Next interest point recommendation method based on multi-relation heterogeneous graph and intent perception

Abstract

This invention belongs to the field of point-of-interest (POI) recommendation and discloses a next-point-of-interest (POI) recommendation method based on multi-relation heterogeneous graphs and intent-awareness. It integrates a multi-relation heterogeneous graph module and an intent-aware tree module, which are used for trajectory representation quality enhancement and user movement intent inference, respectively, jointly improving the accuracy of new POI recommendations. A sample-level loss weighting strategy is proposed, which strengthens the gradient contribution of rare samples to achieve balanced optimization of the model, thereby further improving the recommendation performance of new POIs. The MRHG module uses activity categories as graph nodes and models various heterogeneous relationships between user, POI, and category nodes, thereby capturing rich semantic information in check-in data, enhancing the ability to characterize user travel preferences, and improving the quality of trajectory representation. The IAT module explicitly models the impact of user movement intent on the next POI recommendation to achieve more personalized POI recommendations.

Description

Technical Field

[0001] This invention belongs to the field of interest point recommendation, specifically involving a method for recommending the next interest point based on multi-relationship heterogeneous graphs and intent awareness. Background Technology

[0002] With the widespread adoption of smart mobile devices and the rapid development of Location-Based Social Networks (LBSNs), the way users interact with physical spaces has been profoundly reshaped. Users actively share their Point-of-Interest (POI) visit records through LBSNs, continuously generating massive amounts of mobile check-in data, providing new opportunities for mining human travel patterns and providing personalized services. Against this backdrop, how to effectively learn users' travel preferences from their historical check-in trajectories to achieve high-precision next POI recommendations has become an important research question in the field of personalized location services. However, unlike other sequence recommendation tasks, next POI recommendation not only needs to mine users' travel preferences but also needs to capture their spatiotemporal movement patterns, including temporal periodicity (such as circadian rhythms), spatial constraints (such as distance decay effects), and individual travel habits (such as the "home-office-gym" commuting route).

[0003] To address these characteristics, researchers have widely adopted deep learning techniques to improve the accuracy of next-point-of-purchase (POI) recommendations. Recurrent Neural Networks (RNNs) are effective at modeling the temporal dependencies of user trajectory sequences, but they struggle to capture long-range dependencies and global dependency structure information. To overcome this limitation, models based on Graph Neural Networks (GNNs) and Transformer architectures have emerged in recent years. These models, by introducing graph structure relationships and attention mechanisms, characterize the complex spatiotemporal dependencies and contextual information of user trajectories from a global perspective. While these methods effectively improve the overall performance of next-point-of-purchase recommendation models, several unresolved issues remain.

[0004] First, one of the core goals of location-based services is to recommend the next Point of Interest (POI) that matches a user's travel preferences. Compared to recommending familiar POIs that users have already visited, identifying and recommending new POIs that align with a user's potential interests but are difficult for them to discover autonomously during regular travel patterns is more effective in improving the quality of location-based services. Specifically, users explore a new POI on average every 10 days, and most users explore new POIs at much shorter intervals, with more than half exploring a new POI every 3 days. However, existing POI recommendation methods generally rely on users' historical check-in data, POI co-occurrence frequency, attribute similarity, or geographical proximity to model travel preferences. Among these, RNN-based models (such as DeepMove and Flashback) can effectively capture the temporal sequence dependencies of trajectories, but they often struggle to characterize dynamically changing individual movement patterns. These methods tend to recommend POIs that users have already visited, failing to fully explore potential interests and preferences, thus limiting the effectiveness of new POI recommendations.

[0005] Secondly, users, POIs, and activity categories are the three key entities in LBSNs, interconnected through various heterogeneous relationships that contain rich semantic information. However, existing graph-based modeling methods are mostly limited to dependency modeling using single relationships (such as user-POI or POI-POI), neglecting the role of activity categories in multi-relationship modeling. In most studies, activity categories are only treated as static attributes of POIs to describe their functional type, rather than being explicitly modeled as independent nodes to participate in relationship learning and semantic propagation. This simplification weakens the recommendation model's ability to represent the heterogeneous semantic relationships among the three key entities, thus limiting the completeness of trajectory semantic representation. Therefore, how to treat activity categories as independent graph nodes and leverage GNNs to explore their potential semantic bridging role, thereby promoting the association modeling and semantic transfer between users and POIs, remains to be further explored.

[0006] Furthermore, users' movement intentions play a crucial role in their next POI selection. However, existing recommendation methods generally neglect explicit modeling of movement intentions when generating results, making it difficult to fully reveal the motivations behind user behavior. Classical human mobility theory reveals the statistical regularities of individual movement, indicating that human movement behavior is mainly driven by two basic mechanisms: exploration and prioritizing return. These two mechanisms correspond to two typical movement intentions: exploring new POIs not visited in the historical trajectory, and returning to previously visited familiar POIs. Therefore, introducing movement intentions can explicitly distinguish between users' exploration and return behaviors. Based on this, grouping POIs according to activity category and historical visit records will help narrow the prediction solution space of candidate POIs, thereby improving the accuracy and personalization of POI recommendations. However, users' movement intentions are often influenced by a variety of contextual factors such as spatiotemporal conditions, actual needs, and decision-making psychology. Relying solely on static statistical regularities is insufficient to achieve accurate modeling of user movement intentions. Therefore, there is an urgent need to construct a deep learning model that integrates users' historical trajectory features and dynamic contextual information. Under the theoretical framework of the "exploration-priority return" mechanism, this model can explicitly model and predict users' next movement intentions, thereby deeply exploring users' potential travel patterns. Summary of the Invention

[0007] The purpose of this invention is to provide a next point of interest recommendation method based on multi-relationship heterogeneous graphs and intent awareness, which significantly improves the generalization of next point of interest recommendation and the accuracy of new point of interest recommendation.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A next interest point recommendation method based on multi-relationship heterogeneous graphs and intent awareness includes the following steps:

[0010] Collect user data, point of interest data, and activity category data to construct a multi-relationship heterogeneous graph. Use a multi-relationship heterogeneous graph neural network to learn the node embeddings in the multi-relationship heterogeneous graph. The nodes in the multi-relationship heterogeneous graph include users, points of interest, and activity categories.

[0011] Collect users' historical check-in trajectory and extract the initial embedding of each check-in record in the user's historical check-in trajectory. Use the node embedding fusion in the multi-relationship heterogeneous graph to update the initial embedding of the check-in record. Then, use a sequence encoder to predict the scores of candidate interest points and candidate activity categories.

[0012] Based on the user's historical check-in trajectory and corresponding intent sequence, an intent predictor is used to obtain the probability of the user exploring new activity categories and the probability of exploring new points of interest. A hierarchical weight allocation tree is constructed based on the probability of exploring new activity categories and the probability of exploring new points of interest to obtain the intent perception weight of candidate points of interest and the intent perception weight of candidate activity categories.

[0013] The intention perception weights and scores of candidate interest points are combined to obtain the intention perception scores of candidate interest points. The intention perception weights and scores of candidate activity categories are combined to obtain the intention perception scores of candidate activity categories.

[0014] Based on the intent perception scores of candidate interest points and candidate activity categories, output the recommendation results for the next interest point and the activity category.

[0015] Several alternative methods are provided below, but they are not intended as additional limitations on the overall solution above. They are merely further additions or optimizations. Provided there are no technical or logical contradictions, each alternative method can be combined individually with respect to the overall solution above, or multiple alternative methods can be combined with each other.

[0016] Preferably, the multi-relationship heterogeneous graph includes a node set, an edge set, and a relation matrix set, wherein the relation matrix set includes a user-interest point heterogeneous relation matrix, an interest point-activity category heterogeneous relation matrix, a user-activity category heterogeneous relation matrix, and a user-user relation matrix;

[0017] In the user-interest heterogeneous relationship matrix, the element value corresponding to a user and interest point with an access relationship is 1; otherwise, the element value is 0.

