Interest point recommendation method based on graph enhanced user context information network

Abstract

The invention provides an interest point recommendation method based on a graph-enhanced user context information network. The interest point recommendation method comprises the following steps: acquiring historical track data of a user and preprocessing the historical track data; utilizing an interest point supplementing module to fill the interest point between two adjacent interest points of which the road network distance and the sign-in time difference are greater than a threshold value in the historical trajectory data to obtain enhanced trajectory data; inputting the enhanced track data into a user context information embedding module to extract a context information embedding vector of the user; according to the enhanced trajectory data of all users in the network, constructing a global trajectory graph by taking the interest points as nodes and taking a sign-in sequence of the users to the adjacent interest points as edges; inputting the global trajectory graph into a general travel module, respectively extracting POI embedding and user embedding by utilizing a GCN network and an MCL network to obtain a POI embedding matrix and a user embedding matrix, and fusing the POI embedding matrix and the user embedding matrix to obtain a POI-user embedding vector of the user; and inputting the context information embedded vector of the user, the POI-user embedded vector of the user and the global trajectory graph into a prediction module for prediction to obtain an interest point recommendation result of the user. According to the method, the core problem in the prior art is effectively solved, the accuracy and scene adaptability of recommendation of the next interest point are remarkably improved, and the method has important application value and wide development prospects in the fields of intelligent transportation, personalized position service and the like.

Description

Technical Field

[0001] The present invention belongs to the field of point-of-interest recommendation, and particularly relates to a point-of-interest recommendation method based on a graph-enhanced user context information network. Background Art

[0002] Next Point-Of-Interest Recommendation, as an important research direction in the fields of mobile computing and spatio-temporal data mining, is dedicated to accurately predicting the probability that a user will visit a point of interest (POI) at the next moment based on the user's historical trajectory sequence. With the booming development of mobile Internet and high-precision positioning technology, the massive spatio-temporal behavior data generated by intelligent terminal devices deeply contains users' travel patterns, preference characteristics, and multi-dimensional context correlation information, which not only has important academic research value but also shows great application potential in fields such as intelligent transportation scheduling, personalized location services, and urban resource optimization allocation.

[0003] However, the current technologies in this field face many challenges in practical applications. At the level of spatio-temporal context feature mining, most existing methods adopt a structure that separates temporal modeling and spatial encoding. For example, a recurrent neural network (RNN) is independently used to process time series, and a graph network is used to analyze spatial relationships. This separated architecture leads to a static or decoupled state in spatio-temporal feature fusion, making it difficult to dynamically capture complex spatio-temporal correlations. Although some attention mechanism-based methods can enhance local feature interactions, there are significant deficiencies in co-modeling the dynamic evolution of time sensitivity and spatial dependence, and they cannot accurately depict the dynamically changing spatio-temporal behavior patterns in the user's movement trajectory.

[0004] In terms of user behavior modeling, existing technologies overly rely on the static adjacency matrix constructed by explicit historical interactions, seriously neglecting the dynamic evolution characteristics of potential implicit preferences, such as the influence of similar user groups and semantic associations between points of interest. As a result, the ability to capture user preference drift and implicit interactions is limited. In addition, for the problems of data sparsity and long-tail distribution, although there are already technologies such as matrix completion and similarity filtering to alleviate the data sparsity situation, there are defects in dynamic preference modeling and global semantic association integration, lacking a cross-domain semantic enhancement strategy for long-tail partition POIs, making the recommendation system lack robustness in scenarios with unbalanced data distribution. Summary of the Invention

[0005] In order to solve the problems in the background art, one aspect of the present invention provides a point-of-interest recommendation method based on a graph-enhanced user context information network, which is characterized by including:

[0006] S1: Obtain the historical trajectory data of the user and perform preprocessing. Among them, the historical trajectory data of the user includes: the ID of the user, the ID of the point of interest, the category of the point of interest, the coordinates of the point of interest, and the check-in time of the user at the point of interest;

[0007] S2: Use the point of interest supplement module to fill in points of interest between two adjacent points of interest with a road network distance and check-in time difference greater than the threshold in the historical trajectory data to obtain enhanced trajectory data;

[0008] S3: Input the enhanced trajectory data into the user context information embedding module to extract the user's context information embedding vector;

[0009] S4: According to the enhanced trajectory data of all users in the network, use the point of interest as a node and the check-in order between adjacent points of interest of the user as an edge to construct a global trajectory graph;

[0010] S5: Input the global trajectory graph into the general travel module, use the GCN network and the MCL network to extract the POI embedding and user embedding respectively to obtain the POI embedding matrix and the user embedding matrix, and fuse the POI embedding matrix and the user embedding matrix to obtain the user's POI-user embedding vector;

[0011] S6: Input the user's context information embedding vector, the user's POI-user embedding vector and the global trajectory graph into the prediction module for prediction to obtain the point of interest recommendation result of the user.

