contain information such as the spatial coordinates, category and access frequency of each POI.
[0070] The region trajectory map learning module learns the features of each POI in the region trajectory map through a graph convolutional neural network with gated recurrent units. The specific learning process includes:
[0071] Step 2.1: Calculate the normalized Laplacian matrix of the adjacency matrix of the region trajectory map. ;
[0072] Step 2.2: Feature extraction using a single-layer graph convolutional neural network;
[0073] Step 2.3: Update the hidden state of each node using the gated loop unit;
[0074] Step 2.4: After repeating steps 2.2 and 2.3 multiple times, combine the obtained feature vector with the feature vector extracted in the first graph convolution. The features of each POI are obtained by concatenating and merging the data. The calculation process is expressed by the following formula:
[0075]
[0076] In the formula, This represents a single-layer graph convolution operation. This represents a gated loop unit with time step T. and These represent the weights and biases, respectively. Indicates splicing, For Laplace matrix, These are the eigenvectors.
[0077] Furthermore, the feature learning fusion layer learns the POI-User feature representation for each user's check-in activity through the POI-User fusion module; learns the time-space fusion feature representation for the user's check-in activity using the time-space fusion module; and learns the time-category fusion feature representation for the user's check-in activity using the time-category fusion module. The specific learning process includes:
[0078] Step 2.5: Embed user u's personal information into a user feature vector. Representation; the feature vector of each POI in the user u check-in sequence. With user feature vector The data is concatenated and then nonlinearly transformed in a fully connected layer to obtain the POI-User fusion feature representation for each user's check-in activity. The calculation process is expressed by the following formula:
[0079]
[0080] In the formula, and These represent the weights and biases, respectively. Indicates splicing, LeakyReLU represents the non-linear activation function.
[0081] Step 2.6: Use the encoding strategy of the neural network model CTLE (Context and Time Aware LocationEmbeddings) to obtain the timestamp feature vector. The coordinates of each POI are input into the fully connected layer to obtain spatial feature vectors. The timestamp feature vector and spatial feature vector are concatenated and input into a fully connected layer for feature fusion, constructing a time-space fusion feature representation of user check-in activities. The calculation process is expressed by the following formula:
[0082]
[0083] In the formula, Represents a learnable weight vector. Indicates deviation, LeakyReLU represents a non-linear activation function.
[0084] Step 2.7: Encode the absolute access time using the time2vector neural network model; embed each POI category into a feature vector. Representation; class feature vector and time feature vector The data is then concatenated and fed into a fully connected layer for feature fusion to construct a time-category fusion feature representation of user check-in activities. The calculation process is expressed by the following formula:
[0085]
[0086] In the formula, This represents a learnable weight vector. Indicates deviation, This represents the nonlinear activation function LeakyReLU.
[0087] The user trajectory sequence is defined as: composed of a user's continuous check-in records within 48 hours in chronological order; based on the POI-User fusion feature representation of the user's check-in activities. Temporal-spatial fusion feature representation and time-category fusion feature representation Then, the embedded representation of the user check-in sequence data q is divided into two parts: and ; trajectory sequence of each user Both are embedded from two user check-in sequence data. and express.
[0088] The encoder uses a Transformer to extract user trajectory sequence features. The specific extraction process includes:
[0089] Step 3.1: Embedding of a given user trajectory sequence and ,Will and These are passed as input tensors to the Transformer encoder.
[0090] Step 3.2: For the m-th layer of the encoder, the input... and First, a transformation is performed using a multi-head self-attention layer, and the calculation process is expressed by the following formula:
[0091]
[0092]
[0093]
[0094] In the formula, , , These represent the query weight matrix, key weight matrix, and value weight matrix, respectively. This indicates the correlation between the i-th and j-th check-ins. This indicates the influence weight of other check-in activities on the current check-in, and uses the softmax function to ensure that the sum of attention weights is 1;
[0095] Step 3.3: Stack the outputs from different attention heads and use layer normalization and residual connection techniques. The calculation process is shown in the following formula:
[0096]
[0097] Step 3.4: After the multi-head self-attention layer, a feedforward neural network (FNN) and the ReLU activation function are used to give the model nonlinearity. The calculation process is shown in the following formula:
[0098]
[0099]
[0100] in , Indicates weight, , Indicates deviation, This represents the final output of the encoder at layer m.
[0101] Step 3.5: After stacking multiple multi-head self-attention blocks, the output feature Z of the Transformer encoder is transformed into temporal-space output features. Time-Category Output Features The weighted sum is defined as follows:
[0102]
[0103] in This represents the activation function tanh. and This represents two aggregation hyperparameters.
[0104] The decoder uses three multilayer perceptrons to predict the next POI: Access time period and POI categories The definition is as follows:
[0105]
[0106]
[0107]
[0108] In the formula, Z represents the output characteristic of the Transformer encoder. , , Indicates weight, , , Indicates deviation.