[0018] In the heterogeneous relationship matrix of interest points and activity categories, the element with a corresponding interest point and activity category has a value of 1; otherwise, the element has a value of 0.

[0019] In the user-activity category heterogeneous relationship matrix, for users and points of interest that have an access relationship, the element corresponding to the activity category of the user and the accessed point of interest has a value of 1; otherwise, the element has a value of 0.

[0020] In the user-user relationship matrix, the element corresponding to two users with a social relationship has a value of 1; otherwise, the element has a value of 0.

[0021] Preferably, the multi-relationship heterogeneous graph neural network includes a graph neural network with several stacked layers, each layer of the graph neural network containing a user-interest graph convolutional block, an interest-activity category graph convolutional block, a user-activity category graph attention block, and a user-user graph convolutional block.

[0022] Preferably, the step of updating the initial embedding of the check-in record by fusing node embeddings in a multi-relationship heterogeneous graph includes:

[0023] The initial embedding of the check-in record includes user embedding, point of interest embedding, activity category embedding, intraday time period embedding, intraweek day embedding, and relative geographical location embedding;

[0024] Update the user embedding in the initial embedding by taking the average of the user embedding in the initial embedding and the user node embedding in the multi-relationship heterogeneous graph. Update the interest point embedding in the initial embedding by taking the average of the interest point embedding in the initial embedding and the interest point node embedding in the multi-relationship heterogeneous graph. Update the activity category embedding in the initial embedding by taking the average of the activity category node embedding in the initial embedding and the activity category embedding in the multi-relationship heterogeneous graph.

[0025] Preferably, the score of the candidate interest point and the score of the candidate activity category predicted by the sequence encoder include:

[0026] Add location encoding to the updated check-in records in the user's historical check-in trajectory as the initial input;

[0027] The initial input is passed through a Transformer encoder with several stacked layers to obtain the representation vector of the user's historical check-in trajectory;

[0028] The representation vectors of the user's historical check-in trajectory are input into two fully connected layers to predict the scores of candidate interest points and candidate activity categories, respectively.

[0029] Preferably, the step of using an intent predictor to obtain the probability of a user exploring new activity categories and the probability of exploring new points of interest based on the user's historical check-in trajectory and corresponding intent sequence includes:

[0030] By concatenating the updated embeddings (excluding points of interest), the embedding vectors of category-level intent sequences, and the user's historical activity radius from the check-in records, a category-related feature matrix is ​​obtained.

[0031] By concatenating the updated embeddings (excluding activity categories), the embedding vectors of interest-level intent sequences, and the user's historical activity radius from the sign-in records, an interest-related feature matrix is ​​obtained.

[0032] The category-related feature matrix and the interest-point-related feature matrix are respectively input into a stacked Mamba block, and after attention pooling, the category-level intent-aware trajectory representation vector and the interest-point-level intent-aware trajectory representation vector are obtained respectively. Finally, after passing through a multilayer perceptron, the probability of the user exploring a new activity category and the probability of exploring a new interest point are obtained respectively.

[0033] Preferably, the step of constructing a hierarchical weight allocation tree based on the probability of exploring new activity categories and the probability of exploring new points of interest, to obtain the intent perception weights of candidate points of interest and candidate activity categories, includes:

[0034] A probability enhancement function is used to process the probability of exploring new activity categories and the probability of exploring new points of interest, resulting in the amplified probability of exploring new activity categories and the amplified probability of exploring new points of interest. These are then used as branch conditions for a hierarchical weight allocation tree: the first-level branch condition is to access points of interest belonging to the new activity category with the amplified probability of exploring the new activity category, and to access points of interest belonging to the known activity category with the complementary probability of the amplified probability of exploring the new activity category; the second-level branch condition is to access new points of interest with the amplified probability of exploring the new points of interest, and to access known points of interest with the complementary probability of the amplified probability of exploring the new points of interest.

[0035] For the intermediate nodes obtained from the first-level branch condition decision, the probability value on the path where the intermediate node is located is operated on with the binary mask vector representing whether each activity category is a new activity category to obtain the weight vector of each intermediate node. The weight vectors of each intermediate node are added together to obtain the intention perception weight of all candidate activity categories.

[0036] For the leaf nodes obtained by the second-level branch condition decision, the probability value on the path where the leaf node is located is calculated with the binary mask vector representing whether each interest point belongs to the new activity category and whether it is a new interest point, to obtain the weight vector of each leaf node. The weight vectors of each leaf node are added together to obtain the intention perception weight of all candidate interest points.

[0037] Preferably, the process of fusing the intent perception weights and scores of candidate interest points to obtain the intent perception score of the candidate interest points includes: performing a logarithmic transformation on the intent perception weights of the candidate interest points, and then fusing them with the scores of the candidate interest points to obtain the intent perception score of the candidate interest points.

[0038] The intention perception weights and scores of the candidate activity categories are used to obtain the intention perception score of the candidate activity category, including: adjusting the value range of the intention perception weights of the candidate activity categories to... Then, the scores are combined with the scores of the candidate activity categories to obtain the intention perception score of the candidate activity category.

[0039] As a preferred option, the loss function during training includes a recommendation loss;

[0040] The recommendation loss for a single training sample is obtained by combining the intention-aware score of the candidate interest point and the cross-entropy loss corresponding to the intention-aware score of the candidate activity category;

[0041] The recommendation loss of a single training sample is weighted by the popularity weights of the training samples to obtain the weighted recommendation loss of a single training sample.

[0042] The recommendation loss for each batch is calculated as follows:

[0043]

[0044] In the formula, This represents the recommendation loss corresponding to the batch. Indicates batch size, For training sample index, Indicates the weight of the new point of interest. Indicates the first The weighted recommendation loss for each training sample. As an indicator variable, if the first If the true label of a training sample is a new interest point, the value is 1; otherwise, the value is 0.

[0045] Preferably, the loss function further includes a mobile intent prediction loss, wherein the mobile intent prediction loss for a single training sample is obtained by combining the cross-entropy loss corresponding to the category-level intent-aware trajectory representation vector and the interest-point-level intent-aware trajectory representation vector.

[0046] The present invention provides a next interest point recommendation method based on multi-relationship heterogeneous graphs and intent awareness, which has the following advantages compared with the prior art:

[0047] (1) The Multi-Relational Heterogeneous Graph (MRHG) module and the Intent-Aware Tree (IAT) module are integrated to enhance the quality of trajectory representation and infer user movement intent, respectively. Together, they improve the accuracy of new POI recommendations. In addition, a Sample-Wise Loss Weighting Strategy (SLWS) is proposed to achieve balanced optimization of the model by strengthening the gradient contribution of rare samples, thereby further improving the recommendation performance of new POIs.

[0048] (2) The proposed MRHG module uses activity categories as graph nodes and models various heterogeneous relationships between user, POI, and category nodes to capture rich semantic information in check-in data. This module can enhance the ability to characterize user travel preferences, thereby improving the quality of trajectory representation.

[0049] (3) The proposed IAT module achieves more personalized POI recommendations by explicitly modeling the impact of user movement intentions on the next POI recommendation. This module infers the user's next movement intention in a hierarchical manner, distinguishing between exploring new POIs and returning to known POIs, thereby effectively narrowing the prediction solution space of candidate POIs. Attached Figure Description

[0050] Figure 1 This is a block diagram of a next interest point recommendation method based on multi-relationship heterogeneous graphs and intent awareness according to the present invention.