[0012] Another aspect of the present invention provides a point of interest recommendation system based on a graph-enhanced user context information network, characterized in that the system includes a memory and a processor; the memory is used to store application programs; the processor is used to run the application programs and execute the above-mentioned method for recommending points of interest based on a graph-enhanced user context information network.

[0013] Another aspect of the present invention provides a computer storage medium, characterized in that a remote monitoring program is stored on the computer storage medium, and when the remote monitoring program is executed by a processor, the above-mentioned method for recommending points of interest based on a graph-enhanced user context information network is implemented.

[0014] The present invention has at least the following beneficial effects

[0015] In terms of spatio-temporal context feature mining of the present invention, traditional methods are difficult to dynamically capture complex spatio-temporal correlations. However, the present invention uses the Markov Clustering Algorithm (MCL) and the Graph Convolutional Network (GCN) to perform embedded representation on users and points of interest. On this basis, multi-dimensional information such as the category of points of interest, time, and geographical distance is fused to generate a spatio-temporal context embedding sequence that is fully mined. This method breaks the limitations of static and separate spatio-temporal feature fusion, can dynamically and comprehensively capture the complex spatio-temporal correlations in the user's movement trajectory, accurately reflect the dynamically changing spatio-temporal behavior pattern of the user, and improve the accuracy of predicting the user's next point of interest. For user behavior modeling, existing technologies rely too much on explicit historical interactions and cannot effectively capture implicit preferences. The present invention constructs a more complete user travel trajectory by supplementing the missing points of interest in the user check-in data, which truly reflects the user's travel pattern. At the same time, based on the embedded representation of users and points of interest, implicit correlation relationships are deeply mined, making up for the deficiencies of traditional methods, being able to more accurately capture the user's real behavior pattern, adapt to their preference drift, and thus optimize the recommendation results. In dealing with the problems of data sparsity and long-tail distribution, the present invention constructs a global trajectory flow graph and generates a trajectory probability graph to correct the model results. At the same time, the trajectory enhancement technology is used to mine potential effective information, which not only alleviates the negative impact brought by the unbalanced data distribution, but also significantly improves the user cold start and data long-tail distribution problems, making the recommendation system still have strong robustness in the scenario of data distribution imbalance. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of the overall structure of the present invention;

[0017] Figure 2 is a schematic diagram of the structure of the point-of-interest supplementation module in the present invention;

[0018] Figure 3 is a schematic diagram of the structure of the user context information embedding module in the present invention;

[0019] Figure 4 is a schematic diagram of the structure of the general travel module in the present invention;

[0020] Figure 5 is a schematic diagram of the structure of the prediction module in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0021] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0022] Please refer to Figure 1 , one aspect of the present invention provides a method for recommending points of interest based on a graph-enhanced user context information network, which is characterized by including:

[0023] S1: Obtain the historical trajectory data of the user and perform preprocessing. Among them, the historical trajectory data of the user includes: the ID of the user, the ID of the point of interest, the category of the point of interest, the coordinates of the point of interest, and the check-in time of the user for the point of interest;

[0024] In this embodiment, assume that we have a user who has visited 5 points of interest P4, P3, P5, P2, and P6 within a period of time. We can obtain historical trajectory data in the following form from relevant data sources (such as the user's mobile application, positioning system records, etc.): the ID of the user, the ID of the point of interest, the category of the point of interest, the coordinates of the point of interest, and the check-in time of the user for the point of interest;

[0025] After obtaining the original data, some preprocessing operations are usually required to ensure the quality and usability of the data. The following are some common preprocessing steps and examples: Check whether there are missing fields in the data. For example, if the coordinates or check-in time of a certain point of interest is missing, it may be necessary to supplement or delete the record. Check whether the check-in time is reasonable and whether the coordinates are within the valid range, etc. If it is found that a certain check-in time is significantly abnormal (such as the check-in time is a future time), then it is necessary to correct or delete the record. If the number of historical check-ins of a certain user is less than 3, the relevant data of the user will be deleted. Similarly, if the number of times a certain point of interest is checked in by users is less than 3, the relevant data of the point of interest will be deleted.

