Ideal location preference interest point recommendation method based on multi-factor driving

Abstract

The application provides a multi-factor driven ideal position preference interest point recommendation method, and relates to the technical field of interest point recommendation identification. The method comprises the following steps: firstly, obtaining a target trajectory sequence of a user and constructing a behavior perception interest point graph and a space perception interest point graph; secondly, using a behavior perception graph prompt module, enhancing the node representation of the behavior perception interest point graph through bidirectional information aggregation and time-space interval embedding to generate a user history sequence representation; using a spatial perception hyperbolic graph module to map the space perception interest point graph to a hyperbolic space and generating a user space preference representation through aggregating the node embedding of adjacent interest points; then, using a diffusion model in a diffusion preference contrast module and contrast learning to generate a user ideal position preference; finally, generating an interest point prediction result of the user through multi-objective joint optimization and making an interest point recommendation. The application effectively improves the accuracy of user interest point recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of point-of-interest recommendation, and particularly to a method for recommending ideal position preference points-of-interest based on multiple factors. BACKGROUND

[0002] The rapid development of smart city infrastructure has driven a surge in demand for location-based services (LBS) applications in the fields of tourism, social networks, and daily life. LBS applications use users' geographic location information to provide personalized services, such as navigation, nearby location recommendations, and ride-hailing.

[0003] A point-of-interest (POI) refers to a location point in the real world, such as a restaurant, a station, or a park. Point-of-interest recommendation systems have attracted widespread attention from both industry and academia due to their potential to meet user needs and increase service provider revenue. Point-of-interest recommendation aims to provide personalized suggestions for users' next destinations by utilizing their historical check-in behavior on location-based applications. The temporal and spatial context information generated by check-in behavior is an effective method for understanding users' behavior preferences. For example, temporal context can reveal periodic trends in user behavior, such as users' tendency to visit restaurants during the dinner hour and their consistent visit patterns on weekdays compared to weekends. Similarly, spatial context indicates users' preference for geographic locations, such as their tendency to frequently visit points-of-interest that are close to them.

[0004] However, existing research indiscriminately models subjective and non-subjective factors in user access behavior, making it difficult for models to understand complex user movement behaviors. Since user check-in behavior is influenced by multiple contextual factors and access intentions, some behaviors are driven by subjective preferences, such as frequently going to the gym for exercise or traveling from home to a favorite coffee shop. Conversely, some other behaviors are driven by external non-subjective factors, such as dining based on work tasks or attending a friend's wedding for social reasons. These non-subjective factors embedded in users' complex movement patterns may negatively impact the modeling of their true preferences. Meanwhile, current point-of-interest (POI) recommendation methods simplify the process of understanding user behavior. That is, users may envision their ideal next destination, but the location they ultimately check in at may be a real-world POI closest to this ideal intention. For example, a user who enjoys pasta might ideally seek an authentic Italian restaurant with a quiet atmosphere; this ideal intention may not correspond to any specific restaurant. The user will ultimately choose from several real-world Italian restaurants that approximate their ideal preference. Although user access behavior represents preferences, due to the potential difference between actual behavior and ideal intention, a user may not like every Italian restaurant they visit. Users compromise among numerous existing options to choose the location that best matches their intentions. This ideal location not only incorporates historical preferences but also reflects the user's subconscious intentions, though it may not correspond to specific actual candidate points of interest. Although the distributions of users' ideal and historical preferences may appear similar, implicit preference biases can distort the modeled preference signals, making it challenging to accurately capture users' true behavioral preferences. Therefore, designing a method that considers both subjective and non-subjective factors in user behavior patterns and captures users' ideal location preferences to improve the accuracy of user interest point recommendations is an urgent problem to be solved. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention proposes a multi-factor driven method for recommending points of interest based on ideal location preferences by modeling multiple factors in user access behavior and capturing users' ideal location preferences, aiming to improve the accuracy of user point of interest recommendations.

[0006] This invention proposes a multi-factor driven method for recommending ideal location preference points of interest, which includes the following process:

[0007] By acquiring all points of interest accessed by several users and the access time for each point of interest, a target trajectory sequence for each user is constructed.

[0008] Construct behavior-aware interest point maps and spatial-aware interest point maps based on the user's target trajectory sequence;

[0009] The behavior-aware graph prompting module enhances the node representation of the behavior-aware interest point graph through bidirectional information aggregation and time-space separation embedding, and generates the user historical sequence representation in combination with prompt learning and attention mechanism;

[0010] The spatial-aware hyperbolic graph module maps the spatial-aware interest point graph to the hyperbolic space, and projects back to the hyperbolic space after aggregating the node embeddings of adjacent interest points on the tangent space through a multi-layer graph convolution network, to generate the user spatial preference representation;

[0011] Based on the user historical sequence representation and the user spatial preference representation, the diffusion preference contrast module is used to generate the user geographic preference representation, and based on the user geographic preference representation, a diffusion model and contrast learning are used to generate the user ideal location preference;

[0012] The user historical sequence representation and the user spatial preference representation are used to generate the interest point prediction result of the user through the joint user ideal location preference;

[0013] Based on the generated interest point prediction result, the interest point recommendation is performed for the user.

[0014] Further, the method for constructing the behavior-aware interest point graph and the spatial-aware interest point graph according to the target trajectory sequence of the user is as follows:

[0015] For any user , the user visited each interest point as a vertex, and a directed edge is established according to the time sequence of the user visiting each interest point, so as to construct a directed graph according to the target trajectory sequence of the user and take it as the behavior-aware interest point graph of the user ; for the interest point visited by the user and the interest point , if the user subsequently visited the interest point after visiting the interest point , then there is a directed edge between the vertex representing the interest point and the vertex representing the interest point ;

[0016] According to the target trajectory sequence of all users, all interest points visited by the users are counted; for all counted interest points, each interest point is taken as a vertex, and a vertex set of the spatial-aware interest point graph is constructed; the distance edge is defined to represent any two interest points and The distances between them are determined, and an edge set of a spatially aware point of interest graph is constructed; edge weights are defined. For any two points of interest and Based on the proximity between points, an edge weight set is constructed for the spatially perceived interest point graph. An undirected spatially perceived interest point graph is then constructed using the vertex set, edge set, and edge weight set of the spatially perceived interest point graph. .

[0017] Furthermore, the behavior-aware graph cueing module enhances the node representation of the behavior-aware interest point graph through bidirectional information aggregation and spatiotemporal embedding, and generates the user's historical sequence representation by combining cueing learning and attention mechanisms. The specific content of this representation is as follows:

[0018] For target users The user's behavioral perception interest map Interest points are mapped to vector representations, yielding the input representation of the interest point node embedding. ;

[0019] Aggregating behavior-aware interest point graphs using message functions The bidirectional edge information of each interest point is used, and the input representation of the interest point node embedding is updated by combining the time interval embedding matrix and the spatial interval embedding matrix. The updated node representations in the behavior-aware interest point graph are obtained and used as the original node representations. ;

[0020] Initialize sequence hint vector Using the Conditional Network (CNet) to transform sequence cue vectors and original node representation Dynamic association is performed, and the most relevant prior knowledge of user behavior preferences is extracted using a reweighting mechanism to generate a reweighted sequence cue representation. ;

[0021] Capturing Sequence Cue Representations with Cross-Attention and original node representation The correlation between the data and the generated user sequence-enhanced embedding representations is used to determine the relationship between the data and the data. ;

[0022] A self-attention mechanism is used as the readout layer to further extract the enhanced embedding representation of the user sequence. Fine-grained features are used to generate user history sequence representations. .

