Next point of interest recommendation method, device, electronic device, and storage medium

Abstract

Embodiments of the present specification provide a next point of interest recommendation method and device, electronic equipment and medium, wherein the method comprises: establishing a corresponding relationship between a time-anchored user semantic intention and a specific POI access behavior; modeling long-term preference and short-term preference by adaptively determining a cutoff frequency through a frequency spectrum energy accumulation mechanism; enhancing the coupling relationship between the user semantic intention and the specific POI access behavior in the intention-POI entity joint representation, dynamically fusing the long-term preference and the short-term preference to obtain a user preference representation; constructing a user intention flow chart and a global spatial context chart, combining node features and edge features to decompose edge weights into a spatio-temporal structure term and a semantic interaction term, aggregating node neighborhood information through dynamic edge weights between semantic layers to obtain a user intention representation and a global POI representation; mapping the user preference representation and the user intention representation to a shared latent space, enabling the two types of heterogeneous features to jointly represent user behavior from a complementary perspective to obtain a complete user representation; and calculating a recommendation score using the complete user representation and the global POI representation to obtain a next point of interest recommendation result.

Description

Technical Field

[0001] This document relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for recommending the next point of interest. Background Technology

[0002] With the rapid development of mobile internet, smart terminals, and location-based social networks, users generate a large amount of check-in data with time, location, and category attributes in their daily activities such as travel, dining, shopping, entertainment, and commuting. The next point of interest (POI) recommendation task aims to predict the most likely location a user will visit next based on their historical movement trajectory, supporting applications such as personalized navigation, location services, commercial marketing, and smart city management. Existing technologies mainly learn user preferences through sequence modeling and graph structure modeling methods: sequence modeling methods typically use RNNs, LSTMs, or attention mechanisms to model continuous user check-in sequences to capture the temporal dependencies of POI visits; graph structure modeling methods construct user-POI interaction graphs, trajectory transition graphs, or geographic relationship graphs to mine the interaction relationships between users and POIs and the spatial topological associations between POIs.

[0003] However, existing technologies have the following key limitations in user preference modeling: First, in scenarios with sparse check-in data, most methods only model specific POI entities or only abstract category semantics, resulting in an inability to simultaneously and accurately integrate the correspondence between users' deep semantic intent and specific entity behaviors—relying solely on entity IDs easily leads to insufficient semantic generalization ability, while relying solely on category information loses the personalized preferences of users for specific locations. Second, existing methods divide long-term and short-term preferences based on the time proximity of trajectory sessions, extracting short-term preferences from recent trajectories and long-term preferences from historical trajectories, failing to consider the true rhythm of user behavior—recent repetitive behaviors may reflect long-term stable habits, while early sporadic visits may also map short-term random intents, leading to inaccurate identification of long-term and short-term preferences, and making recommendation results susceptible to noise interference. Third, existing graph structure modeling methods treat semantic information as static attributes of nodes or edges, modeling it separately from spatiotemporal topological relationships. With fixed edge weights, they cannot dynamically respond to the evolution of users' semantic intent, causing graph structures to be unable to effectively depict the dynamic changes in the transfer relationships of user interest points under different semantic states, severely restricting the ability to model user behavior trends in complex scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, electronic device, and storage medium for recommending next points of interest, in order to solve the aforementioned problems in the prior art.

[0005] This invention provides a method for recommending the next point of interest, comprising: Based on continuous temporal coding, interest point category information and interest point entity information are fused respectively to obtain user semantic hierarchical sequence and event hierarchical sequence, establish the correspondence between time-anchored user semantic intent and specific POI access behavior, and construct a joint representation of intent-POI entity; The cutoff frequency is adaptively determined through a spectral energy accumulation mechanism. Low-frequency components below or equal to the cutoff frequency are modeled as long-term preferences, while high-frequency components above the cutoff frequency are modeled as short-term preferences. By using a feature update method that guides entity behavior through semantic features, the coupling relationship between user semantic intent and specific POI access behavior in the intent-POI entity joint representation is enhanced. Combined with an energy-controlled gating fusion mechanism, the long-term preferences and the short-term preferences are dynamically fused to obtain a user preference representation. Based on the user's historical check-in trajectory and the geographical adjacency relationship of global POIs, a user intent flow graph and a global spatial context graph are constructed. Combining node features and edge features, the edge weights are decomposed into spatiotemporal structure items and semantic interaction items. By dynamically aggregating node neighborhood information through semantic layer edge weights, the user intent representation and global POI representation are obtained.