[0051] Figure 2 This is a flowchart of a next interest point recommendation method based on multi-relationship heterogeneous graphs and intent awareness according to the present invention;

[0052] Figure 3 This is a block diagram of the tree perception module of the present invention;

[0053] Figure 4 This is a schematic diagram of the sample type distribution based on different labels in the experiment of this invention;

[0054] Figure 5 This is a graph showing the analysis results of the top 20 recommendations from six baselines and the SeekNew recommendation model in the experiments of this invention;

[0055] Figure 6 This is a graph showing the comparative analysis results of the overall indicators of the baseline model using the IAT module and the model not using the IAT module in the experiment of this invention;

[0056] Figure 7 This is a graph showing the comparative analysis results of the baseline model using the IAT module and the model not using the IAT module on the new POI-related indicators in the experiment of this invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0059] With the widespread application of location-based social networks, users continuously generate a large amount of mobile check-in data on these platforms, providing a rich source of information for research on personalized travel recommendations. Against this backdrop, next point of interest (POI) recommendation has become a core task in achieving personalized services. However, existing deep learning-based models still face problems such as insufficient performance in new POI recommendations, inadequate utilization of heterogeneous semantic relationships, and incomplete modeling of user movement intentions, making it difficult to comprehensively capture users' potential preferences. To address this, this embodiment proposes a next point of interest recommendation method (considered a SeekNew recommendation model) based on multi-relationship heterogeneous graphs and intent-aware tree modules. By integrating a multi-relationship heterogeneous graph module and an intent-aware tree module, it characterizes users' dynamic travel behavior and potential intentions from multiple levels, thereby improving the generalization of next POI recommendations and the accuracy of new POI recommendations. Specifically, the multi-relationship heterogeneous graph module improves the quality of trajectory representation by modeling various heterogeneous relationships between different entities in LBSNs; the intent-aware tree module effectively narrows the prediction solution space of candidate POIs by hierarchically inferring the user's next movement intention. Furthermore, a sample-level loss weighting strategy is designed to strengthen the gradient contribution of rare samples during training, thereby achieving balanced optimization of the model. This embodiment uses real check-in data from three cities for experimental verification. The results show that the SeekNew recommendation model outperforms existing mainstream baseline models in both overall recommendation performance and new POI recommendation performance.

[0060] The next task for recommending points of interest is: based on the user's historical check-in history. And other contextual information to predict users Next most likely visit One POI. Specifically, let's set... , , These represent the user set, the POI set, and the activity category set, respectively, where each POI... Associated with a specific geographic location, using latitude and longitude coordinates express.

[0061] A user's historical check-in history consists of multiple check-in records, each of which... Defined as a triple , indicating user In time Visited POI Among them, POI Belongs to a specific activity category .user The check-in records are arranged in chronological order, forming a user's historical check-in trajectory. ,in This indicates the length of the user's trajectory. Indicates the first One sign-in record.

[0062] For a given user, previously visited Points of Interest (POIs) are defined as known POIs, while unvisited POIs are defined as new POIs. Activity categories associated with at least one known POI are defined as known categories, while activity categories not associated with any known POIs are defined as new categories. Therefore, the set of activity categories... It can be divided into two subsets: 1) new categories; 2) known categories; POI set It can be divided into three subsets: 1) new POIs in new categories; 2) new POIs in known categories; 3) known POIs.

[0063] In this embodiment, movement intent refers to the potential behavioral tendency of a user when deciding to visit a POI. Movement intent is divided into three categories: 1) Cross-category exploration intent: a tendency to visit POIs in a new category that has not been visited before; 2) Intra-category exploration intent: a tendency to continue exploring new POIs in a familiar category; 3) Revisiting intent: a tendency to return to POIs that have been visited in the past. User historical check-in trajectory. Each check-in record Corresponding to a historical movement intention ,and These correspond to three types of mobile intents. An intent sequence is a sequence formed from a user's historical mobile intents, denoted as... .

[0064] like Figure 1 As shown, the SeekNew recommendation model in this embodiment consists of four key modules: 1) Trajectory learning module: Employs a Transformer-based sequence encoder to learn spatiotemporal features and contextual dependencies from users' historical check-in trajectories. 2) MRHG module: By capturing various heterogeneous relationships between different entities, it constructs a multi-relationship heterogeneous graph based on check-in data and applies an MRHG Neural Network (MRHGNN) to learn high-quality node embeddings with strong representational capabilities, thereby improving the quality of trajectory representation. 3) IAT module: Employs an intent predictor based on Mamba (referencing the paper "Mamba: Linear-time sequence modeling with selective statespaces" published in the proceedings of the First Language Modeling Conference) to infer the probability of different movement intentions and designs a hierarchical weight allocation tree to narrow the prediction solution space of candidate POIs. 4) Prediction and optimization module: Simultaneously calculates the training loss for POI prediction and activity category prediction, and uses the SLWS strategy to achieve a balance between samples labeled as unpopular / new POIs and samples labeled as popular / known POIs.

[0065] like Figure 2 As shown in the figure, this embodiment provides a next interest point recommendation method based on multi-relationship heterogeneous graphs and intent awareness, which includes the following steps:

[0066] Step 1: Collect user data, point of interest data, and activity category data, and construct a multi-relationship heterogeneous graph. Use a multi-relationship heterogeneous graph neural network to learn the node embeddings in the multi-relationship heterogeneous graph. The nodes in the multi-relationship heterogeneous graph include users, points of interest, and activity categories.

[0067] The multi-relationship heterogeneous graph module constructs an MRHG based on check-in data and designs a corresponding MRHGNN. By efficiently aggregating information from different nodes in the graph, it learns high-quality node embeddings with strong representational capabilities, thereby further improving the quality of trajectory representation.

[0068] The constructed MRHG is defined as ,in This represents a set of nodes consisting of users, POIs, and activity categories. Represents the set of edges between nodes. This represents a set of relational matrices. Specifically, It contains four relation matrices, defined as follows:

[0069] (1) User-POI heterogeneous relationship matrix In the formula (1), the element value corresponding to the user and interest point with access relationship is 1; otherwise, the element value is 0.

[0070] (1)

[0071] in, Indicates user With POI The value represents the access relationship between the two entities, with 1 indicating an access relationship and 0 indicating no access relationship. Indicates the size of the user set. Indicates the size of the POI set.

[0072] (2) POI-Category Heterogeneity Relationship Matrix In the formula (2), the element value corresponding to the interest point and activity category with a belonging relationship is 1; otherwise, the element value is 0.

[0073] (2)

[0074] in, Indicates POI With category The value represents the affiliation relationship between the parties, with 1 indicating the existence of an affiliation relationship and 0 indicating the absence of an affiliation relationship. This indicates the size of the activity category set.

[0075] (3) User-Category Heterogeneous Relationship Matrix In the context of users and points of interest with an access relationship, the element corresponding to the activity category of the user and the accessed point of interest has a value of 1; otherwise, the element has a value of 0. This can be determined based on... The result is obtained directly, as shown in formula (3).

[0076] (3)

[0077] in, Indicates user With category The value of 1 indicates that an access relationship exists, and the value of 0 indicates that no access relationship exists.

[0078] (4) User-user relationship matrix In the formula, the element value is 1 for two users who have a social relationship; otherwise, the element value is 0, as shown in formula (4).

[0079] (4)

[0080] in, Indicates user With users The social relationship between them, with a value of 1 indicating that a social relationship exists and a value of 0 indicating that a social relationship does not exist.

[0081] To obtain To achieve high-quality embedding of heterogeneous nodes, this embodiment designs a stacked... The layered graph neural network MRHGNN contains four GNN blocks in each layer: User-POI graph convolutional block (User-POI GCN, UPGCN), POI-category graph convolutional block (POI-category GCN, PCGCN), User-category graph attention block (User-category GAT, UCGAT), and User-user graph convolutional block (User-user GCN, UUGCN), which jointly realize the aggregation and propagation of node information.