[0023] Furthermore, the aggregation of behavior-aware interest point maps using message functions... The bidirectional edge information of each interest point is used, and the input representation of the interest point node embedding is updated by combining the time interval embedding matrix and the spatial interval embedding matrix. The specific method is as follows:

[0024] For users Visit any point of interest User behavior-based interest point map Determine points of interest Nearby points of interest and Using message functions from points of interest Collect bidirectional edge information at both the incoming and outgoing edges. And generate points of interest respectively. Nearby Points of Interest and Interactive representation;

[0025] Utilizing users The target trajectory sequence generates a spatial interval matrix and a time interval matrix. The spatial interval matrix and the time interval matrix are truncated by setting a maximum threshold parameter.

[0026] Construct a trainable spatial interval embedding matrix and a time interval embedding matrix. Based on the discrete values ​​obtained after thresholding each element in the spatial interval matrix and the time interval matrix, use the interval mapping operation to extract the interval embedding from the spatial interval embedding matrix and the time interval embedding matrix.

[0027] Update interest points using extracted interval embeddings Nearby Points of Interest and The interaction representation is used to update the input representation of the interest point node embedding, resulting in the updated node representation in the behavior-aware interest point graph, which is then used as the original node representation. .

[0028] Furthermore, the spatially aware hyperbolic graph module maps the spatially aware interest point graph to the hyperbolic space, and after aggregating the node embeddings of neighboring interest points in the tangent space through a multi-layer graph convolutional network, projects them back into the hyperbolic space to generate the specific content of the user spatial preference representation:

[0029] Spatial perception point of interest map Mapped to In a 3D hyperbolic geometric space, Lorentz hyperbolic embeddings are generated for each point of interest.

[0030] The Lorenz hyperbolic embeddings of each point of interest are projected onto the tangent space, and combined with the geographic distance decay mechanism, the node embeddings of each point of interest are updated in the tangent space using a multi-layer graph convolutional network, thereby obtaining the spatial node embeddings of each point of interest.

[0031] All spatial nodes of interest points are embedded and projected back into hyperbolic space to generate a representation of user spatial preferences.

[0032] Furthermore, the specific method for projecting the Lorenz hyperbolic embeddings of each interest point onto the tangent space, and combining the geographic distance decay mechanism, using a multi-layer graph convolutional network to update the node embeddings of each interest point in the tangent space, thereby obtaining the spatial node embeddings of each interest point, is as follows:

[0033] The Lorentz hyperbolic embedding of each interest point is projected into the tangent space, and the initial tangent space vector after the projection of each interest point is used as the input vector of the first layer of graph convolutional network.

[0034] In the first layer of the graph convolutional network, for any point of interest, the node embedding of the point of interest is updated by aggregating the initial tangent space vectors of all neighboring points of interest, in conjunction with the geographical distance decay mechanism, thereby obtaining the output vector of the first layer of the graph convolutional network.

[0035] In the In layered graph convolutional networks, For any point of interest, combining the geographical distance decay mechanism, by aggregating the first... The output vector of the layer graph convolutional network contains the node embeddings of all neighboring interest points of the interest point, combined with the first... The output vector of the layered graph convolutional network is used to update the node features of the interest point, thereby obtaining the first... The output vector of a layered graph convolutional network;

[0036] Average pooling is used to integrate the output vectors of each layer of graph convolutional network to generate spatial node embeddings for each point of interest.

[0037] Furthermore, the specific content of generating user geographic preference representation based on user historical sequence representation and user spatial preference representation using a diffusion preference comparison module, and then generating user ideal location preference based on user geographic preference representation using a diffusion model and contrastive learning, is as follows:

[0038] For any user Acquire users Original node representation As a user Original sequence behavior preferences ;

[0039] Acquiring users User history sequence representation and users User space preference representation A multi-head attention mechanism is used to integrate user historical sequence representations. and user space preference representation Generate user User geographic preference representation ;

[0040] A diffusion model is modeled based on variance-preserving stochastic differential equations, and the forward process of the diffusion model is used to communicate with the user. User geographic preference representation Inject noise to generate a noise location prototype;

[0041] The noise in the noise location prototype is removed using the inverse process of the diffusion model, and the noise is then processed by the user. User history sequence representation and original sequence behavior preferences As a contextual condition, users are generated under different contextual conditions. User's ideal location preference;

[0042] A user contrast loss function is constructed using information contrast loss, and the user contrast loss function is used to correct user behavior under different contextual conditions. The preference deviation between users' ideal location preferences is obtained from the user's User's ideal location preference.

[0043] Furthermore, the diffusion model based on variance-preserving stochastic differential equations utilizes the forward process of the diffusion model to provide information to the user. User geographic preference representation The specific details of injecting noise and generating noise location prototypes are as follows:

[0044] For any point of interest Sampling points of interest Spatial node embedding and as time raw data From time arrive A continuous diffusion process begins, and at each time step, the current time is propagated through a variance-preserving stochastic differential equation. Noise is injected into the original data to generate the current time. The perturbation data is used as the raw data for the next time step. When the diffusion process ends, the time step is... The perturbation data is used as a prototype of the noise location. And there are .

[0045] Furthermore, the generation of users guided by different contextual conditions The specific method for determining users' ideal location preferences is as follows:

[0046] Noise in the noise location prototype is removed through a reverse process based on variance-preserving stochastic differential equations;

[0047] In the noise reduction process, respectively based on user User history sequence representation and original sequence behavior preferences As a contextual condition, a score-based neural network is used to approximate the marginal probability of the score function in the reverse process, thereby obtaining the approximate score function under different contextual conditions;

[0048] The approximate score function under different context conditions is optimized by Fischer divergence, and the optimization objective function under different context conditions is constructed.

[0049] Based on the objective function of the score function under different context conditions, define the total loss function for optimizing the score function;

[0050] The denoising process is optimized using the objective function of the scoring function under different context conditions, generating user profiles under different context conditions. Users' ideal location preferences for points of interest.

[0051] Furthermore, the specific content of generating user interest point prediction results by jointly using user historical sequence representation and user spatial preference representation, through the joint analysis of user ideal location preferences, is as follows:

[0052] For any point of interest According to the user Determine the target trajectory sequence for the user Points of interest The actual access tags;

[0053] By utilizing user historical sequence representations and user spatial preference representations, user ideal location preferences are calculated. Visit points of interest The probability of;

[0054] Based on user Points of interest Real access tags and users Visit points of interest The probability is used to construct a supervised cross-entropy loss function to measure the difference between the probability of a user accessing a point of interest and the actual access tag;

[0055] By utilizing the total loss function of the optimized scoring function, the user contrast loss function, and the supervised cross-entropy loss function, a multi-factor driven objective function for recommending ideal location preference points of interest is constructed. By minimizing the multi-factor driven objective function for recommending ideal location preference points of interest, the optimized behavior-aware graph cueing module, spatially-aware hyperbolic graph module, and diffusion preference contrast module are obtained.

[0056] Based on the optimized behavior-aware graph cues module, spatially-aware hyperbolic graph module, and diffusion preference comparison module, user... The probability of accessing each point of interest and as a user Interest point prediction results.