[0006] By mapping user preference representations and user intent representations to a shared latent space, the two types of heterogeneous features can jointly represent user behavior from a complementary perspective, resulting in a fused user representation. The user's recommendation score for the candidate points of interest is calculated using the user representation and the global POI representation, and the next point of interest recommendation result is obtained by probability sorting.

[0007] This invention provides a next point of interest recommendation device, comprising: The acquisition module is used to fuse point of interest category information and point of interest entity information based on continuous time encoding, acquire user semantic hierarchical sequence and event hierarchical sequence, establish the correspondence between time-anchored user semantic intent and specific POI access behavior, and construct a joint representation of intent-POI entity; The cutoff frequency module is used to adaptively determine the cutoff frequency through a spectral energy accumulation mechanism, modeling low-frequency components below or equal to the cutoff frequency as long-term preferences, and high-frequency components above the cutoff frequency as short-term preferences. The enhanced fusion module is used to guide the feature update method of entity behavior through semantic features, enhance the coupling relationship between user semantic intent and specific POI access behavior in the intent-POI entity joint representation, and combine the energy-controlled gating fusion mechanism to dynamically fuse the long-term preferences and the short-term preferences to obtain the user preference representation. The calculation update module is used to construct a user intent flow graph and a global spatial context graph by combining the user's historical check-in trajectory and the geographical adjacency relationship of global POIs. Combining node features and edge features, the edge weights of the user preference representation are decomposed into spatiotemporal structure items and semantic interaction items. The node neighborhood information is aggregated through dynamic edge weights between semantic layers to obtain the user intent representation and the global POI representation. The mapping module is used to map user preference representations and user intent representations to a shared latent space, enabling the two types of heterogeneous features to jointly represent user behavior from a complementary perspective, resulting in a fused user representation. The scoring module is used to calculate the user's recommendation score for the candidate points of interest using the user representation and the global POI representation, and to obtain the next point of interest recommendation result through probability sorting.

[0008] This invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-described next point of interest recommendation method.

[0009] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described next point of interest recommendation method.