[0082] Compared to traditional GCN models, LightGCN models exhibit superior robustness and lightweight characteristics in the learning process of heterogeneous graph node embeddings. Therefore, in this embodiment, LightGCN operations are used in the UPGCN and PCGCN modules for feature propagation and embedding updates to capture the access relationship between users and POIs, as well as the attribution relationship between POIs and categories, as shown in formulas (5)-(9).

[0083] (5)

[0084] (6)

[0085] (7)

[0086] (8)

[0087] (9)

[0088] in, This represents the LightGCN operation, whose input is the adjacency matrix. and node embedding matrix ; express a diagonal matrix; This represents the adjacency matrix between users and POIs. Represents the user-POI heterogeneous relationship matrix The transpose of the matrix; This represents the adjacency matrix between POIs and their categories. Represents the POI-category heterogeneous relation matrix The transpose of the matrix; , and They represent the first Embedding matrices for user, POI, and category nodes in a layered MRHGNN; This represents a matrix concatenation operation; and Indicates the first Two embedding matrices formed by concatenation in a layered MRHGNN; and Indicates the first The updated user interest point concatenation embedding matrix and interest point activity category concatenation embedding matrix in the layered MRHGNN; and These represent the LightGCN operations in the UPGCN and PCGCN modules, respectively.

[0089] The UCGAT module characterizes the access relationship between users and categories by calculating attention scores, thereby modeling users' potential travel preferences for different categories of POIs, as shown in formulas (10)-(13).

[0090] (10)

[0091] (11)

[0092] (12)

[0093] (13)

[0094] in, Indicates user With category Attention score between; This represents the sigmoid activation function; Represents the attention weight matrix. express The transpose of the matrix; This represents the vector concatenation operation; and They represent the first User and category node embeddings in layered MRHGNN; Indicates dynamic weighting ; A matrix representing the association between users and categories; express The transpose of the matrix; express and In the The embedding matrix formed by concatenation in the MRHGNN layer; Indicates the first The updated user activity category concatenation embedding matrix in the layered MRHGNN; This indicates the LightGCN operation in the UCGAT module.

[0095] Due to the importance of social influence in POI recommendations, the UUGCN module facilitates message passing between socially connected users, as shown in formula (14).

[0096] (14)

[0097] in, Represents the learnable weight matrix; Indicates the first The updated user embedding matrix in the MRHGNN layer.

[0098] The node features are propagated and aggregated through four GNN blocks. The node embedding matrix in the MRHGNN layer is updated and then passed to the next layer of the MRHGNN, as shown in Equation (15).

[0099] (15)

[0100] In the formula, , and They represent the first Embedding matrices for user, POI, and category nodes in a layered MRHGNN.

[0101] Step 2: Collect the user's historical check-in trajectory and extract the initial embedding of each check-in record in the user's historical check-in trajectory. Use the node embedding fusion in the multi-relationship heterogeneous graph to update the initial embedding of the check-in record, and use the sequence encoder to predict the score of the candidate interest point and the score of the candidate activity category.

[0102] The trajectory learning module in this embodiment converts each check-in record in the user's historical check-in trajectory into a low-dimensional embedding vector and uses a sequence encoder to learn the complex sequence dependencies between the check-in records, thereby effectively capturing the user's travel preferences and movement patterns.

[0103] This module first introduces an embedding layer, where the user ID, POI ID, and category ID in each check-in record are transformed into dimensions. The embedding vectors are used to encode the check-in time using two time dimensions: intraday time period (discretized in 24-hour format) and intraweek days (discretized in 7-day format). The discrete encodings for the corresponding intraday time period and the corresponding intraweek days are used based on the check-in time of the check-in record. Furthermore, the relative geographical location of each POI (i.e., its Euclidean distance from the city center) is encoded using a Multilayer Perceptron (MLP). Finally, all the encoded embedding vectors are concatenated to form a complete check-in record representation, as shown in Equation (16).

[0104] (16)

[0105] in, Indicating the first position in the trajectory Embedding of each check-in record (this is the initial embedding); For splicing operations; They represent the first The check-in record includes the user, POI, activity category, time period within the day, number of days within the week, and relative geographical location.

[0106] Then, the node embedding output of MRHGNN is used to improve the representation quality of the trajectory. Specifically, the user embedding in equation (16) POI embedding and activity category embedding Respectively with Figure Embeddings of different nodes are merged to achieve embedding updates, as shown in formula (17).

[0107] (17)

[0108] in, , and These represent the embeddings of user nodes, POI nodes, and category nodes output by MRHGNN, respectively, taken from the final output embedding matrices of user, POI, and category nodes of MRHGNN. The updated user embeddings are then used in conjunction with these matrices. POI embedding and activity category embedding renew The reconstructed and updated check-in embedding is obtained. Subsequent calculations are based on the reconstructed and updated check-in embedding. conduct.

[0109] This embodiment obtains the reconstructed and updated check-in embedding. Subsequently, a Transformer-based sequence encoder was introduced for encoding learning. The sequence encoder first adds positional encoding to the generated check-in embedding sequence to preserve the order information of historical check-ins, and then, based on stacking... The layered Transformer encoder models sequence dependencies as shown in equations (18) and (19).

[0110] (18)

[0111] (19)

[0112] in, This is the initial input; This represents the check-in embedded sequence, by The refactored and updated check-in embedding constitute; Represents a sequence of position-encoded vectors; Indicates input to the first Embedded sequences of layer Transformer encoders Represents the first in the embedded sequence An embedding; This represents a Transformer encoder layer. Due to the widespread adoption of the Transformer architecture, this embodiment will not further elaborate on... For a detailed description, please refer to the paper "Attention is all you need" from the proceedings of the 2017 Conference on Neural Information Processing Systems.

[0113] Subsequently, the Transformer encoder's ( ) ) layer output Average pooling is performed along the sequence dimension to obtain the representation vector of the user's historical check-in trajectory. As shown in formula (20).

[0114] (20)

[0115] Finally, based on the representation vector of the user's historical check-in trajectory The sequence encoder uses two fully connected layers ( and Generate unnormalized predicted scores (POI logits) for POIs and their categories. and category logits ), respectively used to predict the next POI and its activity category, as shown in formula (21).

[0116] (twenty one)

[0117] Step 3: Based on the user's historical check-in trajectory and corresponding intent sequence, use an intent predictor to obtain the probability of the user exploring new activity categories and the probability of exploring new points of interest. Construct a hierarchical weight allocation tree based on the probability of exploring new activity categories and the probability of exploring new points of interest to obtain the intent perception weight of candidate points of interest and the intent perception weight of candidate activity categories. Combine the intent perception weight and score of candidate points of interest to obtain the intent perception score of candidate points of interest. Combine the intent perception weight and score of candidate activity categories to obtain the intent perception score of candidate activity categories.

[0118] To ensure that recommended Points of Interest (POIs) align with users' personalized travel preferences, it is essential to understand the motivations and decision-making tendencies behind their next move. Therefore, explicitly modeling the user's next movement intention is crucial. Based on the theoretical framework of the exploration-priority return mechanism, this embodiment proposes an IAT module that integrates users' historical trajectory features and dynamic contextual information to predict users' next movement intentions. Specifically, this module employs a Mamba-based intention predictor to infer the probabilities of the user's exploration and return behaviors, and designs a hierarchical weight allocation tree to adaptively assign weights to candidate POIs, thereby narrowing the prediction solution space for candidate POIs. This module is described as follows: Figure 3 As shown.

[0119] The input to the Mamba-based intent predictor includes the user's historical check-in history and their intent sequence. The intention sequence can be further decomposed into the following two binary sequences.