[0057] The beneficial effects of adopting the above technical solution are as follows:

[0058] Existing technologies, by uniformly processing access factors in user behavior, often lack tailored modeling of both subjective and non-subjective factors, thus limiting the ability to understand dynamic user behavior. This invention, however, designs a behavior-aware graph cueing module. This module utilizes bidirectional information from user behavior sequences (i.e., user target trajectory sequences) and designs a preference filter to dynamically adjust user behavior sequences to tailor modeling of non-subjective user access factors, thereby capturing user behavior sequence representations. Furthermore, this invention considers spatiotemporal context and designs a spatially aware hyperbolic graph module that captures high-order and non-linear geographic associations between points of interest to tailor modeling subjective user access factors from multiple perspectives, capturing user spatial preference representations. Finally, this invention uses user behavior sequence representations and user spatial preference representations to capture users' historical geographic preferences, thereby improving the accuracy of user point-of-interest recommendations.

[0059] Existing methods neglect users' ideal intentions, leading to inaccurate identification of users' true preferences. This invention addresses this by designing a diffusion preference comparison module. This module employs diffusion sampling to generate customized ideal location preferences for the target user, guided by different perspectives of the user's historical behavioral preferences. Through contrastive learning, it corrects preference biases in these ideal location preferences, extracting consistent ideal location preferences across different perspectives. Furthermore, by integrating users' historical preferences, it more accurately captures their true behavioral preferences, thereby further improving the accuracy of user interest point recommendations. Attached Figure Description

[0060] Figure 1 This is a flowchart of an ideal location preference point of interest recommendation method based on multi-factor driven principles in this embodiment;

[0061] Figure 2This is a schematic diagram illustrating the operational logic of the ideal location preference point of interest recommendation method in this embodiment. Detailed Implementation

[0062] To facilitate understanding of this application, specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. The following embodiments are illustrative of the invention but are not intended to limit its scope. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0063] This implementation addresses the limitations of existing methods in understanding dynamic user behavior paradigms by tailoring modeling to both subjective and subjective user access factors to collaboratively capture users' historical location preferences. Specifically, it first constructs directed behavior-aware interest point graphs and undirected spatially-aware interest point graphs. Then, it designs a behavior-aware graph cueing module, modeling non-subjective user access factors from the dual perspectives of bidirectional sequence patterns and sequence representation preference filters. Simultaneously, it integrates spatiotemporal context and proposes a spatially-aware hyperbolic graph module to overcome the limitations of Euclidean space in representing complex relational structures, modeling users' subjective access preferences. Building upon this, this implementation designs a diffusion preference comparison module to utilize historical location preferences to obtain users' potential ideal location preferences, thereby improving the accuracy of capturing users' true preferences. Specifically, it uses diffusion sampling to generate different ideal location preferences based on different views of users' historical behavior preferences. Finally, it corrects the preference biases of these ideal location preferences through comparative learning, extracting consistent ideal location preference representations across different views and combining them with users' historical preferences to achieve more accurate capture of true behavioral preferences.

[0064] This embodiment presents a method for recommending points of interest based on ideal location preferences driven by multiple factors, such as... Figure 1 As shown, the method includes the following steps:

[0065] By acquiring all points of interest visited by several users and the time of access to each point of interest, a target trajectory sequence for each user is constructed.

[0066] The method for constructing the user's target trajectory sequence is as follows: For the user set Any user in Obtain all the points of interest visited by the user and the time spent visiting each point of interest to construct a user profile. target trajectory sequence , recorded as ,in This represents the total number of points of interest accessed by the user. The index value representing the point of interest, and ; Indicates user In time Visit points of interest Points of interest are represented using geographic coordinate tuples. , recorded as ;in Indicate points of interest longitude; Indicate points of interest Latitude.

[0067] In this implementation, embedding is achieved using contextual information features across four dimensions: user, points of interest, time, and space. All users acquired are from a known, existing user set. , denoted as: All user-accessed interest points are derived from a known set of interest points. , denoted as: .

[0068] Construct behavior-aware interest maps and spatial-aware interest maps based on the user's target trajectory sequence.

[0069] In this embodiment, such as Figure 2 As shown, based on the determined user target trajectory sequence, i.e. the user's check-in trajectory sequence, two interest point maps are constructed from the user's target trajectory sequence, considering the spatial distance between the access sequence and the location, respectively.

[0070] The specific content of constructing behavior-aware interest point maps and spatial-aware interest point maps based on the user's target trajectory sequence includes:

[0071] For any user , will users Each visited point of interest is treated as a vertex, based on the user. Directed edges are established based on the time sequence of visits to each point of interest, thereby allowing users to... Construct a directed graph from the target trajectory sequence and use it as a user Behavioral perception interest point map , denoted as: ;in The set of vertices representing the behavior-aware interest point graph; This represents the set of edges in a behavior-aware interest graph; for users Points of interest for visits and points of interest If the user Visiting points of interest Then I visited the points of interest. Then, in the area representing the point of interest Vertices and representative points of interest There is a directed edge between the vertices. .

[0072] In this embodiment, the behavior-aware interest point map Revealed users A sequence paradigm for accessing various locations, where the vertex set includes all points of interest in the user's target trajectory sequence.

[0073] Based on the target trajectory sequences of all users, all points of interest visited by users are counted; for each counted point of interest, each point of interest is treated as a vertex, and a vertex set of a spatially aware point of interest map is constructed. Define the distance edge Represent any two points of interest and The distances between them are used to construct the edge set of the spatially perceived interest point graph. ;in , and All are points of interest; edge weights are defined. For any two points of interest and Based on the proximity between points of interest, a set of edge weights is constructed for the spatially perceived point of interest graph. ,and ;Utilizing the vertex set of a spatially perceptible point of interest map The set of edges in a spatially perceptive interest point graph The set of edge weights for spatially aware point of interest graphs Constructing an undirected spatially perceptible point of interest map , denoted as: .

[0074] In this embodiment, the order in which users check in effectively reflects changes in their interests. However, certain non-subjective factors in users' complex movement patterns can negatively impact the modeling of their true preferences. For example, attending a friend's wedding or dining out for work are not actions based on subjective choice. To capture the sequential behavior representation of users, this embodiment proposes a behavior-aware graph cueing module. This module models the historical behavior sequence of the target user through a directed behavior-aware interest point graph, considering the impact of non-subjective access factors and spatiotemporal context on the accuracy of capturing user preferences.

[0075] By utilizing a behavior-aware graph cueing module, the node representation of the behavior-aware point of interest graph is enhanced through bidirectional information aggregation and spatiotemporal embedding, and user historical sequence representation is generated by combining cueing learning and attention mechanisms.

[0076] In the embodiment, the behavior-aware graph prompting module considers the non-subjective access factors of the user by aggregating the bidirectional information in the user target trajectory sequence and the design preference filter, and considers the subjective access factors of the user by constructing the time-space interval matrix, so as to capture the user historical sequence representation.

[0077] The behavior-aware graph prompting module enhances the node representation of the behavior-aware interest point graph by bidirectional information aggregation and time-space interval embedding, and the specific content of generating the user historical sequence representation by combining prompt learning and attention mechanism is as follows:

[0078] For the target user , the behavior-aware interest point graph of the user is mapped into a vector representation to obtain an input representation of the interest point node embedding . ; wherein represents the vector representation of the interest point after mapping; represents the number of interest points in the behavior-aware interest point graph.

[0079] In the design of the behavior-aware graph prompting module, the embodiment solves the non-subjective access factors in the user check-in behavior sequence from two angles: (1) the rich bidirectional information in the user behavior sequence, i.e., the target trajectory sequence of the user; and (2) the design of the preference filter to capture more robust user sequence behavior preferences.