[0026] Please refer to Figure 2 , S2: Use the point-of-interest supplement module to fill in points of interest between two adjacent points of interest with a road network distance and a check-in time difference greater than a threshold value in the historical trajectory data to obtain enhanced trajectory data;

[0027] In this embodiment, first, based on the user's historical trajectory data U = {P4, P3, P5, P2, P6}, calculate the road network distance and check-in time between two adjacent points of interest in the historical trajectory data; determine whether the road network distance and check-in time between two adjacent points of interest in the historical trajectory data are greater than the set threshold. For example, when the road network distance (usually taking the average of the walking road network distance, cycling road network distance, and driving road network distance) between point of interest P5 and point of interest P2 is greater than the set threshold, and the check-in time difference between point of interest P5 and point of interest P2 is greater than the set threshold, then an intermediate point of interest needs to be filled in at the intermediate moment between point of interest P5 and point of interest P2. Suppose the filled point of interest is calculated as P1 through steps S21 - S25, then the user's historical trajectory data U1 after filling is {P4, P3, P5, P1, P2, P6}. After filling, fill the user's historical trajectory data U1 again. Through calculation, it is found that the road network distance and check-in time difference between adjacent points of interest in sequence U1 do not meet the conditions, so no further filling is performed; if there are still two adjacent points of interest in U1 that meet the above conditions, steps S21 - S25 can continue to be used to fill between the adjacent points of interest.

[0028] Preferably, step S2 includes:

[0029] S21: Use the Masking mechanism to generate the location coordinate mask and category mask of the filled point of interest at the intermediate moment between two points of interest;

[0030] In this embodiment, if it is found that an intermediate point of interest needs to be filled between point of interest P5 and point of interest P2, at the intermediate moment of their check-in times, such as 12:45, use the Masking mechanism to generate the location coordinate mask and category mask of the filled point of interest. The location coordinate mask is used to determine the position of the filled point of interest subsequently, and the category mask is used to determine its category.

[0031] S22: Embed the location coordinate sequence of the points of interest in the masked trajectory data to obtain the location embedding vector of the masked trajectory data, and embed the user ID to obtain the user embedding vector; embed the category sequence of the points of interest in the masked trajectory data to obtain the category embedding vector of the masked trajectory data;

[0032] Embedding the location coordinate sequence of the point of interest in the masked trajectory data in this embodiment includes: First, calculate the mean and standard deviation of the coordinates of all points of interest in the network, standardize the coordinates of the points of interest based on the calculated mean and standard deviation, and use a linear model such as a fully connected layer to map them to a high-dimensional space to obtain the location embedding; Concatenate the location embeddings of all points of interest in the trajectory data to obtain the location embedding vector of the masked trajectory data; Similarly, an embedding model can also be directly used for embedding. For the user ID, an embedding model can be directly used for embedding; The category sequence of the points of interest in the masked trajectory data is directly embedded using an embedding model.

[0033] S23: Multiply the location embedding vector of the trajectory data, the user embedding vector, and the category embedding vector of the masked trajectory data to obtain an intermediate feature vector, and add position markers to each feature of the intermediate feature vector to obtain the input feature vector;

[0034] In this embodiment, multiply the location embedding vector of the trajectory data, the user embedding vector, and the category embedding vector of the masked trajectory data to obtain an intermediate feature vector. Then, add position markers to each feature of the intermediate feature vector to form the input feature vector. This position marker can reflect the position information of each feature in the sequence.

[0035] S24: Process the input feature vector sequentially through a multi-head bidirectional attention network, a GRU module, and an activation function to obtain the missing point of interest;

[0036] In this embodiment, process the input feature vector sequentially through a multi-head bidirectional attention network, a GRU module, and an activation function. The multi-head bidirectional attention network can capture the dependencies between different features, the GRU module can handle the temporal dependencies in the sequence data, and the activation function performs a non-linear transformation on the output. Finally, relevant information about the missing point of interest, such as location coordinates and category, is obtained.

[0037] S25: Fill the missing point of interest into the intermediate time between two points of interest to obtain the enhanced trajectory data.