[0120] (1) Category-level intent sequence If the sign-in record If the activity category of a POI has not appeared in the user's previous check-in history, then the first category is defined as... Category-level mobile intent Otherwise .

[0121] (2) POI-level intent sequence If the sign-in record If the POI in the check-in history has not appeared before, then the first check-in is defined as... POI-level mobile intent Otherwise .

[0122] As the number of visited locations increases, the probability of an individual exploring new locations decreases with a power law. Therefore, this embodiment extracts several discrete user features, including the user's historical activity radius, the total number of unique activity categories visited, and the total number of unique POIs, to improve the accuracy of predicting movement intentions. The above features are concatenated with the user's historical check-in trajectory and intention sequence information, and then transformed into two independent feature matrices, as shown in formulas (22)-(25).

[0123] (twenty two)

[0124] (twenty three)

[0125] (twenty four)

[0126] (25)

[0127] in, and The concatenated feature vectors represent the th... The category-related information and POI-related information for each check-in record; and They represent the first The embedding vectors of category-level intent and POI-level mobile intent for each check-in record are obtained from the embedding layer; and Representing users respectively Number of activity categories and POIs visited in the past; Indicates user The radius of historical activity; and These represent the category-related feature matrix and the POI-related feature matrix, respectively. and Represents the learnable weight matrix; and Indicates the bias term; This indicates the dimension of the concatenated feature vectors. This indicates the input dimension of the subsequent Mamba block.

[0128] Subsequently, because Mamba's selective state update mechanism can flexibly fuse discrete and continuous features, and efficiently capture contextual information with low time complexity, the IAT module adopts... Each Mamba block serves as the encoder for the intent predictor, as shown in equations (26)-(29).

[0129] (26)

[0130] (27)

[0131] (28)

[0132] (29)

[0133] in, The input feature matrix (is) or ); This indicates that the layer normalization operation has been performed. The processed feature matrix; This indicates a one-dimensional convolution along the sequence dimension. The processed feature matrix; This indicates the state-space model. The processed feature matrix; Indicates connection via residual The output feature matrix (as well as) corresponding or with corresponding ).

[0134] In addition, to dynamically focus on the user's main historical movement patterns, the IAT module applies attention pooling operations to the category-related feature matrix output by the Mamba block. Feature matrix related to POI Thus, the category-level intent-aware trajectory representation vector and the POI-level intent-aware trajectory representation vector are obtained respectively. and As shown in formulas (30) and (31).

[0135] (30)

[0136] (31)

[0137] in, This represents the attention pooling operation, whose input is a learnable weight matrix. and characteristic matrix ; Representation of the characteristic matrix The transpose of the matrix; This represents the normalized exponential function; and This represents the learnable weight matrix corresponding to the attention pooling operation, used to generate... and .

[0138] Finally, through the multilayer perceptron corresponding to the category and the POI (for... and )right and Feature processing is performed separately to infer the probability of the user's next category level and POI level mobile intent, as shown in Equation (32).

[0139] (32)

[0140] in, and These represent the probability of a user exploring a new activity category and the probability of exploring a new POI, respectively.

[0141] To improve the predictive model's ability to discriminate movement intentions, a probability augmentation function is used. Applied to and This is used to amplify the difference between high-confidence probabilities and low-confidence probabilities. For a given probability... The amplification process is shown in formula (33).

[0142] (33)

[0143] in, Represents a symbolic function; This indicates an adjustable index, used to control the degree of amplification.

[0144] A hierarchical weighted allocation tree can conceptually be viewed as a binary decision tree with a two-level structure, utilizing amplified probabilities. and Constructed as a branch condition.

[0145] (1) First-level branch: Category-level movement intent. The first-level category branch operation determines the subsequent access... The probability of accessing a POI belonging to a new category, or based on... The probability of accessing a POI belonging to a known category.

[0146] (2) Second-level branching: POI-level movement intent. Within each category branch, the second-level POI branching operation determines whether subsequent visits explore new POIs or return to known POIs, where the probability... Indicates an exploratory tendency. It indicates a tendency to return.

[0147] (3) Leaf nodes: Intent-aware weights. Each leaf node is associated with a specific subset of candidate POIs, and its weight is calculated by combining the probability of the user's next category level and the POI level mobile intent.

[0148] Subsequently, the model calculates the weight of each subset. Specifically, for the leaf nodes obtained by the second-level branch conditional decision, the probability value on the path where the leaf node is located is calculated with the binary mask vector representing whether each interest point belongs to the new activity category and whether it is a new interest point, to obtain the weight vector of each leaf node. The weight vectors of each leaf node are added together to obtain the intention perception weight of all candidate interest points, as shown in formulas (34)-(36).

[0149] (34)

[0150] (35)

[0151] (36)

[0152] in, The binary mask vectors are used to represent whether the corresponding candidate POI belongs to one of the three subsets of the leaf node (i.e., the new POI in the new category, the new POI in the known category, and the known POI in the known category). The length of the three binary mask vectors is the total number of candidate POIs. If the corresponding POI belongs to a certain category, the element value in the corresponding binary mask vector is 1, and the element value in other binary mask vectors is 0. That is, the element value of 1 indicates that the condition is true, and 0 indicates that the condition is false. Let be the weight vector, whose element values ​​represent the intent-aware weights for exploring new POIs from new categories, exploring new POIs from known categories, and returning known POIs from known categories, respectively. Subsequently, the intent-aware weights of the candidate POIs can be obtained, as shown in formula (37).

[0153] (37)

[0154] in, Let be the weight vector, representing the intent-aware weights of all candidate POIs.

[0155] Based on the aforementioned binary decision tree, candidate POIs are divided into specific subsets, each assigned an adaptive intent-aware weight to guide the subsequent re-ranking process. During this process, the calculated weight vector... First, perform a logarithmic transformation. Then logits with POI The two are integrated to achieve intention-aware reordering of candidate POIs in the logarithmic space, as shown in Equation (38).

[0156] (38)

[0157] in, This represents the intent-aware logits of candidate POIs; This represents a constant used to ensure numerical stability.

[0158] Furthermore, to fully utilize the hierarchical semantic relationships in the tree structure, this embodiment is based on the probability of the user's next-level category mobile intent. Candidate activity categories are reordered to generate category logits with intent awareness capabilities, providing richer supervision signals for the next POI recommendation. Specifically, for the intermediate nodes obtained from the first-level branch conditional decision, the probability value on the path where the intermediate node is located is calculated with the binary mask vector representing whether each activity category is a new activity category, to obtain the weight vector of each intermediate node. The weight vectors of each intermediate node are added together to obtain the intent awareness weight of all candidate activity categories, as shown in formulas (39)-(42).

[0159] (39)

[0160] (40)

[0161] (41)

[0162] (42)

[0163] in, This is a binary mask vector, indicating whether each activity category is a new category. Elements that are new categories have a value of 1, and those that are new categories have a value of 0. and This is a weight vector, whose elements represent the intent-aware weights assigned to exploring new categories and returning to known categories, respectively. Let be the weight vector, representing the intent-aware weights of all candidate activity categories; by... The range of values ​​is from Adjust to Used to adjust category logits This allows us to obtain the intent-aware logits of candidate activity categories. .

[0164] The hierarchical weight allocation tree proposed in this embodiment effectively reduces the prediction solution space of candidate POIs by modeling user movement intentions at the category level and POI level, thereby improving the diversity of recommendations.

[0165] Step 5: Based on the intent perception scores of candidate interest points and candidate activity categories, output the recommendation results for the next interest point and the activity category.

[0166] After obtaining the scores, they can be sorted from highest to lowest, and the top scores can be taken. The final prediction result consists of candidate points of interest and candidate activity categories, with the selected points of interest and activity categories corresponding in a sorted order. Alternatively, in other embodiments, the prediction results can be filtered based on scores according to a predefined filtering method.