[0080] The bidirectional information in the user target trajectory sequence is crucial in the interest point recommendation scenario, which focuses on the locations visited by the user and captures the potential future preferences. Considering the synergy between the historical and future preferences, the influence of the non-subjective access driven by external factors can be reduced. In the embodiment, the message function is used to collect the rich bidirectional information in the user behavior sequence, so as to obtain the node representation after solving the non-subjective access factors in the user check-in behavior sequence.

[0081] The message function is used to aggregate the bidirectional edge information of each interest point in the behavior-aware interest point graph , and the input representation of the interest point node embedding is updated in combination with the time interval embedding matrix and the space interval embedding matrix , so as to obtain the updated node representation in the behavior-aware interest point graph and serve as the original node representation .

[0082] The specific method for aggregating the bidirectional edge information of each interest point in the behavior-aware interest point graph by using the message function, and updating the input representation of the interest point node embedding in combination with the time interval embedding matrix and the space interval embedding matrix is as follows: ​​

[0083] For a user arbitrary points of interest visited , based on the user's behavior perception of interest point graph determine the interest point the neighboring interest point and , using the message function from the incoming edge and outgoing edge of interest point Collect bidirectional edge information , and generate the interaction representation of interest point and neighboring interest point and .

[0084] The method of using the message function to collect bidirectional information from the incoming edge and outgoing edge of interest point is expressed as:

[0085] (1);

[0086] (2);

[0087] (3);

[0088] wherein represents the corresponding node of interest point in the behavior perception interest point graph ; represents the vector representation of node after mapping; is the vector representation of the previous node pointed to by node ; represents the attention weight of the neighboring interest point ; is the vector representation of the next node pointed to by node ; represents the attention weight of the neighboring interest point ; represents the attention weight when node collects bidirectional information from node and ; is a normalization function; and are two learnable projection matrices initialized; represents the connection operation; represents the element product operation; represents the node and node​ Interactive representation; Represents a node and nodes Interactive representation.

[0089] In this implementation, users generally prefer points of interest that are closer to them and are more easily influenced by points of interest with shorter time intervals. Furthermore, the relationship between non-adjacent locations and non-contiguous visits is equally important. For example, if a user visits a shopping mall farther from their home on a Saturday morning and then dines at a restaurant near the mall, there is a significant spatial distance between the user's home and the shopping mall. Regarding time, dining at a restaurant in the evening is not directly related to going to a movie theater near the mall in the afternoon. Instead, it is more closely related to having dinner at a different location the previous evening. Therefore, this implementation, by analyzing the user's access sequence... We construct trainable time interval embeddings and spatial interval embeddings to consider subjective access factors in user access behavior.

[0090] Utilizing users The target trajectory sequence generates a spatial interval matrix and a time interval matrix. The spatial interval matrix and the time interval matrix are truncated by setting a maximum threshold parameter.

[0091] The spatial interval matrix and the time interval matrix are represented as follows:

[0092] (4);

[0093] in Represents the spatial interval matrix; Represents the time interval matrix; The first in the spatial interval matrix Line 1 The elements of the column are used to represent users. Points of interest for visits and The relative physical distance between them, and have , express and The distance between them; It is the result of normalization and rounding down to the nearest discrete integer, representing the user's... target trajectory sequence The minimum transfer distance in; This represents the floor function; similarly, Used to represent the minimum transition time in the access sequence, and ; Represents the target trajectory sequence The minimum transition time in the process.

[0094] The process of threshold truncation of the spatial interval matrix and the time interval matrix by setting a maximum threshold parameter is expressed as follows:

[0095] (5);

[0096] in express Discrete values ​​after threshold truncation; This is the maximum threshold parameter for the spatial interval; express Discrete values ​​after threshold truncation; This is the maximum threshold parameter for the time interval; in this embodiment... and All were fixed at 256, when the calculated interval value or When the value exceeds the corresponding maximum threshold parameter, it is limited to the maximum allowed value.

[0097] Construct trainable spatial interval embedding matrices and time interval embedding matrices. Based on the discrete values ​​obtained after thresholding each element in the spatial interval matrix and time interval matrix, extract the interval embedding from the spatial interval embedding matrix and time interval embedding matrix using an interval mapping operation, as follows:

[0098] (6);

[0099] (7);

[0100] in Represents the spatial interval embedding matrix; Represents the time interval embedding matrix; It is a hidden dimension; and Let represent the interval embedding of the spatial interval embedding matrix and the interval embedding of the time interval embedding matrix, respectively. ; Indicates the extraction of spatial interval embedding matrix The embedding vector after all column elements in the row; Indicates the extraction time interval embedding matrix The embedding vector after all column elements in the row.

[0101] The interval vector can be obtained from these embedding matrices through the interval embedding matrix. In this embodiment, the interval embedding of any interval matrix is ​​obtained through the interval mapping operation.

[0102] Update interest points using extracted interval embeddings Nearby Points of Interest and the interaction representation of the interest point node, and then update the input representation of the interest point node to obtain the updated node representation in the behavior-aware interest point graph , and , for representing the output after integrating bidirectional sequence information and considering the spatio-temporal context to update the node representation in the behavior-aware interest point graph, is the updated vector representation.

[0103] The interest point is updated by using the extracted interval embedding and the interaction representation of the adjacent interest point .

[0104] (8) ;

[0105] (9) ;

[0106] wherein and are interval embeddings, and .

[0107] Inspired by the success of prompt learning in static graphs, the embodiment first solves the problem of non-subjective access noise in the interest point recommendation field by using a prompt learning method to enhance the user behavior sequence representation. A learnable vector is defined as a sequence prompt, and a conditional network CNet is used to combine the historical user behavior to ensure that the differences are differentiated while the original sequence preferences of the user are retained. The most relevant prior knowledge is extracted from the user's behavior preferences to generate a sequence prompt representation with a reweighting mechanism. As shown in Figure 2 , the embodiment dynamically adjusts according to the complete behavior sequence of the user, sets a preference filter for extracting the most relevant prior knowledge from the user's behavior preferences, and dynamically captures the deep correlation between the sequence prompt representation and the original node representation through a cross-attention mechanism, so as to more accurately capture the subtle differences in user behavior dependence.

[0108] The sequence prompt vector is initialized, the sequence prompt vector and the original node representation are dynamically associated by using the conditional network CNet, and the most relevant prior knowledge in the user's behavior preferences is extracted by using the reweighting mechanism to generate a sequence prompt representation with a reweighting mechanism .

[0109] The sequence prompt representation is represented as:

[0110] (10);

[0111] where CNet is a multi-layer perceptron; as a sequence prompt vector, is a learnable vector and follows a uniform distribution, and The dimension of and the original node representation The dimension of and the original node representation can be regarded as a reweighting mechanism for extracting the most relevant prior knowledge from the user's behavior preferences.

[0112] The cross-attention is used to capture the relevance between the sequence prompt representation and the original node representation to generate the enhanced embedding representation of the user sequence .

[0113] In this embodiment, the cross-attention is used to capture the relevance between the sequence prompt representation and the original node representation to further solve the non-subjective access factors in the user check-in behavior sequence and enhance the representation ability of the original node representation .

[0114] The enhanced embedding representation of the user sequence is represented as:

[0115] (11);

[0116] where , represents the embedding vector of the interest point after sequence enhancement; is a normalization function.