[0109] The overall loss function of the POI recommendation model is defined as follows:
[0110]
[0111] In the formula, This represents the overall loss function of the network model. and Let these represent the loss functions for predicting the next POI and POI category, respectively, using the common cross-entropy loss. The mean squared error (MSE) was used for performance evaluation of access time prediction.
[0112] Step 3: Input the user check-in data and regional trajectory map obtained in Step 1 into the model obtained in Step 2 to make the next POI recommendation;
[0113] Specifically, the network was trained using the Adam optimizer with an initial learning rate of 0.0001, a batch size of 16, and for 150 epochs, with hyperparameters aggregated. The value is 0.7. The value is 0.3, obtained through multiple tests. See Tables 1 and 2 for details.
[0114] Accuracy (Acc@k) and mean reciprocal rank (MRR) were used as metrics on the FourSquare dataset. Tables 1 and 2 compare the metrics of the proposed method with other methods for POI recommendation on the NYC and TKY datasets. As shown in Tables 1 and 2, the proposed method outperforms other methods in both accuracy (Acc@k) and mean reciprocal rank (MRR). Compared to GETNext, the proposed method improves Acc@1 by 1.96%, Acc@5 by 2.3%, Acc@10 by 2.21%, Acc@20 by 1.9%, and MRR by 2.08% on the NYC dataset; on the TKY dataset, Acc@1 improves by 0.65%, Acc@5 by 1.5%, Acc@10 by 1.6%, Acc@20 by 3.7%, and MRR by 0.99%.
[0115] Table 1. Overall performance comparison of different methods on the NYC dataset.
[0116]
[0117] Table 2. Overall performance comparison of different methods on the TKY dataset
[0118]
[0119] The effectiveness of the POI recommendation network GGCN-STC of this invention was verified by changing the length n of the input user trajectory. Figure 4 (a) To show the effect of the input user trajectory length n on Acc@10 in the NYC dataset. Figure 4 (b) The impact of the input user trajectory length n on Acc@10 in the TKY dataset. To test the model's performance in handling the cold start problem for new users, the value of n was set from 2 to 12 to simulate different sparsity levels of user check-in data, while keeping other parameters constant. The trend in the figure shows that when the input user trajectory length n=2, the GGCN-STC model significantly outperforms the baseline model on both the NYC and TKY datasets, and the recommendation performance of the model continues to improve as the input trajectory length increases. This indicates that the GGCN-STC model can effectively learn and utilize common user behavior patterns to provide more accurate recommendations for target users, thereby alleviating the cold start problem for new users.
[0120] This invention relates to a POI recommendation method that integrates crowd movement trajectories and spatiotemporal-category data. This method improves upon the GETNext model (GETNext: Trajectory Flow Map Enhanced Transformer for Next POI Recommendation) by introducing a gated graph convolutional neural network module, thus solving the technical problem in existing technologies that cannot effectively filter interference noise when modeling common user trajectories. Furthermore, this invention combines temporal-spatial fusion and temporal-category fusion modules to deeply mine spatiotemporal features from multiple perspectives. This invention improves the accuracy of next POI recommendations while simplifying the GETNext model structure and significantly reducing model computation time.
[0121] It should be noted that embodiments of the present invention can be provided as methods, systems, and / or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0122] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.
[0123] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.
[0124] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.
[0125] Contents not described in detail in this specification are prior art known to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications are readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of this invention are protected.
Claims
1. A POI recommendation method that integrates crowd movement trajectories and spatiotemporal-category information, characterized in that: Includes the following steps: Step 1: Obtain user check-in sequence data and perform data preprocessing and regional trajectory map construction; Step 2: Construct a POI recommendation model that integrates crowd movement trajectories and spatiotemporal-category data; Step 3: Input the user check-in sequence data and regional trajectory map obtained in Step 1 into the model obtained in Step 2 to make the next POI recommendation; The region trajectory graph mentioned in step 1 is defined as: a directed graph consisting of a node set V, an edge set E, node attributes S, and weights W; the node set V consists of POIs visited historically by the user; the edge set E is a set of directed edges, where the user has visited two POIs consecutively within a certain period of time: and Then in and Add a directed edge e between them, wherein the directed edge e is the user's... arrive The weight W of each directed edge is equal to the number of times the sub-trajectory appears in the region trajectory graph; The node attribute S contains the spatial coordinates, category, and access frequency of each POI; The POI recommendation model in step 2 includes a region trajectory map learning module, a feature learning fusion layer, and an encoder-decoder structure. The region trajectory map learning module performs feature aggregation and noise filtering on the region trajectory map through a graph convolutional neural network with gated recurrent units, and learns the features of each POI. The feature learning fusion layer performs preliminary feature extraction and fusion of information from different dimensions in the user check-in sequence data through the POI-User fusion module, the time-space fusion module, and the time-category fusion module; the encoder extracts features from the user trajectory sequence; and the decoder predicts the next POI, the access time period, and the POI category through three multilayer perceptrons. The region trajectory map learning module performs feature aggregation and noise filtering on the region trajectory map through a graph convolutional neural network with gated recurrent units, learning the features of each POI. The specific learning process includes: Step 2.1: Calculate the normalized Laplacian matrix of the adjacency matrix of the region trajectory map. ; Step 2.2: Feature extraction using a single-layer graph convolutional neural network; Step 2.3: Update the hidden state of each node using the gated loop unit; Step 2.4: After repeating steps 2.2 and 2.3 multiple times, combine the obtained feature vector with the feature vector extracted in the first graph convolution. The features of each POI are obtained by concatenating and merging the data. The calculation process is expressed by the following formula: ； In the formula, This represents a single-layer graph convolution operation. This represents a gated loop unit with time step T. and These represent the weights and biases, respectively. Indicates splicing, For Laplace matrix, For feature vectors; The user trajectory sequence is defined as: composed of a user's continuous check-in records within 48 hours in chronological order; based on the POI-User fusion feature representation of the user's check-in activities. Temporal-spatial fusion feature representation and time-category fusion feature representation Then, the embedded representation of the user check-in sequence data q is divided into two parts: and ; trajectory sequence of each user Both are embedded from two user check-in sequence data. and express.