[0167] In addition to outputting the prediction results, the prediction and optimization module of this embodiment is also responsible for calculating the loss and updating all parameters to be learned during the training process. During the training process, this embodiment uses cross-entropy as the loss function for POI prediction and activity category prediction. Specifically, the recommendation loss of a single training sample is obtained by combining the cross-entropy loss corresponding to the intent perception score of the candidate interest point and the intent perception score of the candidate activity category, as shown in formula (43).

[0168] (43)

[0169] in, This represents the recommendation loss for a single training sample; Represents the cross-entropy loss function; Indicates the true label of the next POI; Indicates the true label for the next activity category; This represents the loss weight for predicting activity categories.

[0170] In the training data, the number of samples with true labels of unpopular or new POIs is far less than that of samples with popular or known POIs. To achieve balanced optimization of the model, this embodiment proposes a SLWS strategy based on popularity weighting and new POI weight adapter to amplify the gradient contribution of rare samples, as shown in formulas (44)-(46).

[0171] (44)

[0172] (45)

[0173] (46)

[0174] in, Represents the popularity weights of the training samples; This represents the true label of the POI of the training sample. The popularity (in this embodiment, the popularity is the number of times the POI appears in all training samples); This represents the weighted recommendation loss for a single training sample; Indicates batch size; For training sample index; This represents the new POI weight adapter, whose value increases with the number of training rounds, causing the recommendation model to focus on learning samples labeled with the new POI in the later stages of training; As an indicator variable, if the first If the true label of a training sample is a new interest point, the value is 1; otherwise, the value is 0. This indicates the recommended loss to be assessed for this batch.

[0175] Meanwhile, this embodiment uses binary cross-entropy as the loss function for the intent predictor, with a total loss of Recommendation loss And movement intention prediction loss The two components are combined to achieve a balance between the main task of POI recommendation and the auxiliary task of mobile intent prediction, as shown in formulas (47)-(49).

[0176] (47)

[0177] (48)

[0178] (49)

[0179] in, The motion intent prediction loss for a single training sample is obtained by combining the cross-entropy loss corresponding to the category-level intent-aware trajectory representation vector and the interest-point-level intent-aware trajectory representation vector. Represents the binary cross-entropy loss function; The true label indicating the next POI level movement intention (marking whether it is exploring a new POI or returning to a known POI); The true label indicating the next category-level movement intention (marking whether the visit is to a POI belonging to a new activity category or to a POI belonging to a known activity category); Represents the loss weights for category-level motion intent prediction; Indicates the first Loss for predicting movement intent for each training sample; Indicates the term used to balance the loss. and Dynamic weights.

[0180] experiment:

[0181] This experiment details the experimental setup and performance evaluation results of the SeekNew recommendation model and the baseline model. All experiments were conducted based on real check-in data from three cities.

[0182] (1) Dataset:

[0183] The experiment was based on real check-in data from three cities, derived from a global check-in dataset collected on the Foursquare platform, which contains approximately 22 months of user check-in records (April 2012 to January 2014), as well as users' social friend information.

[0184] To improve data quality and meet model training requirements, the following preprocessing steps were taken in this experiment: 1) Remove POIs with fewer than 5 visits and users with fewer than 50 check-ins; 2) Use a sliding window to divide the complete check-in sequence of each user into multiple data samples, each containing 20 consecutive check-in records; 3) Divide all data samples into training set, validation set and test set in chronological order, with a division ratio of 8:1:1; 4) Remove check-in records of POIs that only appear in the validation set and test set.

[0185] Table 1 shows the statistics of the preprocessed dataset. Figure 4 This shows the distribution of sample types based on different labels in three datasets, where NP, KP, NC, KC, NP-NC, and NP-KC represent samples labeled as new POI, known POI, new category, known category, new POI within the new category, and new POI within the known category, respectively. Figure 4 As can be seen, at least 26.3% of the samples in these datasets are NP-complete (NP), and the sample distribution across the training, validation, and test sets is relatively consistent. This indicates that the datasets used can effectively support the experimental analysis in this paper and can be used to evaluate the model's performance in recommending new POIs. Furthermore, the three datasets also exhibit significant differences, which helps to comprehensively evaluate the model's recommendation performance under different conditions. The key features of the datasets are as follows:

[0186] Table 1. Statistical information of the dataset

[0187]

[0188] (1) The total number of samples in the NYC dataset is significantly less than that in the TKY and IST datasets, but its check-in distribution is the densest (sparsest). The number of check-ins in the TKY and IST datasets is similar, but the number of users is significantly different, and the average number of check-ins per user in the TKY dataset is much higher than that in the IST dataset.

[0189] (2) In terms of the distribution of NP, KP, NC, and KC samples, the NYC and IST datasets are quite similar, while they differ somewhat from TKY. Users in TKY tend to return known POIs. Because each user in TKY has a higher average number of check-ins, as the number of POI visits increases, the probability of users exploring new POIs decreases, thus generating more KP samples.

[0190] (3) There are significant differences in the sample type distribution between the NYC and IST datasets. In the NYC dataset, the NP-NC class accounts for more than 50%, which is higher than that in the IST dataset (41.9%). This indicates that compared to IST, users in NYC are more inclined to visit new POIs in new categories and have a stronger exploratory tendency.

[0191] (2) Baseline model:

[0192] To evaluate the recommendation performance of the SeekNew recommendation model, the following baseline models were selected for comparison:

[0193] STRNN, an RNN-based model, integrates spatiotemporal context information and employs a time- and distance-based transition matrix to improve the accuracy of predicting a user's next destination. DeepMove, an attention-based RNN model designed to predict human movement behavior, captures complex sequence transitions and multi-level periodic features from large amounts of sparse trajectory data. LSTPM, a model for next POI recommendation, combines nonlocal networks and geographically extended RNNs to capture long-term user preferences and model short-term spatiotemporal dependencies, respectively. Flashback, an RNN-based model, searches for historical hidden states through spatiotemporal context to fully utilize historical movement data, thereby improving the accuracy of next location prediction. PLSPL, a model for next POI recommendation, uses an attention mechanism to model long-term preferences and employs a parallel LSTM model to characterize short-term preferences in location and category sequences, while incorporating personalized weight adjustment mechanisms to enhance the model's ability to express both long-term and short-term user preferences. G-Flashback, a graph-based model, combines learned POI transition maps with GCNs to enhance the quality of POI representations and integrates spatiotemporal data and user travel preference data to improve the accuracy of next POI recommendations. GetNext, a graph-augmented Transformer model, effectively captures individual and group movement patterns by utilizing global trajectory flow graphs and spatiotemporal embedding methods, thus accurately predicting the user's next POI. ExNext, a self-explanatory POI recommendation model, incorporates information bottleneck theory to reduce spatiotemporal data redundancy, thereby improving the accuracy and interpretability of recommendations. HGARN, a model for next location prediction, utilizes hierarchical graph attention mechanisms to capture complex dependencies between time, activity, and location, and employs historical confidence labels to balance the importance of historical locations. iPCM, a model for next POI recommendations, achieves personalized spatiotemporal clustering by integrating global historical trajectory data with user feature embeddings, and uses Transformer encoding and MLP decoding mechanisms to improve prediction accuracy.

[0194] To ensure fairness in model comparison, all baseline models follow the parameter configurations provided in their original papers and open-source code.