[0117] In this embodiment, as shown in Figure 2 , the preference filter is used to extract the most relevant prior knowledge from the user's behavior preferences, and the cross-attention is used to capture the relevance between the sequence prompt representation and the original node representation to further solve the non-subjective access factors in the user check-in behavior sequence to obtain the enhanced embedding representation of the user sequence .

[0118] The self-attention mechanism is used as the readout layer to further extract the fine-grained features in the enhanced embedding representation of the user sequence to generate the user history sequence representation .

[0119] In this embodiment, as shown in Figure 2As shown, a self-attention mechanism is used as the readout layer to model user behavior sequence representations at a finer granularity. The user history sequence representation... The calculation method is as follows:

[0120] (12);

[0121] (13);

[0122] in This represents the self-attention mechanism; All are trainable vector matrices; , and These are the Query, Key, and Value matrices, respectively.

[0123] The spatially aware hyperbolic graph module is used to map the spatially aware interest point graph to the hyperbolic space. Then, after aggregating the node embeddings of neighboring interest points in the tangent space through a multi-layer graph convolutional network, the graph is projected back into the hyperbolic space to generate a user spatial preference representation.

[0124] In this embodiment, a spatially aware hyperbolic graph module maps a spatially aware point of interest map onto a hyperbolic space, taking into account higher-order geographical influences to model subjective access factors during user access, thereby capturing the user's spatial preference representation.

[0125] The spatially aware hyperbolic graph module maps the spatially aware interest point map to the hyperbolic space, and after aggregating the node embeddings of neighboring interest points in the tangent space through a multi-layer graph convolutional network, projects them back into the hyperbolic space to generate the specific content of the user spatial preference representation:

[0126] Spatial perception point of interest map Mapped to In a 3D hyperbolic geometric space, Lorentz hyperbolic embeddings for each point of interest are generated.

[0127] The complex geographical relationships between points of interest are crucial for understanding user preferences. Therefore, this implementation focuses on modeling high-order and nonlinear geographical dynamics in the spatially aware hyperbolic graph module. This allows for the capture of distance-based geographical representations to account for the impact of subjective access factors on the accuracy of capturing user preferences. A spatially aware point of interest map constructed based on physical distance is presented. This reflects the spatial relationships between points of global interest. (Due to hyperbolic space) Inspired by the ability to capture complex hierarchical relationships, this implementation method... Mapped to In a hyperbolic geometric space, Lorentz hyperbolic embeddings of each interest point are generated to effectively model the complex high-order relationships between interest points. This mapping process is represented as:

[0128] (14);

[0129] (15);

[0130] (16);

[0131] in Indicate points of interest and The final embedding vector in the Lorentz hyperbolic space is the Lorentz hyperbolic embedding. ; The Lorentz index mapping is used to project embedded vectors in Euclidean space onto hyperbolic space; according to hyperbolic space... The inverse of curvature is used to select a reference point. , ,in and , For hyperbolic space The curvature; based on the reference point Define tangent space ; Representing hyperbolic space At the reference point tangent space The vector in, and ; ; , The index value represents the dimension; the Euclidean embedding vector is located in the tangent space. Sampling in a multivariate Gaussian distribution Perform initialization. and Representing points of interest respectively and In tangent space The feature vector sampled from the multivariate Gaussian distribution on the surface; and Representing points of interest respectively and In tangent space The initial Euclidean embedding vector in the vector.

[0132] The Lorenz hyperbolic embeddings of each point of interest are projected onto the tangent space, and combined with the geographic distance decay mechanism, the node embeddings of each point of interest are updated in the tangent space using a multi-layer graph convolutional network, thereby obtaining the spatial node embeddings of each point of interest.

[0133] In this embodiment, such as Figure 2As shown, the graph convolution network (Graph Convolution Network) aggregates features from neighbor nodes, so that the model can gradually capture global relationships, thereby improving the understanding of complex interaction relationships between points of interest and enriching information representation. Therefore, in the embodiment, a multi-layer graph convolution network is used as a graph encoder to model in tangent space, gradually capture global relationships, enrich information representation, and integrate the output information of each layer by using average pooling operation .

[0134] The specific method for projecting the Lorentz hyperbolic embedding of each point of interest into tangent space and combining the attenuation mechanism of geographical distance to update the node embedding of each point of interest in tangent space by using a multi-layer graph convolution network is as follows:

[0135] The Lorentz hyperbolic embedding of each point of interest is projected into tangent space, and the initial tangent space vector of each point of interest after projection is used as the input vector of the first layer graph convolution network.

[0136] In the first layer graph convolution network, for any point of interest, the node embedding of the point of interest is updated by aggregating the initial tangent space vectors of all neighboring points of interest in combination with the attenuation mechanism of geographical distance, and then the output vector of the first layer graph convolution network is obtained.

[0137] In the first layer graph convolution network, , for any point of interest, the node embedding of the point of interest is updated by aggregating the node embeddings of all neighboring points of interest in the output vector of the first layer graph convolution network in combination with the attenuation mechanism of geographical distance, and then the output vector of the first layer graph convolution network is obtained.

[0138] (17) ;

[0139] (18) ;

[0140] (19) ;

[0141] wherein is a Lorentz logarithmic mapping for projecting the Lorentz hyperbolic embedding into tangent space; are input vectors of the first layer graph convolution network, ; is a point in hyperbolic space, and ; represents the output point of interest of the first layer graph convolution network​ Node characteristics; Indicates the first Interest points output by layer graph convolutional networks Node characteristics; Indicate points of interest The set of adjacent points of interest; Indicate points of interest The set of adjacent points of interest, and having ; yes Convolution kernels in layered graph convolutional networks; The weighting factor in the graph convolutional network represents the distance relationship, enabling the update of node features to incorporate the decay mechanism of geographical distance to fully utilize information from nearby points of interest.

[0142] Mean pooling is used to integrate the output vectors of each layer of graph convolutional networks to generate spatial node embeddings for each interest point.

[0143] All spatial nodes of interest points are embedded and projected back into hyperbolic space to generate a representation of user spatial preferences.

[0144] (20);

[0145] in Indicate points of interest Spatial node embedding; The index value represents the number of layers in the graph convolutional network. The set of spatial node embeddings for all interest points is represented as... The spatial node embedding of points of interest, i.e., geographic representation embedding, is achieved through... Projecting back into hyperbolic space yields a representation of user spatial preferences, which is the final geographic representation.

[0146] Based on user historical sequence representation and user spatial preference representation, a user geographic preference representation is generated using a diffusion preference comparison module. Then, based on the user geographic preference representation, a diffusion model and contrastive learning are used to generate the user's ideal location preference.

[0147] The embodiment designs a diffusion preference contrast module to capture the ideal location preference of the user from the generation perspective and align it with the historical geographic preference of the user, that is, to generate the ideal location preference of the user aligned with the historical location preference of the user modeling subjective and non-subjective access factors. Specifically, by using a score function to approximate the score, while using a variance preserving stochastic differential equation (VP-SDE) framework to represent the sampling process. The diffusion model is a kind of generative model based on probability model, the core idea of which is to gradually convert data into noise through a forward process, and then generate data samples highly similar to the original data distribution from noise through an inverse process, that is, to generate samples through a parameterized denoising process. Finally, a consistent representation of ideal location preference guided by different context conditions is obtained by using contrast learning. The score diffusion modeling process based on stochastic differential equation is as follows:

[0148] In the diffusion model, the forward and reverse processes are usually used in combination to achieve the goal of generative modeling. The forward process describes how the data is affected by random disturbance over time , and the stochastic differential equation (SDE) can be used to describe the continuous change process of the data affected by random disturbance over time , and its solution can be represented as Itô SDE:

[0149] (21);

[0150] where is the diffusion process, used to represent the state of time ; is the drift coefficient, used to represent the deterministic trend of the process; is the diffusion coefficient, which changes over time to control the noise intensity; is the increment of the standard Wiener process, used to introduce random noise.