[0038] In this embodiment, by calculating the road network distance and check-in time difference between adjacent points of interest, it is determined whether points of interest need to be filled, making the originally possibly missing trajectories become coherent and more truly reflecting the user's travel path, providing a more reliable data basis for subsequent analysis and applications. The location coordinate mask and category mask generated by the Masking mechanism help to accurately locate and determine the attributes of the filled points of interest. A variety of vector embedding operations convert different types of data into computable vector forms, and then by multiplying features and adding position markers, multi-dimensional information is fused, enabling better capture of the spatio-temporal correlation between points of interest and the user's behavior pattern. In terms of model processing, the input feature vector is processed by a multi-head bidirectional attention network, a GRU module, and an activation function, fully considering the dependency relationship between features and time dependency, and the output missing point of interest information is more in line with the actual situation. Finally, through continuous filling and iteration, the enhanced trajectory data obtained can improve the accuracy and reliability of the next point of interest recommendation. For fields such as intelligent transportation scheduling and personalized location services, these more complete and accurate trajectory data can support more scientific decision-making and provide users with higher-quality and more demand-tailored services.

[0039] Please refer to Figure 3 , S3: Input the enhanced trajectory data into the user context information embedding module to extract the user's context information embedding vector;

[0040] Preferably, the extraction of the user's context information embedding vector includes:

[0041] S31: Perform feature embedding on the category sequence of the points of interest in the enhanced trajectory data to obtain the POI category embedding vector;

[0042] First, encode the category sequence based on one-hot to generate the one-hot vector of the category sequence, and multiply the one-hot vector by the adjustable embedding matrix one-hot vector W , ,

[0043] , , , embed ,

[0042] , T , ,

[0046] , i ,

[0045] , ,

[0044] , Then perform embeding to obtain the POI category embedding.

[0043] S32: Perform feature embedding on the check-in time of the user for the points of interest in the enhanced trajectory data to obtain the time embedding vector;

[0044] Divide the 24 hours of a day into I time slots, and the feature embedding representation of each time slot i is:

[0045]

[0046] Among them, represents the time embedding of the i-th time slot, t i represents the i-th time slot; cos represents the function, f embed represents the embeding model; wi and θ i Learnable parameters; obtain time embeddings based on the user's check-in time at points of interest.

[0047] S33: Perform feature embedding on the road network distance between two adjacent points of interest in the enhanced trajectory data to obtain a road network distance embedding vector;

[0048] Preferably, the network distance embedding vector includes:

[0049] Embed d = f embed (mmn(α1d walk + α2d bike + α3d drive ))

[0050] α1 + α2 + α3 = 1

[0051] where Embed d represents the road network distance embedding between two adjacent points of interest, f embed represents the embedding model; mmn represents min-max normalization; d walk represents the road network distance between two adjacent points of interest by walking; d bike represents the road network distance between two adjacent points of interest by cycling; d drive represents the road network distance between two adjacent points of interest by riding a bike.

[0052] S34: Input the POI category embedding vector, time embedding vector, and road network distance embedding vector into the fully connected layer respectively, map them to the same dimension, and then perform feature addition to obtain the user's context information embedding vector.

[0053] From a data processing perspective, this step in this embodiment transforms complex contextual information into a computable vector form by embedding features across different types of data, making the data easier to process and analyze by subsequent models. By embedding the POI category sequence, the user's check-in time at the POI, and the road network distance between two adjacent POIs, it effectively integrates multiple pieces of information and comprehensively reflects the user's behavioral characteristics in both temporal and spatial dimensions. Regarding feature mining, the specific embedding method is highly targeted and effective. For the POI category sequence, one-hot encoding and an adjustable embedding matrix are used to highlight the differences between different categories and uncover potential correlations between POI categories and user behavior. When processing check-in times, the day is divided into time slots and feature embedding is performed, taking into account the periodicity and regularity of time, helping to capture user behavior patterns across different time periods. For the embedding of road network distance, the distances of walking, cycling, and driving are comprehensively considered and normalized, accurately reflecting the spatial relationships between POIs and exploring the impact of spatial factors on user behavior. From a practical application perspective, the extracted contextual information embedding vector can provide richer and more accurate information for the next POI recommendation. These vectors can better characterize user behavior and preferences, enabling recommendation systems to more accurately predict a user's next interest, improving the accuracy and personalization of recommendations. Furthermore, these vectors can help make more informed decisions in areas such as smart traffic scheduling and urban resource allocation, and have broad application value.