[0010] This invention employs embodiments that, based on continuous-time coding, fuse point-of-interest (POI) category information and POI entity information to construct time-anchored user semantic hierarchical sequences and event hierarchical sequences. This achieves a precise correspondence between user semantic intent and specific POI access behavior, thereby establishing a joint representation of intent and POI entities. This effectively solves the problem in existing technologies where user preference modeling fails to simultaneously fuse semantic intent and entity behavior. Furthermore, it adaptively determines the cutoff frequency through a spectral energy accumulation mechanism, modeling low-frequency components as long-term preferences and high-frequency components as short-term preferences. This enables dynamic differentiation and modeling of long-term user habits and short-term intents, overcoming the shortcomings of existing methods that rely on the temporal distance of trajectory sessions to distinguish between long-term and short-term user preferences. Subsequently, it introduces a feature update method that guides entity behavior with semantic features, strengthening the coupling relationship between semantic intent and behavioral entities, and combining this with an energy-controlled gating fusion mechanism. This approach achieves adaptive dynamic fusion of long-term and short-term preferences driven by energy distribution, resulting in a more accurate representation of user preferences. Simultaneously, based on users' historical check-in trajectories and the geographical adjacency of global POIs, it constructs a user intent flow graph and a global spatial context graph. Combining node and edge features, edge weights are decomposed into spatiotemporal structure items and semantic interaction items. Through a dynamic edge weight mechanism between semantic layers, the graph structure achieves real-time response and adaptive updates to semantically driven user behavior transfer relationships, effectively aggregating neighborhood information and generating more expressive user intent and global POI representations. Finally, user preference and user intent representations are mapped to a shared latent space, enabling two types of heterogeneous features to collaboratively represent user behavior from complementary behavioral and semantic perspectives, forming a more integrated and generalized user representation that comprehensively improves the accuracy and personalization of next point of interest recommendations. Therefore, this solution systematically solves three core problems in existing technologies: user preference modeling cannot collaboratively integrate semantic intent and entity behavior; the division of long-term and short-term preferences relies too much on the time distance of trajectory sessions; and graph structures cannot dynamically respond to semantically driven changes in relationships. It significantly improves the semantic understanding ability, temporal adaptability, and structural evolution response ability of recommendation systems, achieving more accurate and intelligent recommendations for the next POI. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of the next point of interest recommendation method according to an embodiment of the present invention; Figure 2This is a schematic diagram of the next point of interest recommendation device according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0014] Method Implementation Examples According to embodiments of the present invention, a method for recommending next points of interest is provided, specifically, a method for recommending next points of interest based on time-frequency preference separation and dynamic semantic topology. Figure 1 This is a flowchart of the next point of interest recommendation method according to an embodiment of the present invention, such as... Figure 1 As shown, the next point of interest recommendation method according to an embodiment of the present invention specifically includes: Step S101: Based on continuous time encoding, interest point category information and interest point entity information are fused respectively to obtain user semantic level sequence and event level sequence, establish the correspondence between time-anchored user semantic intent and specific POI access behavior, and construct joint representation of intent-POI entity; In this embodiment, based on continuous temporal encoding, interest point category information and interest point entity information are fused separately to obtain user semantic-level sequences and event-level sequences. A correspondence between time-anchored user semantic intent and specific POI access behavior is established, constructing a joint representation of intent and POI entity. By fusing the embedded representations corresponding to each interest point category and interest point entity with temporal information processed by continuous temporal encoding, two types of sequences are formed: one centered on category information, reflecting the user's access pattern at the semantic intent level; the other centered on specific POI entities, reflecting the user's access choices at the behavioral event level. Temporal encoding ensures that the time point of each record is transformed into a continuous vector, thus preserving the dynamic characteristics of behavior evolution over time without relying on discrete time slices. Both types of sequences use time as an anchor point, enabling precise alignment of semantic intent and entity behavior in the temporal dimension, forming a joint representation system that simultaneously carries user intent and specific choices, providing structured input for subsequent preference modeling.

[0015] Step S102: The cutoff frequency is adaptively determined through the spectrum energy accumulation mechanism. Low-frequency components below or equal to the cutoff frequency are modeled as long-term preferences, and high-frequency components above the cutoff frequency are modeled as short-term preferences. In this embodiment, the cutoff frequency is adaptively determined through a spectral energy accumulation mechanism. Low-frequency components below or equal to the cutoff frequency are modeled as long-term preferences, while high-frequency components above the cutoff frequency are modeled as short-term preferences. This process is based on performing a frequency domain transformation on the user check-in behavior sequence, calculating the energy distribution of each frequency component, and accumulating energy along the frequency axis until the accumulated proportion reaches a preset threshold. This naturally forms a dividing point that dynamically adapts to the rhythm of user behavior. Frequency components below this dividing point are uniformly classified as long-term preferences reflecting stable periodic patterns, while frequency components above this dividing point are identified as short-term preferences reflecting instantaneous fluctuations. This approach does not rely on fixed time windows or manually set thresholds; the frequency domain division of preference types can be achieved solely through the mathematical process of energy accumulation.