[0151] The reverse process aims to denoise the data to restore the original data distribution. Its corresponding reverse-time SDE changes over time as follows:

[0152] (twenty two);

[0153] in The gradient term of the logarithmic marginal distribution plays a key role in correcting the drift term to approximate the original distribution, and is used to guide the denoising process so that the model can approximate the original distribution; This is the standard Wiener process. Furthermore, this embodiment also introduces a parameterized scoring neural network. To estimate the gradient term This completes the score matching task in the diffusion process.

[0154] The process of generating user geographic preference representations based on user historical sequence representations and user spatial preference representations using a diffusion preference comparison module, and then generating user ideal location preferences based on user geographic preference representations using a diffusion model and contrastive learning, specifically includes the following:

[0155] For any user Acquire users Original node representation As a user Original sequence behavior preferences .

[0156] Acquiring users User history sequence representation and users User space preference representation A multi-head attention mechanism is used to integrate user historical sequence representations. and user space preference representation Generate user User geographic preference representation .

[0157] In this embodiment, the behavior-aware graph cues module models the relationships between local points of interest, thereby capturing user preferences and obtaining user... User history sequence representation The spatially aware hyperbolic graph module, on the other hand, mines the global potential correlations of user-visited points of interest. Specifically, it uses a spatially aware point of interest graph to generate user data by aggregating information about the user's visited points of interest and their neighboring points of interest. User space preference representation Therefore, this implementation method integrates local and global geographic influences to construct a more refined representation of user geographic preferences. .

[0158] The method employs a multi-head attention mechanism to integrate user historical sequence representations. and user space preference representation Generate user User geographic preference representation The process is represented as:

[0159] (twenty three);

[0160] (twenty four);

[0161] in Indicates the first The output of the size, and ; The number of heads in the multi-head attention mechanism, and in this embodiment ; , and All are trainable matrices. This represents the attention mechanism, and , This indicates normalization processing; It is a weight matrix used to ensure and The dimensions between them are consistent; Represented as user Personalized location prototype vector Initialization, used to calculate user The probability of visiting a point of interest is expressed as:

[0162] (25);

[0163] in Indicates user Visit points of interest The probability of; Indicates in the user Personalized location prototype vector In the middle, points of interest The probability of occurrence, and , The cosine similarity function; Let be the posterior distribution, representing the distribution for a given user. target trajectory sequence Spatial perception point of interest map Under these conditions, users Personalized location prototype vector The probability of occurrence. The goal of this implementation is to obtain the probability of occurrence from the posterior distribution. sampling To calculate .

[0164] A diffusion model is modeled based on variance-preserving stochastic differential equations, and the forward process of the diffusion model is used to communicate with the user. User geographic preference representation Inject noise to generate a noise location prototype.

[0165] For any point of interest Sampling points of interest Spatial node embedding and as time raw data From time arrive A continuous diffusion process begins, and at each time step, the current time is propagated through a variance-preserving stochastic differential equation. Noise is injected into the original data to generate the current time. The perturbation data is used as the raw data for the next time step. When the diffusion process ends, the time step is... The perturbation data is used as a prototype of the noise location. And there are .

[0166] In this embodiment, the forward process of the diffusion model includes: sampling the current point of interest. Position embedding, from time 0 to The continuous diffusion process begins, starting from... Ending , This is a noise location prototype, which is set in this embodiment. The value of is 100, meaning the diffusion process requires 100 time steps. The diffusion process based on stochastic differential equations is modeled using the following formula:

[0167] (26);

[0168] (27);

[0169] in Indicates the current time The disturbance data; Indicates the noise dispatch coefficient; and All are hyperparameters; is a hyperparameter used to represent the total time steps of the diffusion process; It is the increment of the Wiener process at each time step; in this implementation, it is first sent to the target data point. Noise is injected to obtain perturbation data.

[0170] The noise in the noise location prototype is removed using the inverse process of the diffusion model, and the noise is then processed by the user. User history sequence representation and the original sequence behavior preference a user ideal location preference under different context conditions.

[0171] the user ideal location preference under different context conditions.

[0172] Removing noise from the noisy location prototype by the inverse process based on the variance-preserving stochastic differential equation.

[0173] In this embodiment, the user geographic preference representation is denoised by the inverse process based on the variance-preserving stochastic differential equation. The inverse process is both a denoising process and a diffusion process, and the mathematical expression of the inverse stochastic differential equation process is as follows:

[0174] (28) ;

[0175] wherein is the gradient term of the score function, i.e., the logarithmic marginal distribution; is the drift term; is the diffusion term.

[0176] In the denoising process, the user historical sequence representation and the original sequence behavior preference are taken as context conditions, and a score-based neural network is used to approximate the marginal probability of the score function in the inverse process, so as to obtain the approximate score function under different context conditions. The approximate score function under different context conditions is optimized by the Fisher divergence, and the optimization objective function under different context conditions is constructed.

[0177] Since the ideal interest point in the user's consciousness is highly consistent with the interest point that has been checked in, and can reflect the user's preference, the goal of this embodiment is to generate personalized ideal interest points for the user. Therefore, the sampling result should approximate the real interest distribution of the user as much as possible. However, directly calculating the marginal probability of the score function

[0178] is a computationally intensive task. To solve this problem, this embodiment uses a score-based neural network to approximate the marginal probability of the score function , and takes the user historical sequence representation as the context condition, and the approximate score function is expressed as:

[0179] (29) ;

[0180] wherein​​​ representing the user history sequence as the context condition; is a multi-layer perception; representing the concatenation operation; and as the context condition, guiding the denoising process to generate the user's personalized ideal interest point.

[0181] optimizing the score function by Fishe Divergence so that the approximate score function is closer to the actual sample score. Accordingly, when the user history sequence is represented as the context condition, the corresponding optimization objective function is as follows:

[0182] (30);

[0183] wherein is a uniformly sampled time step; represents the expectation of the time step ; represents norm; represents the conditional probability density function with respect to the gradient of

[0184] Through the above process, the embodiment adopts a diffusion-based sampling method to capture the user's ideal interest point distribution, making it as close as possible to the user's actual access interest point distribution, which effectively solves the problem of inherent subjective preference differences in user behavior. In order to improve the sensitivity of the model to subtle differences between samples, the embodiment designs diversified context condition prompts to guide the model to capture the user's geographical preference from different perspectives.

[0185] The embodiment takes the original node representation of the user as the original sequence behavior preference of the user , i.e. the sequence preference without preference filter. Since directly reflects the user's actual access behavior, the original sequence behavior preference of the user is taken as a guide condition to generate a personalized location prototype vector to ensure that the diffusion model can accurately generate samples conforming to the data distribution. At this time, when a score-based neural network is used to approximate the edge probability of the score function , the original sequence behavior preference is taken as the context condition, and the approximate score function is represented as:

[0186] (31);

[0187] wherein denotes the preference for the original sequence as the marginal probability of the score function as the contextual condition.