[0054] S4: Based on the enhanced trajectory data of all users in the network, a global trajectory graph is constructed by taking points of interest as nodes and the check-in order between adjacent points of interest as edges;

[0055] In this embodiment, the global trajectory graph is constructed based on the enhanced trajectory data of all users. For example, if the historical trajectory data of a user U1 = {P4, P3, P5, P1, P2, P6}, then there is an edge from point of interest P4 to P3 in the global trajectory graph. If there is another user who also checks in sequentially from point of interest P4 to P3, the weights of points of interest P4 and P3 are increased, thereby constructing the global trajectory graph. The global trajectory graph is a weighted directed graph structure.

[0056] See also Figure 4 ,S5: Input the global trajectory graph into the general travel module, use the GCN network and MCL network to extract POI embedding and user embedding respectively to obtain the POI embedding matrix and user embedding matrix, and fuse the POI embedding matrix and the user embedding matrix to obtain the user's POI-user embedding vector;

[0057] Preferably, step S5 includes:

[0058] S51: Embed the information of the point of interest to obtain the initial features of the point of interest, and construct the feature matrix H0 of the global trajectory map;

[0059] S52: Use the GCN network to perform feature aggregation on the feature matrix H0 of the global trajectory map and the adjacency matrix to obtain the POI embedding matrix of the global trajectory map;

[0060] S53: Calculate the POI category similarity between users according to the categories of points of interest visited by different users, and calculate the check-in time similarity between users according to the check-in time of different users for the points of interest;

[0061] S54: Perform weighted fusion on the POI category similarity and the check-in time similarity between users to obtain the comprehensive similarity between users, and construct the user comprehensive similarity matrix S;

[0062] Preferably, the user comprehensive similarity S ij includes:

[0063]

[0064] where W1 and W2 represent weight matrices; C i represents the set of check-in POI categories of user i; C j represents the set of POI categories checked in by user j; S ij ∈S represents the comprehensive similarity between user i and user j; T i represents the set of check-in times of user i, and T j represents the set of check-in times of user j.

[0065] S55: Input the user comprehensive similarity matrix into the MCL network for random walk to obtain a convergence matrix, and perform feature embedding on the convergence matrix to obtain a user embedding matrix;

[0066] In this embodiment, the Markov clustering algorithm MCL is introduced. First, the Jaccard similarity is used to calculate the POI category similarity and the check-in time similarity of the user trajectories respectively, and they are linearly fused to obtain the user similarity. According to the user similarity S ij construct the similarity matrix A, perform power operation on the similarity matrix A to simulate the expansion of user random walk, and then perform contraction through the exponential operation and normalization of matrix elements to strengthen the connections with higher similarity. Alternate the expansion and contraction steps to obtain the convergence matrix B:

[0067]

[0068] where p represents the exponent of the power operation; j represents the index of summation, indicating summing the elements of each column (or row, depending on the rules of matrix operations) in the matrix.

[0069] S56: Input the POI embedding matrix and the user embedding matrix into the linear fusion module to calculate the dynamic weight matrix of the user for the point of interest;

[0070] S57: Update the adjacency matrix of the global trajectory map according to the dynamic weight of the user for the point of interest;

[0071] S58: Based on the user embedding matrix and the adjacency matrix of the updated global trajectory map, use the GCN network for feature aggregation to obtain the enhanced POI embedding matrix of the global trajectory map;

[0072] Preferably, the adjacency matrix of the updated global trajectory map includes:

[0073] W dy = σ(Embed u T WEmbed p + b)

[0074] A en = A + γ·W dy

[0075] where, W dy represents the dynamic weight matrix of the user for the point of interest; σ represents the activation function; Embed u represents the user embedding matrix; Embed p represents the POI embedding matrix; W and γ represent the weight parameter matrices; b represents the bias matrix; A en represents the adjacency matrix of the updated global trajectory map; A represents the adjacency matrix of the global trajectory map before update.

[0076] S59: Add the enhanced POI embedding matrix and the user embedding matrix to obtain the POI - user embedding matrix, and extract the POI - user embedding vector of the user from the POI - user embedding matrix.