[0016] Step S103: Enhance the coupling relationship between user semantic intent and specific POI access behavior in the intent-POI entity joint representation by using the feature update method guided by semantic features, and dynamically fuse the long-term preferences and the short-term preferences in combination with the energy control gating fusion mechanism to obtain the user preference representation. In this embodiment, a cross-layer attention mechanism is used, with semantic features serving as keys and values ​​in the attention process and event features serving as queries. This completes the feature update method that guides entity behavior through semantic features, enhancing the coupling between user semantic intent and specific POI access behavior. Semantic features act as key guiding signals, influencing entity access behavior and causing it to adjust at the representation level driven by semantic intent. Simultaneously, a gating fusion mechanism with energy control is used to dynamically generate gating coefficients based on the difference in signal energy intensity between long-term and short-term preferences at the current moment. This achieves a weighted fusion of long-term and short-term preferences, ultimately resulting in a user preference representation that reflects the co-evolution of intent and behavior.

[0017] Step S104: Based on the user's historical check-in trajectory and the geographical adjacency relationship of global POIs, a user intent flow graph and a global spatial context graph are constructed. Combining node features and edge features, the edge weights are decomposed into spatiotemporal structure items and semantic interaction items. Node neighborhood information is aggregated through dynamic edge weight aggregation between semantic layers to obtain the user intent representation and the global POI representation. In this embodiment, a user intent flow graph and a global spatial context graph are constructed respectively, and the edge weights in the graphs are decomposed into spatiotemporal structure items and semantic interaction items. The spatiotemporal structure item, combined with edge features, describes the prior influence of time intervals and spatial distances on the node association strength. The semantic interaction item dynamically depicts the changing interaction relationships between nodes as the semantic state evolves, based on the node features updated layer by layer during graph convolution. By multiplicatively coupling spatiotemporal structural terms and semantic interaction terms, the edge weights in the graph can be dynamically adjusted according to the semantic state of nodes while preserving objective spatiotemporal topological constraints. This allows for a more accurate representation of the intent flow patterns in user access behavior and the spatial contextual relationships between points of interest. Furthermore, based on the different flow characteristics of the user intent flow graph and the global spatial context graph, two different aggregation methods are designed to aggregate neighborhood information, thereby obtaining user intent representations and global point of interest representations. This provides feature support for subsequent prediction of the next point of interest, taking into account both individual dynamic transfer patterns and global spatial constraints.

[0018] Step S105: Map user preference representation and user intent representation to a shared latent space, so that the two types of heterogeneous features can jointly represent user behavior from a complementary perspective, and obtain the fused user representation.

[0019] In this embodiment, user preference representations and user intent representations are mapped to a shared latent space, enabling these two heterogeneous features to jointly characterize user behavior from a complementary perspective, resulting in a fused user representation. This process uses a nonlinear transformation to map the previously obtained user preference and user intent representations to the same representation space, ensuring comparability and synergy within the same dimension. This achieves a joint characterization of user behavior from two different dimensions: the regularity of behavioral rhythms and the trend of intent flow. The mapped fused representation no longer independently retains the original structures of the two types of features but forms a composite representation. Its content simultaneously carries the time-frequency regularity information extracted by preference separation and the intent flow pattern captured by dynamic graph convolution. This allows the final user representation to comprehensively reflect the stability and dynamism of user behavior, providing a unified and multidimensional input basis for subsequent recommendation score calculation.

[0020] Step S106: Calculate the user's recommendation score for the candidate points of interest using the user representation and the global POI representation, and obtain the next point of interest recommendation result by probability sorting.

[0021] In this embodiment, based on the constructed fused user representation and global POI representation, the system calculates the recommendation scores of users for candidate points of interest and performs probability ranking, directly transforming users' long-term and short-term preferences and user intent flow patterns into executable recommendation decisions. This achieves a closed-loop output from high-dimensional behavioral features to specific point of interest ranking results, solving the technical gap of only including users' long-term and short-term preferences or only including users' flow intentions. This enables the system to accurately predict the points of interest that users are most likely to visit in the next moment based on long-term and short-term preferences that conform to the rhythm of user behavior and combined with the dynamic graph topology evolution results, thereby significantly improving the personalization accuracy and temporal adaptability of recommendations and effectively supporting the need for continuous behavior prediction driven by dynamic intent in complex scenarios.