[0188] Likewise, the score function is optimized by the Fishe Divergence, so that the approximation score function of the score function is closer to the score of the actual sample, and the optimization objective function of the score function under the guidance condition is represented as , which is similar to formula (30), and the corresponding approximation score function is replaced.

[0189] According to the optimization objective function of the score function under different contextual conditions, the total loss function of the optimized score function is defined.

[0190] In view of the importance of the two diffusion enhancement perspectives to the model, the total loss function of the optimized score function is defined as:

[0191] (32);

[0192] The denoising process is optimized by using the optimization objective function of the score function under different contextual conditions, respectively, to generate the user's ideal location preference for the point of interest under different contextual conditions.

[0193] The user contrast loss function is constructed by using the information contrast loss, and the preference deviation between the user's ideal location preferences for the point of interest guided by different contextual conditions is corrected through the user contrast loss function, to obtain the user's ideal location preference.

[0194] The user contrast loss function is represented as:

[0195] (33);

[0196] wherein denotes the batch sample for training; is a temperature parameter, and ; denotes the transpose of the user's ideal location preference for the first point of interest; denotes the user's ideal location preference for the first point of interest guided by the contextual condition.

[0197] ​​​​​​​​​In the embodiment, the information contrast loss InfoNCE is used to establish the contrast learning, and the user contrast loss function aims to maximize the similarity of positive samples, i.e., the similarity between and , and minimize the similarity of negative samples, i.e., the similarity between and .

[0198] The user historical sequence representation and the user spatial preference representation are used to generate the interest point prediction result of the user by combining the ideal location preference of the user.

[0199] The specific content of the user historical sequence representation and the user spatial preference representation is as follows:

[0200] For any interest point , the real access label of the user to the interest point is determined according to the target trajectory sequence of the user .

[0201] The probability of the user accessing the interest point is calculated by combining the ideal location preference of the user .

[0202] In the embodiment, by solving the difference between the non-subjective access noise and the subjective preference, the historical sequence representation of the user and the geographical preference representation of the user are obtained. On this basis, by comprehensively considering the preference of the user, the probability of the user accessing the interest point is modeled , which is represented as:

[0203] (34).

[0204] wherein is a hyperparameter.

[0205] Based on the real access label of the user to the interest point and the probability of the user accessing the interest point , a supervised cross-entropy loss function for measuring the difference between the probability of the user accessing the interest point and the real access label is constructed.

[0206] The supervised cross-entropy loss function is represented as:

[0207] (35);

[0208] wherein represents a supervised cross-entropy loss function; represents a target trajectory sequence of a user ; is a weight of regularization, and ; ; represents a trainable parameter; represents a real access label of a user to a point of interest .

[0209] An ideal location preference point of interest recommendation target function based on multi-factor driving is constructed by using a total loss function of an optimization score function, a user contrast loss function and a supervised cross-entropy loss function, and an optimized behavior perception graph prompt module, a spatial perception hyperbolic graph module and a diffusion preference contrast module are obtained by minimizing the ideal location preference point of interest recommendation target function based on multi-factor driving.

[0210] The ideal location preference point of interest recommendation target function based on multi-factor driving is represented as:

[0211] (36);

[0212] wherein represents an ideal location preference point of interest recommendation target function; and are balance coefficients of different terms in the loss function, and , .

[0213] Based on the optimized behavior perception graph prompt module, the spatial perception hyperbolic graph module and the diffusion preference contrast module, the probabilities of a user visiting each point of interest are calculated and used as the point of interest prediction results of the user .

[0214] In this embodiment, by constructing an ideal location preference point of interest recommendation target function based on multi-factor driving, multi-objective joint optimization of the total loss function of the optimization score function, the user contrast loss function and the supervised cross-entropy loss function is realized. The ideal location preference point of interest recommendation method based on multi-factor driving proposed in this embodiment is trained by using the target function, which is embodied in that the model parameters of the behavior perception graph prompt module, the spatial perception hyperbolic graph module and the diffusion preference contrast module are trained to generate point of interest prediction results closest to the real behavior preferences of a user. The training process is as follows:

[0215] According to the target trajectory sequence of all users, the real access labels of each user to each interest point are determined respectively, and the target trajectory sequence of each user and the real access label of each user to each interest point are taken as a set of sample data, and then a sample set is constructed.

[0216] The training set is taken as input data to start iterative training. In each iteration process, first, the behavior-aware interest point graph and the space-aware interest point graph are constructed according to the input sample set; second, the behavior-aware interest point graph is input into the behavior-aware graph prompt module to generate user historical sequence representation, and the space-aware interest point graph is input into the space-aware hyperbolic graph module to generate user space preference representation; then, the user historical sequence representation and the user space preference representation are input into the diffusion preference comparison module to obtain user geographic preference representation and user ideal location preference; finally, the user historical sequence representation and the user space preference representation are used to calculate the probability of user accessing any interest point through joint user ideal location preference, and a multi-factor driven ideal location preference interest point recommendation target function in the current iteration process is obtained.

[0217] When the multi-factor driven ideal location preference interest point recommendation target function is minimized, the iteration is stopped, and the training optimization of the multi-factor driven ideal location preference interest point recommendation method is completed.

[0218] Based on the generated interest point prediction result, the interest point recommendation is performed for the user.

[0219] In the embodiment, the interest point prediction result output by the multi-factor driven ideal location preference interest point recommendation method is taken as the interest point recommendation for the corresponding user, and the user The probabilities of accessing each interest point are sorted in descending order, and the top K interest points are taken as the recommendation result.

[0220] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the present application.