[0077] In this embodiment, a feature matrix is constructed by performing feature embedding on the point-of-interest (POI) information, and a POI embedding matrix is obtained by aggregating features using a graph convolutional network (GCN), which can fully explore the associations and features among POIs. Meanwhile, the similarity is calculated based on the category and check-in time of the POIs visited by the user to construct a comprehensive user similarity matrix, and then a user embedding matrix is obtained through random walk using the Markov clustering (MCL) network, comprehensively capturing the user's behavioral characteristics and preferences. The POI embedding matrix and the user embedding matrix are linearly fused to calculate a dynamic weight matrix, the adjacency matrix of the global trajectory graph is updated according to the dynamic weight, and then an enhanced POI embedding matrix is obtained using the GCN network. Finally, it is added to the user embedding matrix to obtain a POI-user embedding matrix and extract vectors. This series of operations enables a more accurate characterization of the relationship between the user and the POI, and can better reflect the user's dynamic preferences for different POIs. The obtained POI-user embedding vectors provide richer and more accurate information for the recommendation of the next POI. Compared with traditional methods, this approach that combines multi-faceted features and a dynamic update mechanism can better adapt to the behavioral patterns and preference changes of different users, improving the accuracy and personalization of recommendations. At the same time, for fields such as intelligent transportation scheduling and urban resource allocation, it can also provide more valuable decision-making basis, helping to optimize resource allocation and improve service quality. By dynamically updating the adjacency matrix and performing multiple feature aggregations, the adaptability and robustness of the model to data changes are enhanced, and it can still maintain good performance in the presence of certain noise or uncertainty in the data.

[0078] Please refer to Figure 5 , S6: Input the user's context information embedding vector, the user's POI-user embedding vector, and the global trajectory graph into the prediction module to obtain the user's POI recommendation result.

[0079] Preferably, the step S6 includes:

[0080] S61: Input the feature matrix H0 of the global trajectory graph into the self-attention module to capture the feature information of the global trajectory graph, then calculate the normalized Laplacian matrix according to the adjacency matrix A of the global trajectory graph, and finally obtain the global probability graph by combining the output of the self-attention module

[0081] Q = HW Q , K = HW K , V = HW V

[0082]

[0083] where W Q 、W K and W V represent weight matrices; d krepresents the scaling factor, D represents the degree matrix of the adjacency matrix A; I N represents the identity matrix;

[0084] S62: Add the feature of the user's context information embedding vector and the user's POI-user embedding vector to obtain a sequence information embedding vector;

[0085] S63: Input the sequence information embedding vector into the Transformer encoder for encoding to obtain encoded features, input the encoded features into the MLP for decoding to obtain an initial prediction result, and correct the initial prediction result through the global probability graph to obtain the POI recommendation result of the user.

[0086] In this embodiment, by combining the self-attention module and the normalized Laplacian matrix to process the global trajectory graph, the feature information and the relationship between nodes in the global trajectory graph can be accurately captured. The self-attention mechanism enables the model to focus on the importance of different parts of the trajectory graph, and the normalized Laplacian matrix further processes the adjacency matrix, enabling the effective utilization of the graph structure information. The finally obtained global probability graph provides comprehensive and accurate global information for the subsequent correction of the prediction result, which helps the recommendation result to fit the actual trajectory distribution. Adding the feature of the user's context information embedding vector and the POI-user embedding vector realizes the fusion of multi-dimensional information. The context information embedding vector contains context information such as POI category, check-in time, and road network distance, while the POI-user embedding vector reflects the dynamic relationship between the user and the POI. The sequence information embedding vector after the fusion of the two combines the user's behavioral characteristics, preferences, and the association between POIs, providing richer and more representative information for the subsequent prediction. Using the Transformer encoder and the MLP for encoding and decoding gives full play to the advantages of the Transformer in processing sequence data, and can learn the long-distance dependence relationship in the sequence, so as to obtain a more accurate initial prediction result. Then, through the global probability graph corrects the initial prediction result, further improving the accuracy and reliability of the recommendation result. This prediction method combining multiple modules and multiple information sources can better adapt to the behavioral patterns and preference changes of different users, provide more personalized and accurate POI recommendations for users, and improve the performance and user experience of the recommendation system.

[0087] Another aspect of the present invention provides a POI recommendation system based on a graph-enhanced user context information network, characterized in that the system includes a memory and a processor; the memory is used to store application programs; the processor is used to run the application programs and execute the above-mentioned method for a POI recommendation method based on a graph-enhanced user context information network.

[0088] In another aspect of the present invention, there is provided a computer storage medium, characterized in that a remote monitoring program is stored on the computer storage medium, and when the remote monitoring program is executed by a processor, the above-mentioned method for recommending points of interest based on a graph-enhanced user context information network is implemented.