[0022] By applying the technical solution of this embodiment, continuous-time coding is used to fuse point of interest (POI) categories and entity information to construct a joint representation between time-anchored semantic intent and POI access behavior, accurately achieving collaborative modeling of semantic intent and entity behavior. Furthermore, a spectral energy accumulation mechanism is used to adaptively determine the cutoff frequency, modeling low-frequency and high-frequency components as long-term and short-term preferences respectively, breaking the limitation of existing methods that rely on the temporal distance of trajectory sessions to classify user preferences. Based on this, a feature update method that guides entity behavior through semantic features enhances the coupling between intent and behavior, and energy-controlled gating is implemented. The fusion module dynamically weights long-term and short-term preferences to form a more discriminative user preference representation. Furthermore, it constructs a user intent flow graph and a global spatial context graph, combining node and edge features to decompose edge weights into spatiotemporal structure items and semantic interaction items. Through dynamic edge weight aggregation of node neighborhood information across semantic layers, semantic features are integrated into the dynamic evolution of the spatiotemporal topology. Finally, the user preference representation and the user intent representation obtained from the dynamic graph aggregation are mapped to a shared latent space, enabling these two heterogeneous representations to collaboratively model user behavior from complementary perspectives of behavioral inertia and intent evolution, forming a unified and refined user representation. This comprehensively solves core problems in existing technologies, such as the inaccurate fusion of semantic intent and entity behavior, the over-reliance on the temporal distance of trajectory sessions for long- and short-term preference division, and the lack of dynamic changes in user trajectory transition relationships with semantic state evolution in graph structure modeling methods. This significantly improves the accuracy and personalization level of next point of interest recommendations.

[0023] Step S101 specifically includes: constructing learnable independent embedding matrices for POIs and categories respectively. , ,in This represents the total number of POIs and categories. This represents the POI identifier corresponding to the i-th check-in in the given sequence of embedding dimensions. and category identifier The dense vector representation is as follows: Formula 1; According to Formula 2, time is normalized within the sequence: Formula 2; in It is normalized time. User All check-in times t i The maximum and minimum values ​​in the set. It is a very small constant; According to Formula 3, Time2Vec is used to scalar time. Vectorized encoding is performed to capture both the linear trend and multi-scale periodic patterns of user behavior over time. Formula 3; in, The mapped continuous time vector Indicates the embedding dimension. As learnable parameters, the first dimension represents the non-periodic long-term trend through a linear transformation, while the remaining dimensions use sinusoidal basis functions to capture the periodic dependence at different frequencies. A two-layer coupled representation is constructed using time as an anchor point: the semantic layer, as shown in Equation 4, focuses on capturing the user's semantic intent by concatenating category embeddings with temporal embeddings; the event layer, as shown in Equation 5, focuses on the user's specific entity selection behavior by concatenating POI entity embeddings with temporal embeddings. Through this concatenated representation, the model learns the user's semantic intent and specific access behavior at the corresponding time point. Formula 4; Formula 5; in, This represents a vector concatenation operation; Initial node features for constructing a full graph of nodes within the same embedding space: for any graph node Its corresponding POI identifier and category identifier are The initial characteristics of the nodes are: Formula 6; Step S102 specifically includes: using Fast Fourier Transform to transform the semantic sequence and event sequence Transform from the time domain to the frequency domain: Formula 7; in, For sequence length, It is the hidden layer dimension; Before calculation The cumulative energy percentage of each frequency component, when it exceeds a certain proportion When the frequency is equal to the cutoff frequency, the corresponding frequency is called the cutoff frequency. Set the energy of frequency 0 to 0: Formula 8; in, Indicates the first The and the first Fourier coefficients at discrete frequency positions; Will Normalization to The cutoff frequency is obtained within the interval. Cutoff frequency calculated from semantic sequence The global threshold is also applied to the event sequence. After determining the cutoff frequency threshold, a Butterworth low-pass filter is used to filter the signal. The Butterworth low-pass filter controls the frequency roll-off characteristic through its order, achieving a good balance between cutoff characteristics and smoothness. Formula 9; Where n represents the order of the Butterworth low-pass filter; Indicates the filter at the 1st Amplitude-frequency response coefficient at each frequency component; Construct a filter mask using the Butterworth low-pass filter. Extracting low-frequency long-term components using complementary masks Extracting high-frequency short-term components: Formula 10; in, This indicates the long / short-term preference of semantic sequences in the frequency domain. This indicates the long / short-term preference of the event sequence in the frequency domain. This indicates element-wise multiplication; The decoupled long-term and short-term components are restored back to the time domain using IFFT: Formula 11; in, This indicates the long / short-term preference of semantic sequences in the time domain. This indicates the long / short-term preference of the event sequence in the time domain; After obtaining the decomposed representations of the semantic sequence and the event sequence, the event component is used as the query, and the corresponding semantic component is used as the key and value. The original sequence features are preserved through residual connections. Formula 12; in, It captured long-term preferences driven by users' lifestyle habits. It captures short-term intents triggered by users' immediate needs, where Q represents the query vector, K represents the key vector, and V represents the value vector; Step S103 specifically includes: employing a signal energy control gating mechanism to generate gating coefficients by calculating the logarithmic energy difference between low-frequency and high-frequency signals at the current moment. : Formula 13; in, and They represent The energy intensity of low-frequency and high-frequency components at any given time. For a given energy ratio, These are learnable parameters; Final User Preference Representation Represented by long-term preferences and short-term preference representation The weighted fusion yields: Formula 14.