Claims

1. A method for ideal location preference interest point recommendation based on multi-factor driving, characterized in that, The method comprises the following processes: The method comprises the following processes: According to the target trajectory sequence of the user, a behavior-aware interest point graph and a space-aware interest point graph are constructed; A behavior-aware graph prompt module is used to enhance the node representation of the behavior-aware interest point graph through bidirectional information aggregation and time-space interval embedding, and a user historical sequence representation is generated by combining prompt learning and attention mechanism; A spatially-aware hyperbolic graph module is used to map the spatially-aware interest point graph to a hyperbolic space, and a user spatial preference representation is generated by projecting the node embedding of adjacent interest points in the tangent space back to the hyperbolic space through a multi-layer graph convolution network; the specific content comprises: Spatially-aware point-of-interest map mapping to a four-dimensional hyperbolic geometry space, generating a lorentzian hyperbolic embedding for each point of interest; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The spatial node embedding of all interest points is projected back to the hyperbolic space to generate the user spatial preference representation; Based on the user historical sequence representation and the user spatial preference representation, a user geographical preference representation is generated by using a diffusion preference contrast module, and based on the user geographical preference representation, a user ideal location preference is generated by using a diffusion model and contrast learning; The user historical sequence representation and the user spatial preference representation are used to generate an interest point prediction result of the user by combining the user ideal location preference; Based on the generated interest point prediction result, an interest point recommendation is made for the user. 2.The method of claim 1, wherein, The method for constructing the behavior-aware interest point graph and the space-aware interest point graph according to the target trajectory sequence of the user comprises: For any user , will users Each visited point of interest is treated as a vertex, based on the user. Directed edges are established based on the time sequence of visits to each point of interest, thereby allowing users to... Construct a directed graph from the target trajectory sequence and use it as a user Behavioral perception interest point map For users Points of interest for visits and points of interest If the user Visiting points of interest Then I visited the points of interest. Then, in the area representing the point of interest Vertices and representative points of interest There is a directed edge between the vertices. ; According to the target trajectory sequence of all users, all visited interest points of the users are counted; for all counted interest points, each interest point is taken as a vertex to construct a vertex set of the spatial awareness interest point graph; a distance edge between any two interest points and is defined and an edge set of the spatial awareness interest point graph is constructed; a distance edge weight between any two interest points and based on the proximity between the distances is defined and an edge weight set of the spatial awareness interest point graph is constructed; a spatial awareness interest point graph is constructed by using the vertex set of the spatial awareness interest point graph, the edge set of the spatial awareness interest point graph and the edge weight set of the spatial awareness interest point graph . 3.The method of claim 2, wherein, The specific content of the behavior-aware graph prompt module for enhancing the node representation of the behavior-aware interest point graph through bidirectional information aggregation and time-space interval embedding and generating the user historical sequence representation by combining prompt learning and attention mechanism comprises: For target user map the points of interest in the user's behavior-aware interest point graph into vector representations, obtaining input representations of interest point node embeddings ; Aggregating point-of-interest graph with behavior-aware interest points using message passing The input representation of the point-of-interest node embedding is updated by combining the bidirectional edge information of each point-of-interest in the behavior-aware point-of-interest graph with a time interval embedding matrix and a spatial interval embedding matrix , obtaining the updated node representation in the behavior-aware point-of-interest graph and taking it as the original node representation ; Initialization sequence prompt vector , the sequence prompt vector and the original node representation are dynamically associated by using the conditional network CNet, and the most relevant prior knowledge in the user behavior preference is extracted by using the reweighting mechanism to generate the sequence prompt representation of the reweighting mechanism ; Capturing sequence cue representations with cross-attention and original node representations correlation, generating enhanced embedding representations for user sequences ; The self-attention mechanism is adopted as a reading layer to further extract fine-grained features in the enhanced embedding representation of the user sequence to generate a user history sequence representation . 4.The method of claim 3, wherein, The method comprises the following steps of: acquiring a plurality of behaviors of a plurality of users in a plurality of interest points; and aggregating the behaviors of the users in the interest points by using a message function to obtain an interest point graph. The input representation of the interest point node embedding is updated by combining the bidirectional edge information of each interest point with a time interval embedding matrix and a space interval embedding matrix The specific method is as follows: For users Visit any point of interest User behavior-based interest point map Determine points of interest Nearby points of interest and Using message functions from points of interest Collect bidirectional edge information at both the incoming and outgoing edges. And generate points of interest respectively. Nearby Points of Interest and Interactive representation; Utilizing a user a target trajectory sequence generation space interval matrix and time interval matrix, by setting the maximum threshold parameter to respectively threshold interval matrix and time interval matrix; A trainable spatial interval embedding matrix and a time interval embedding matrix are constructed, and interval embedding is extracted from the spatial interval embedding matrix and the time interval embedding matrix by using interval mapping operation according to the discrete values obtained by threshold truncation of elements in the spatial interval matrix and the time interval matrix; Updating a point of interest with extracted interval embeddings and interactions with neighboring points of interest and of interest, thereby updating the input representation of the point of interest node embedding, obtaining an updated node representation in the behavior-aware point of interest graph and as the original node representation . 5.The method of claim 4, wherein, The specific method for projecting the Lorenz hyperbolic embedding of each interest point to the tangent space, updating the node embedding of each interest point in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, and then obtaining the spatial node embedding of each interest point comprises: The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; In the first layer graph convolution network, for any interest point, the node embedding of all neighboring interest points of the interest point in the output vector of the first layer graph convolution network is aggregated in combination with the decay mechanism of geographical distance, and the node feature of the interest point is updated in combination with the output vector of the first layer graph convolution network, and then the output vector of the first layer graph convolution network is obtained. The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space by using a multi-layer graph convolution network combined with a geographical distance attenuation mechanism, thereby obtaining the spatial node embedding of each interest point; The Lorenz hyperbolic embedding of each interest point is projected to the tangent space, and the node embedding of each interest point is updated in the tangent space 6.The method of claim 5, wherein, The user history sequence representation and the user space preference representation are used to generate a user geographical preference representation by a diffusion preference contrast module, and based on the user geographical preference representation, a diffusion model and contrast learning are used to generate specific content of a user ideal location preference, which is: For any user , obtain the original node representation of the user as the original sequential behavior preference of the user ; Acquiring users User history sequence representation and users User space preference representation A multi-head attention mechanism is used to integrate user historical sequence representations. and user space preference representation Generate user User geographic preference representation ; Modeling a diffusion model based on a variance preserving stochastic differential equation, using a forward process of the diffusion model to a user's geographical preference representation injecting noise, generating noise location prototypes; Using the inverse process of the diffusion model to remove noise from the noise position prototype, and respectively taking the user's history sequence representation and the original sequence behavior preference as the context condition, the user's ideal position preference guided by different context conditions is generated ; A user contrast loss function is constructed using an information contrast loss, and the user contrast loss function is used to correct preference deviation between user ideal position preferences guided by different context conditions to obtain user ideal position preferences of the user . 7.The method of claim 6, wherein, The diffusion model based on the variance preserving type stochastic differential equation models, utilizes the forward process of the diffusion model to the user user geographic preference representation The specific content of injecting noise to generate noise position prototype is that: for any point of interest , sample spatial node embeddings of the point of interest and as temporal raw data , start a continuous diffusion process from time to , at each time step, inject noise into the raw data at the current time by a variance-maintaining stochastic differential equation, generate perturbed data at the current time and as raw data for the next time step, when the diffusion process ends, the perturbed data at time as noise location prototypes , and have . 8.The method of claim 7, wherein, The generated different context condition guided user The specific method of user ideal position preference is: Noise is removed from the noise position prototype by a reverse process based on a variance preserving stochastic differential equation; In the noise reduction process, respectively based on user User history sequence representation and original sequence behavior preferences As a contextual condition, a score-based neural network is used to approximate the marginal probability of the score function in the reverse process, thereby obtaining the approximate score function under different contextual conditions; An approximate score function under different context conditions is optimized by a Fisher divergence, and an optimization objective function under different context conditions is constructed; According to the optimization objective function of the score function under different context conditions, a total loss function of the optimization score function is defined; The denoising process is optimized by using an optimization objective function of the scoring function under different context conditions to generate user User ideal location preferences for points of interest. 9.The method of claim 8, wherein, The specific content of the user history sequence representation and the user space preference representation is used to generate a point of interest prediction result of a user by a joint user ideal location preference, which is: For any point of interest , a real visit label of the user to the point of interest is determined according to a target trajectory sequence of the user ​ The user history sequence and the user spatial preference are used to calculate the user's ideal location preference, and the user's ideal location preference is used to calculate the probability of visiting the point of interest Accessing a point of interest ​ Based on user To a point of interest Real access label and user Access probability of a point of interest Construct a supervised cross-entropy loss function for measuring the difference between the probability of a user accessing a point of interest and the real access label. An ideal location preference point of interest recommendation objective function based on multiple factors is constructed by using the total loss function of the optimization score function, a user contrast loss function and a supervised cross entropy loss function, and the ideal location preference point of interest recommendation objective function based on multiple factors is minimized to obtain an optimized behavior perception graph prompt module, a spatial perception hyperbolic graph module and a diffusion preference contrast module. Based on the optimized behavior perception, the graph prompt module, the spatial perception hyperbolic graph module and the diffusion preference comparison module calculate the user probability of accessing each point of interest as the point of interest prediction result of the user .

Images