[0195] (3) Experimental details:

[0196] To evaluate model performance, the experiment used three metrics commonly used in POI recommendation research: Mean Reciprocal Rank (MRR), Recall (R@K), and Normalized Decay Cumulative Gain (N@K). MRR measures the rank of the true label among all candidate POIs; R@K assesses whether the true label appears in the top K POIs recommended by the model; and N@K reflects the ranking quality of the top K POIs. To further evaluate the model's ability to recommend new POIs, the experiment calculated the above metrics separately for NP samples. This paper refers to the metrics calculated based on all samples as the overall metrics (MRR, R@K, and N@K), and the metrics calculated based on NP samples as the new POI-related metrics (N-MRR, NR@K, and NN@K). In the experiment, the K value was set to 10 and 20; higher values ​​for all metrics indicate better recommendation performance.

[0197] The SeekNew recommendation model of this invention is implemented using the PyTorch framework. Its training details and key hyperparameter settings are as follows: the number of training epochs is set to 20, the Adam optimizer is used, and the initial learning rate is set to... And it gradually decays during training. The dimensions of each ID embedded in the check-in record. Dimensions of Mamba block input embedding All are set to 128. (Number of Transformer encoder layers) MRHGNN layer number and Mamba block count The weights are set to 3, 4, and 3 respectively. Loss weights for activity category prediction. Loss weights compared to category-level motion intent prediction All are set to 0.5. Dynamic weights. The initial value is set to 50, which gradually decays to 0 during the training process, so that the model strengthens the learning of the intent predictor in the early stage of training, thereby providing effective support for subsequent POI recommendation optimization.

[0198] (4) Overall performance:

[0199] Tables 2-4 show the overall performance comparison between the SeekNew recommendation model and existing state-of-the-art baseline models, from which the following conclusions can be drawn:

[0200] Table 2 Overall performance of SeekNew and baseline models on the NYC dataset.

[0201]

[0202] Table 3 Overall performance of SeekNew and baseline models on the TKY dataset.

[0203]

[0204] Table 4 Overall performance of SeekNew and baseline models on the IST dataset.

[0205]

[0206] In Tables 2-4, the bolded and underlined values ​​represent the best and second-best performance among all comparison methods, respectively, while the italicized values ​​represent the improvement of SeekNew relative to the best baseline. Based on Tables 2-4, the following conclusions can be drawn:

[0207] (1) On all three datasets, SeekNew outperformed the baseline model in all metrics, indicating that its overall performance was superior. In particular, SeekNew improved by more than 16.10% in the relevant metrics for new POI recommendation (N-MRR, NR@K and NN@K), far exceeding the improvement of the corresponding overall metrics, indicating that it has a stronger ability to recommend new POIs.

[0208] (2) Due to factors such as dataset size, sparsity, and the distribution of different sample types, SeekNew's performance varies across the three datasets. On the NYC dataset, SeekNew shows a greater improvement in overall metrics, but a smaller improvement in metrics related to new POIs. This is because the NYC dataset contains a higher proportion of NP-NC samples, which are more difficult to predict than new POI samples labeled with known categories. SeekNew shows a smaller improvement in overall metrics on the IST dataset because the IST dataset has higher sparsity, making accurate POI recommendations more difficult. However, on the TKY dataset, SeekNew achieves significant improvements in both overall metrics and metrics related to new POIs. This is because TKY has a higher average number of check-ins per user, allowing the model to more effectively capture these users' travel preferences and movement patterns.

[0209] (3) Among the baseline models, Flashback and PLSPL outperform graph-based models (such as G-Flashback, GetNext, and HGARN) on the NYC dataset. However, they perform poorly on the TKY and IST datasets, indicating that while RNN-based models can effectively capture user travel preferences in small datasets, they still have limitations in large-scale scenarios. In contrast, graph-based models capture user travel preferences from a global perspective, thus performing better on large-scale datasets.

[0210] (4) All baseline models showed similar but poor performance on relevant metrics for new POI recommendations. This phenomenon reflects that the baseline models have difficulty effectively modeling the impact of user mobile intent on POI recommendations, leading them to tend to recommend popular or previously visited POIs, rather than accurately predicting new POIs.

[0211] (5) Analysis of recommendation results and module generalization ability:

[0212] The experiment further analyzed the performance of the SeekNew recommendation model and six baseline models (LSTPM, Flashback, PLSPL, ExNext, HGARN, and iPCM) to fully demonstrate the advantages of the SeekNew recommendation model and its IAT module, and to reveal the necessity of modeling user mobile intent.

[0213] (5.1) Analysis of the coverage of new POI recommendations:

[0214] First, the experiment specifically analyzed the differences in the top K POIs recommended by the models. The experiment recorded the recommendation lists generated by six baseline models and the SeekNew recommendation model on the test set. For each test sample, the experiment selected the top 20 POIs from the recommendation list and calculated the following three metrics reflecting the coverage of new POIs: the number of new POIs (New POIs@20), the number of new categories (New Categories@20), and the activity category prediction accuracy (Correct Categories@20).

[0215] like Figure 5 As shown in the first column, SeekNew achieves higher New POIs@20 values ​​on all three datasets than the baseline model, indicating superior performance in recommending new POIs. This result is consistent with the evaluation results of new POI-related metrics. Further analysis is needed. Figure 5 The second and third columns show that although SeekNew did not significantly outperform the baseline model on the New Categories@20 metric, it performed better on the Correct Categories@20 metric. Particularly on the TKY dataset, more than half of the top 20 POIs recommended by SeekNew belonged to the activity categories with the true labels. These results demonstrate that SeekNew possesses superior activity category prediction capabilities and performs significantly well in predicting new POIs, achieving accurate and personalized POI recommendations.

[0216] (5.2) Performance analysis of intent-aware modeling:

[0217] The experiment integrated the IAT module into six baseline models and compared the recommendation performance of the models before and after integrating the IAT module on three overall metrics (R@20, N@20 and MRR) and three new POI-related metrics (NR@20, NN@20 and N-MRR) to evaluate the generalization ability of the IAT module.

[0218] like Figure 6 and Figure 7 As shown, the baseline model using IAT outperforms the baseline model without IAT across all three overall metrics and three new POI-related metrics. The performance difference is even more significant for the three new POI-related metrics. For example, on the NYC dataset, the ExNext model improves the new POI-related metrics by 74.85%, 107.48%, and 107.80%, respectively; while on the IST dataset, the LSTPM model improves the corresponding metrics by 36.09%, 46.84%, and 48.80%, respectively. These results clearly demonstrate the effectiveness and generalization ability of the IAT module, which enhances the new POI recommendation performance by explicitly modeling the user's next move intention.

[0219] This invention addresses several technical bottlenecks in existing methods, including poor recommendation performance for new Points of Interest (POIs), limited utilization of heterogeneous semantic relationships, and insufficient consideration of user mobile intent. By integrating MRHG and IAT modules, SeekNew effectively captures rich semantic information from heterogeneous relationships and fully considers the impact of user mobile intent on POI recommendations, thereby improving the accuracy of the next POI recommendation. Furthermore, the IAT module proposed in this invention can be integrated into various baseline models to enhance their personalized recommendation performance, fully demonstrating the importance of modeling user mobile intent for the next POI recommendation.

[0220] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0221] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A next interest point recommendation method based on multi-relationship heterogeneous graphs and intent awareness, characterized in that, Includes the following steps: Collect user data, point of interest data, and activity category data to construct a multi-relationship heterogeneous graph. Use a multi-relationship heterogeneous graph neural network to learn the node embeddings in the multi-relationship heterogeneous graph. The nodes in the multi-relationship heterogeneous graph include users, points of interest, and activity categories. Collect users' historical check-in trajectory and extract the initial embedding of each check-in record in the user's historical check-in trajectory. Use the node embedding fusion in the multi-relationship heterogeneous graph to update the initial embedding of the check-in record. Then, use a sequence encoder to predict the scores of candidate interest points and candidate activity categories. Based on the user's historical check-in trajectory and corresponding intent sequence, an intent predictor is used to obtain the probability of the user exploring new activity categories and the probability of exploring new points of interest. A hierarchical weight allocation tree is constructed based on the probability of exploring new activity categories and the probability of exploring new points of interest to obtain the intent perception weight of candidate points of interest and the intent perception weight of candidate activity categories. The intention perception weights and scores of candidate interest points are combined to obtain the intention perception scores of candidate interest points. The intention perception weights and scores of candidate activity categories are combined to obtain the intention perception scores of candidate activity categories. Based on the intent perception scores of candidate interest points and candidate activity categories, output the recommendation results for the next interest point and the activity category.