[0024] Step S104 specifically includes: normalizing the spatiotemporal attributes, and the normalized time interval is: Formula 15 in, This represents the minimum value in the sequence of user check-in time intervals. It is a very small constant; User records The spherical distance between two adjacent POI visits is the spatial interval, and the latitude and longitude of each POI are recorded as follows: The spherical distance is calculated using the Haversine distance formula, and the normalized spatial interval is: Formula 16; in, This represents the minimum non-zero value in the sequence of spatial distances between adjacent points of interest visited by a user. Based on graph node features Define two graph structures: SeqDSE-GCN( ): A node is a unique POI visited in the user's history. A directed edge is established between two consecutively visited POIs to capture the trend of the user's trajectory. The edge feature includes a normalized time interval. and spatial distance , Represents the user intent flow diagram; GeoDSE-GCN ( ): Nodes are all POIs in the dataset. If the distance between two POIs is less than a preset threshold... Then, construct undirected edges, with the edge characteristic being the normalized spatial distance. , Represents a global spatial context graph; for To express the prior influence of spatiotemporal proximity on transition probabilities, we define... For nodes and nodes The spatiotemporal structural edge weights between them are specifically a joint exponential decay function of distance and time: Formula 17; in, This is a learnable attenuation coefficient used to balance the relative importance of time and space factors. Represents a node and nodes The time interval, Represents a node and nodes Spatial interval; definition For the first Layer nodes and nodes Dynamic semantic edge weights between: Formula 18; in, For learnable parameters, For the sigmoid function, The heat coefficient, Indicates the first Layer nodes and nodes Node characteristics; The final edge weight is the product of two terms. ; for Similarly, the construction uses structural terms plus semantic terms. The structural edge weights are initialized with a Gaussian kernel distance decay function. This structure is used to express the spatial prior that the closer the elements are, the stronger the relationship. Formula 19; in, Represents a node and nodes The spherical distance representing the POI; Dynamic semantic edge weights are constructed using a Gaussian kernel function in the feature space: Formula 20; in, The projection matrix is ​​learnable. Softplus is used to ensure positivity; the closer the endpoints are in the semantic space, the closer the semantic term is to 1, and the final edge weight is... ; For user intent flow graph and global spatial context diagram Different flow characteristics for directed graphs Calculate the independent normalization coefficients: Formula 21 in, and They are nodes In the The in-degree normalization factor and out-degree normalization factor of the layer, and Representing the in-degree neighbor set and the out-degree neighbor set, the graph convolution aggregation process is represented as: Formula 22; in, and These are the original node features before and after the update. It is a feature of neighboring nodes. It is an activation function. It is a learnable weight matrix, which is aggregated through graph convolution to ultimately obtain a representation of user intent; For undirected graphs , No. The weighting degree of the layer is The aggregation intensity is adjusted using a symmetric normalization operator. The graph convolution aggregation process is represented as follows: Formula 23; in, The weight matrix is ​​a learnable matrix, after... After layer propagation, a global POI representation is obtained.