[0089] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0090] In summary, in terms of spatio-temporal context feature mining in the present invention, traditional methods are difficult to dynamically capture complex spatio-temporal correlations. However, the present invention uses the Markov Clustering Algorithm (MCL) and Graph Convolutional Network (GCN) to perform embedded representations of users and points of interest. On this basis, multi-dimensional information such as point-of-interest categories, time, and geographical distance is fused to generate a spatio-temporal context embedded sequence that is fully mined. This approach breaks the limitations of static and separate spatio-temporal feature fusion, can dynamically and comprehensively capture the complex spatio-temporal correlations in the user's movement trajectory, accurately reflect the dynamically changing spatio-temporal behavior pattern of the user, and improve the accuracy of predicting the user's next point of interest. For user behavior modeling, existing technologies rely too much on explicit historical interactions and cannot effectively capture implicit preferences. The present invention constructs a more complete user travel trajectory by supplementing the missing points of interest in the user check-in data, which truly reflects the user's travel pattern. At the same time, based on the embedded representations of users and points of interest, implicit correlation relationships are deeply mined, making up for the deficiencies of traditional methods, being able to more accurately capture the true behavior pattern of the user, adapt to their preference drift, and thus optimize the recommendation results. In dealing with the problems of data sparsity and long-tail distribution, the present invention constructs a global trajectory flow graph and generates a trajectory probability graph to correct the model results. At the same time, the trajectory enhancement technology is used to mine potential effective information, which not only alleviates the negative impact brought by the unbalanced data distribution, but also significantly improves the user cold start and data long-tail distribution problems, making the recommendation system still have strong robustness in the scenario of unbalanced data distribution.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for recommending points of interest based on a graph-enhanced user context information network, characterized in that Including: S1: Obtain the historical trajectory data of the user and perform preprocessing. Among them, the historical trajectory data of the user includes: the ID of the user, the ID of the point of interest, the category of the point of interest, the coordinates of the point of interest, and the check-in time of the user at the point of interest; S2: Use the point of interest supplement module to fill in points of interest between two adjacent points of interest with a road network distance and check-in time difference greater than a threshold value in the historical trajectory data to obtain enhanced trajectory data; S3: Input the enhanced trajectory data into the user context information embedding module to extract the user's context information embedding vector; S4: According to the enhanced trajectory data of all users in the network, use the point of interest as a node and the check-in order between adjacent points of interest of the user as an edge to construct a global trajectory graph; S5: Input the global trajectory graph into the general travel module, use the GCN network and the MCL network to extract the POI embedding and user embedding respectively to obtain the POI embedding matrix and the user embedding matrix, and fuse the POI embedding matrix and the user embedding matrix to obtain the user's POI-user embedding vector; S6: Input the user's context information embedding vector, the user's POI-user embedding vector and the global trajectory graph into the prediction module for prediction to obtain the point of interest recommendation result of the user.

2. The method for recommending points of interest based on a graph-enhanced user context information network according to claim 1, wherein The step S2 includes: S21: Use the Masking mechanism to generate a location coordinate mask for filling in the point of interest and a category mask for filling in the point of interest at the intermediate moment between two points of interest; S22: Embed the location coordinate sequence of the point of interest in the masked trajectory data to obtain the location embedding vector of the masked trajectory data, and embed the ID of the user to obtain the user embedding vector; Embed the category sequence of the point of interest in the masked trajectory data to obtain the category embedding vector of the masked trajectory data; S23: Multiply the features of the location embedding vector of the trajectory data, the user embedding vector and the category embedding vector of the masked trajectory data to obtain an intermediate feature vector, and add a position marker to each feature of the intermediate feature vector to obtain the input feature vector; S24: Process the input feature vector through a multi-head bidirectional attention network, a GRU module and an activation function in sequence to obtain the missing point of interest; S25: Fill in the missing point of interest at the intermediate moment between two points of interest to obtain enhanced trajectory data.

3. The method for recommending points of interest based on a graph-enhanced user context information network according to claim 1, characterized in that, The extraction of the user's context information embedding vector includes: S31: Embed the category sequence of the point of interest in the enhanced trajectory data to obtain the POI category embedding vector; S32: Embed the check-in time of the user at the point of interest in the enhanced trajectory data to obtain the time embedding vector; S33: Embed the road network distance between two adjacent points of interest in the enhanced trajectory data to obtain the road network distance embedding vector; S34: Input the POI category embedding vector, the time embedding vector and the road network distance embedding vector into the fully connected layer to map them to the same dimension and then add the features to obtain the user's context information embedding vector.