2. The next interest point recommendation method based on multi-relationship heterogeneous graph and intent awareness as described in claim 1, characterized in that, The multi-relationship heterogeneous graph includes a node set, an edge set, and a relation matrix set. The relation matrix set includes a user-interest point heterogeneous relation matrix, an interest point-activity category heterogeneous relation matrix, a user-activity category heterogeneous relation matrix, and a user-user relation matrix. In the user-interest heterogeneous relationship matrix, the element value corresponding to the user and interest point with access relationship is 1; Otherwise, the element value is 0; In the heterogeneous relationship matrix of interest points and activity categories, the element value corresponding to the interest points and activity categories that have a belonging relationship is 1; Otherwise, the element value is 0; In the user-activity category heterogeneous relationship matrix, for users and points of interest that have access relationships, the element value corresponding to the activity category of the user and the accessed point of interest is 1; Otherwise, the element value is 0; In the user-user relationship matrix, the element value corresponding to two users who have a social relationship is 1; Otherwise, the element value is 0.

3. The next interest point recommendation method based on multi-relationship heterogeneous graph and intent awareness according to claim 1, characterized in that, The multi-relationship heterogeneous graph neural network includes several stacked graph neural networks, each of which contains a user-interest graph convolutional block, an interest-activity category graph convolutional block, a user-activity category graph attention block, and a user-user graph convolutional block.

4. The next point of interest recommendation method based on multi-relationship heterogeneous graph and intent awareness according to claim 1, characterized in that, The method of updating the initial embedding of the check-in record by fusing node embeddings in a multi-relationship heterogeneous graph includes: The initial embedding of the check-in record includes user embedding, point of interest embedding, activity category embedding, intraday time period embedding, intraweek day embedding, and relative geographical location embedding; Update the user embedding in the initial embedding by taking the average of the user embedding in the initial embedding and the user node embedding in the multi-relationship heterogeneous graph. Update the interest point embedding in the initial embedding by taking the average of the interest point embedding in the initial embedding and the interest point node embedding in the multi-relationship heterogeneous graph. Update the activity category embedding in the initial embedding by taking the average of the activity category node embedding in the initial embedding and the activity category embedding in the multi-relationship heterogeneous graph.

5. The next point of interest recommendation method based on multi-relationship heterogeneous graph and intent awareness according to claim 1, characterized in that, The scores for candidate interest points and candidate activity categories predicted by the sequence encoder include: Add location encoding to the updated check-in records in the user's historical check-in trajectory as the initial input; The initial input is passed through a Transformer encoder with several stacked layers to obtain the representation vector of the user's historical check-in trajectory; The representation vectors of the user's historical check-in trajectory are input into two fully connected layers to predict the scores of candidate interest points and candidate activity categories, respectively.

6. The next interest point recommendation method based on multi-relationship heterogeneous graph and intent awareness according to claim 1, characterized in that, The process of using an intent predictor based on the user's historical check-in trajectory and corresponding intent sequence to obtain the probability of the user exploring new activity categories and new points of interest includes: By concatenating the updated embeddings (excluding points of interest), the embedding vectors of category-level intent sequences, and the user's historical activity radius from the check-in records, a category-related feature matrix is ​​obtained. By concatenating the updated embeddings (excluding activity categories), the embedding vectors of interest-level intent sequences, and the user's historical activity radius from the sign-in records, an interest-related feature matrix is ​​obtained. The category-related feature matrix and the interest-point-related feature matrix are respectively input into a stacked Mamba block, and after attention pooling, the category-level intent-aware trajectory representation vector and the interest-point-level intent-aware trajectory representation vector are obtained respectively. Finally, after passing through a multilayer perceptron, the probability of the user exploring a new activity category and the probability of exploring a new interest point are obtained respectively.

7. The next interest point recommendation method based on multi-relationship heterogeneous graph and intent awareness according to claim 1, characterized in that, The process of constructing a hierarchical weight allocation tree based on the probability of exploring new activity categories and the probability of exploring new points of interest, to obtain the intent perception weights of candidate points of interest and candidate activity categories, includes: A probability enhancement function is used to process the probability of exploring new activity categories and the probability of exploring new points of interest, resulting in the amplified probability of exploring new activity categories and the amplified probability of exploring new points of interest. These are then used as branch conditions for a hierarchical weight allocation tree: the first-level branch condition is to access points of interest belonging to the new activity category with the amplified probability of exploring the new activity category, and to access points of interest belonging to the known activity category with the complementary probability of the amplified probability of exploring the new activity category; the second-level branch condition is to access new points of interest with the amplified probability of exploring the new points of interest, and to access known points of interest with the complementary probability of the amplified probability of exploring the new points of interest. For the intermediate nodes obtained from the first-level branch condition decision, the probability value on the path where the intermediate node is located is operated on with the binary mask vector representing whether each activity category is a new activity category to obtain the weight vector of each intermediate node. The weight vectors of each intermediate node are added together to obtain the intention perception weight of all candidate activity categories. For the leaf nodes obtained by the second-level branch condition decision, the probability value on the path where the leaf node is located is calculated with the binary mask vector representing whether each interest point belongs to the new activity category and whether it is a new interest point, to obtain the weight vector of each leaf node. The weight vectors of each leaf node are added together to obtain the intention perception weight of all candidate interest points.

8. The next interest point recommendation method based on multi-relationship heterogeneous graph and intent awareness according to claim 1, characterized in that, The process of fusing the intent perception weights and scores of candidate interest points to obtain the intent perception score of the candidate interest points includes: performing a logarithmic transformation on the intent perception weights of the candidate interest points, and then fusing them with the scores of the candidate interest points to obtain the intent perception score of the candidate interest points. The intention perception weights and scores of the candidate activity categories are used to obtain the intention perception score of the candidate activity category, including: adjusting the value range of the intention perception weights of the candidate activity categories to... Then, the scores are combined with the scores of the candidate activity categories to obtain the intention perception score of the candidate activity category.

9. The next interest point recommendation method based on multi-relationship heterogeneous graph and intent awareness according to claim 1, characterized in that, During training, the loss function includes the recommendation loss; The recommendation loss for a single training sample is obtained by combining the intention-aware score of the candidate interest point and the cross-entropy loss corresponding to the intention-aware score of the candidate activity category; The recommendation loss of a single training sample is weighted by the popularity weights of the training samples to obtain the weighted recommendation loss of a single training sample. The recommendation loss for each batch is calculated as follows: In the formula, This represents the recommendation loss corresponding to the batch. Indicates batch size, For training sample index, Indicates the weight of the new point of interest. Indicates the first The weighted recommendation loss for each training sample. As an indicator variable, if the first If the true label of a training sample is a new interest point, the value is 1; otherwise, the value is 0.

10. The next interest point recommendation method based on multi-relationship heterogeneous graph and intent awareness according to claim 9, characterized in that, The loss function also includes a mobile intent prediction loss, wherein the mobile intent prediction loss for a single training sample is obtained by combining the cross-entropy loss corresponding to the category-level intent-aware trajectory representation vector and the interest-point-level intent-aware trajectory representation vector.

Images