[0025] Step S105 specifically includes: In the prediction stage, firstly, a nonlinear fusion strategy is used to map heterogeneous features to a shared latent space: Formula 24; in, Representation layer normalization, This represents the activation function. For learnable parameters, To represent user preferences, To represent user intent, As a complete user representation, it includes both short-term and long-term user preferences and trends in user intent. Through user representation Representation of Global POI The predicted score is calculated to obtain the recommendation result, where, Provided by the global spatial context map encoder, it contains objective geographical constraint information. Predicted score Defined as: Formula 25; in, yes The geographic embedding representation, and the final access probability distribution Obtained through the Softmax function Indicates transpose; To address the sparsity of check-in data and enhance model interpretability, a loss function that incorporates both the primary recommendation task and dual auxiliary constraints is employed. The primary recommendation task utilizes labeled, smoothed cross-entropy loss to suppress overfitting on sparse data and improve robustness to noisy labels. Formula 26; in, To recommend losses, To predict scores, It is a model that predicts POI The probability of access. It is a smoothing factor. It is the number of candidate POIs; To prevent the model from focusing too much on POI IDs and ignoring the underlying semantic intent, a semantic auxiliary loss is introduced: Formula 27; in, It is the total number of POI categories. It is the one-hot tag of the category to which the real POI belongs. These are learnable hyperparameters; To ensure the long-term components obtained from frequency domain decomposition and short-term components In accordance with physical principles, energy constraints are introduced: Formula 28; in, Indicates the sequence at time step The long-term and short-term eigenvectors, These are the balance coefficients. The first term in Formula 28 is a smoothness constraint, ensuring that long-term preferences remain stable over time. The second term in Formula 28 is a recency constraint. The mask vector marks historical time steps far from the current time as 1, which penalizes the energy residue of short-term high-frequency signals in distant history. The model's final objective loss function is a weighted sum of the above components: Formula 29; in, It is a weighting factor used to balance semantic guidance and energy constraint sharing.

[0026] Device Example 1 According to an embodiment of the present invention, a next point of interest recommendation device is provided. Figure 2 This is a schematic diagram of the next point of interest recommendation device according to an embodiment of the present invention, as shown below. Figure 2 As shown, the next point of interest recommendation device according to an embodiment of the present invention specifically includes: The acquisition module 20 is used to fuse point of interest category information and point of interest entity information based on continuous time encoding, acquire user semantic hierarchical sequence and event hierarchical sequence, establish the correspondence between time-anchored user semantic intent and specific POI access behavior, and construct a joint representation of intent-POI entity; The cutoff frequency module 22 is used to adaptively determine the cutoff frequency through a spectrum energy accumulation mechanism, modeling low-frequency components below or equal to the cutoff frequency as long-term preferences, and high-frequency components above the cutoff frequency as short-term preferences. The enhanced fusion module 24 is used to guide the feature update method of entity behavior through semantic features, enhance the coupling relationship between user semantic intent and specific POI access behavior in the joint representation of intent-POI entity, and combine the energy control gating fusion mechanism to dynamically fuse long-term preferences and short-term preferences to obtain user preference representation. The calculation update module 26 is used to construct a user intent flow graph and a global spatial context graph based on the user's historical check-in trajectory and the geographical adjacency relationship of global POIs. Combining node features and edge features, the edge weights are decomposed into spatiotemporal structure items and semantic interaction items. The node neighborhood information is aggregated through dynamic edge weights between semantic layers to obtain the user intent representation and the global POI representation. Mapping module 28 is used to map user preference representation and user intent representation to a shared latent space, so that the two types of heterogeneous features can jointly represent user behavior from a complementary perspective, and obtain a fused user representation. The scoring module 29 is used to calculate the user's recommendation score for the candidate points of interest using the user representation and the global POI representation, and to obtain the next point of interest recommendation result by probability sorting.