2. The POI recommendation method that integrates crowd movement trajectories and spatiotemporal-category analysis according to claim 1, characterized in that: The feature learning fusion layer performs preliminary feature extraction and fusion on information from different dimensions in the user check-in sequence data through the POI-User fusion module, the time-space fusion module, and the time-category fusion module. The specific learning process includes: Step 2.5: Embed user u's personal information into a user feature vector. Representation; the feature vector of each POI in the user check-in data. With user feature vector The data is concatenated and then nonlinearly transformed in a fully connected layer to obtain the POI-User fusion feature representation for each user's check-in activity. The calculation process is expressed by the following formula: ； In the formula, and These represent the weights and biases, respectively. Indicates splicing, LeakyReLU represents the non-linear activation function. Step 2.6: Obtain timestamp feature vectors using the encoding strategy of a neural network model. The coordinates of each POI are input into the fully connected layer to obtain spatial feature vectors. The timestamp feature vector and spatial feature vector are concatenated and input into a fully connected layer for feature fusion, constructing a time-space fusion feature representation of user check-in activities. The calculation process is expressed by the following formula: ； In the formula, Represents a learnable weight vector. Indicates deviation, Leaky ReLU represents the non-linear activation function; Step 2.7: Encode the absolute access time using a neural network model encoding strategy; embed each POI category into a feature vector. Representation; class feature vector and time feature vector The data is then concatenated and fed into a fully connected layer for feature fusion to construct a time-category fusion feature representation of user check-in activities. The calculation process is expressed by the following formula: ； In the formula, This represents a learnable weight vector. Indicates deviation, This represents the non-linear activation function Leaky ReLU.
3. The POI recommendation method according to claim 2, which integrates crowd movement trajectories and spatiotemporal-category analysis, is characterized in that: The encoder uses a Transformer to extract user trajectory sequence features. The specific extraction process includes: Step 3.1: Embedding of a given user trajectory sequence and ,Will and These are passed as input tensors to the Transformer encoder. Step 3.2: For the m-th layer of the encoder, the input... and First, a transformation is performed using a multi-head self-attention layer, and the calculation process is expressed by the following formula: ； ； ； In the formula, , , These represent the query weight matrix, key weight matrix, and value weight matrix, respectively. This indicates the correlation between the i-th and j-th check-ins. This indicates the influence weight of other check-in activities on the current check-in, and uses the softmax function to ensure that the sum of attention weights is 1; Step 3.3: Stack the outputs from different attention heads and use layer normalization and residual connection techniques. The calculation process is shown in the following formula: ； Step 3.4: After the multi-head self-attention layer, a feedforward neural network (FNN) and the ReLU activation function are used to give the model nonlinearity. The calculation process is shown in the following formula: ； ； in , Indicates weight, , Indicates deviation, This represents the final output of the encoder at layer m. Step 3.5: After stacking multiple multi-head self-attention blocks, the output feature Z of the Transformer encoder is transformed into temporal-space output features. Time-Category Output Features The weighted sum is defined as follows: ； in This represents the activation function tanh. and This represents two aggregation hyperparameters.
4. The POI recommendation method that integrates crowd movement trajectories and spatiotemporal-category analysis according to claim 3, characterized in that: The decoder uses three multilayer perceptrons to predict the next POI: Access time period and POI categories The definition is as follows: ； ； ； In the formula, Z represents the output characteristic of the Transformer encoder. , , Indicates weight, , , Indicates deviation.
5. The POI recommendation method according to claim 4, which integrates crowd movement trajectories and spatiotemporal-category analysis, is characterized in that: The overall loss function of the POI recommendation model is defined as follows: ； In the formula, This represents the overall loss function of the network model. and Let these represent the loss functions for predicting the next POI and POI category, respectively, using the common cross-entropy loss. The mean squared error (MSE) was used for performance evaluation of access time prediction.