4. The method for recommending points of interest based on a graph-enhanced user context information network according to claim 3, wherein The road network distance embedding vector includes: Embed d = f embed (mmn(α1d walk + α2d bike + α3d drive )) α1+α2+α3＝1 Among them, Embed d represents the road network distance embedding between two adjacent points of interest, and f embed represents the embedding model; mmn represents the maximum-minimum normalization; d walk represents the road network distance between two adjacent points of interest by walking; d bike represents the road network distance between two adjacent points of interest by cycling; d drive represents the road network distance between two adjacent points of interest by riding a bike.

5. The method for recommending points of interest based on a graph-enhanced user context information network according to claim 1, wherein The step S5 includes: S51: Embed the information of the point of interest to obtain the initial feature of the point of interest, and construct the feature matrix H0 of the global trajectory graph; S52: Aggregate features using the GCN network based on the feature matrix H0 and the adjacency matrix of the global trajectory map to obtain the POI embedding matrix of the global trajectory map; S53: Calculate the POI category similarity between users according to the categories of points of interest visited by different users, and calculate the check-in time similarity between users according to the check-in time of users for points of interest; S54: Perform weighted fusion on the POI category similarity and the check-in time similarity between users to obtain the comprehensive similarity between users, and construct a user comprehensive similarity matrix; S55: Input the user comprehensive similarity matrix into the MCL network for random walk to obtain a convergence matrix, and perform feature embedding on the convergence matrix to obtain a user embedding matrix; S56: Input the POI embedding matrix and the user embedding matrix into a linear fusion module to calculate the dynamic weight matrix of users for points of interest; S57: Update the adjacency matrix of the global trajectory map according to the dynamic weights of users for points of interest; S58: Based on the user embedding matrix and the updated adjacency matrix of the global trajectory map, aggregate features using the GCN network to obtain the enhanced POI embedding matrix of the global trajectory map; S59: Add the enhanced POI embedding matrix and the user embedding matrix to obtain a POI-user embedding matrix, and extract the POI-user embedding vector of the user from the POI-user embedding matrix.

6. The method for recommending points of interest based on a graph-enhanced user context information network according to claim 5, characterized in that The updated adjacency matrix of the global trajectory map includes: w dy = σ(Embed u T WEmbed p + b) a en = a + γ·W dy Among them, W dy represents the dynamic weight matrix of the user for the point of interest; σ represents the activation function; Embed u represents the user embedding matrix; Embed p represents the POI embedding matrix; W and γ represent the weight parameter matrices; b represents the bias matrix; A en represents the adjacency matrix of the updated global trajectory graph; A represents the adjacency matrix of the global trajectory graph before update.

7. The method for recommending points of interest based on a graph-enhanced user context information network according to claim 1, wherein The step S6 includes: S61: Input the feature matrix H0 of the global trajectory map into the self-attention module to capture the feature information of the global trajectory map, then calculate the normalized Laplacian matrix according to the adjacency matrix A of the global trajectory map, and finally obtain the global probability map by combining the output of the self-attention module. Q = HW Q , K = HW K , V = HW V Among them, W Q , W K and W V represent weight matrices; d k represents a scaling factor, and D represents the degree matrix of the adjacency matrix A; I N represents the identity matrix; S62: Perform feature addition on the context information embedding vector of the user and the POI-user embedding vector of the user to obtain a sequence information embedding vector; S63: Embed the sequence information into a vector and input it into the Transformer encoder for encoding to obtain encoded features. Input the encoded features into the MLP for decoding to obtain an initial prediction result, and correct the initial prediction result to obtain the POI recommendation result of the user.

8. A point of interest recommendation system based on a graph-enhanced user context information network, characterized in that The system includes a memory and a processor; the memory is used to store application programs; the processor is used to run the application programs and execute a method for recommending points of interest based on a graph-enhanced user context information network according to any one of claims 1 to 7.

9. A computer storage medium, characterized in that, A remote monitoring program is stored on the computer storage medium, and when the remote monitoring program is executed by the processor, a method for recommending points of interest based on a graph-enhanced user context information network according to any one of claims 1 to 7 is implemented.

Images