[0027] The embodiments of the present invention are device embodiments corresponding to the above method embodiments. The specific operation of each module can be understood with reference to the description of the method embodiments, and will not be repeated here.

[0028] Device Example 2 This invention provides an electronic device, such as... Figure 3 As shown, it includes: a memory 30, a processor 32, and a computer program stored in the memory 30 and executable on the processor 32, wherein the computer program, when executed by the processor 32, performs the steps as described in the method embodiment.

[0029] Device Example 3 This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor 32, performs the steps described in the method embodiment.

[0030] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

[0031] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for recommending the next point of interest, characterized in that, include: Based on continuous temporal coding, interest point category information and interest point entity information are fused respectively to obtain user semantic hierarchical sequence and event hierarchical sequence, establish the correspondence between time-anchored user semantic intent and specific POI access behavior, and construct a joint representation of intent-POI entity; The cutoff frequency is adaptively determined through a spectral energy accumulation mechanism. Low-frequency components below or equal to the cutoff frequency are modeled as long-term preferences, while high-frequency components above the cutoff frequency are modeled as short-term preferences. By using a feature update method that guides entity behavior through semantic features, the coupling relationship between user semantic intent and specific POI access behavior in the intent-POI entity joint representation is enhanced. Combined with an energy-controlled gating fusion mechanism, the long-term preferences and the short-term preferences are dynamically fused to obtain a user preference representation. Based on the user's historical check-in trajectory and the geographical adjacency relationship of global POIs, a user intent flow graph and a global spatial context graph are constructed. Combining node features and edge features, the edge weights are decomposed into spatiotemporal structure items and semantic interaction items. By dynamically aggregating node neighborhood information through semantic layer edge weights, user intent representation and global POI representation are obtained. By mapping user preference representations and user intent representations to a shared latent space, the two types of heterogeneous features can jointly represent user behavior from a complementary perspective, resulting in a fused user representation. The user's recommendation score for the candidate points of interest is calculated using the user representation and the global POI representation, and the next point of interest recommendation result is obtained by probability sorting.

2. The method according to claim 1, characterized in that, Based on continuous temporal encoding, interest point category information and interest point entity information are fused separately to obtain user semantic hierarchical sequences and event hierarchical sequences. A correspondence is established between time-anchored user semantic intent and specific POI access behavior, and a joint representation of intent-POI entities is constructed, specifically including: Construct learnable, independent embedding matrices for POIs and categories respectively. , ,in This represents the total number of POIs and categories. Represents the embedding dimension. Given the i-th check-in in a sequence, its corresponding POI identifier. and category identifier The dense vector representation is as follows: Official 1; According to Formula 2, time is normalized within the sequence: Official 2; in It is normalized time. User The maximum and minimum values ​​in the set of all check-in timestamps. It is a very small constant; According to Formula 3, Time2Vec is used to scalar time. Vectorized encoding is performed to capture both the linear trend and multi-scale periodic patterns of user behavior over time. Official 3; in, It is the mapped continuous time vector. Indicates the embedding dimension. As learnable parameters, the first dimension represents the non-periodic long-term trend through a linear transformation, while the remaining dimensions use sinusoidal basis functions to capture the periodic dependence at different frequencies. A two-layer coupled representation is constructed using time as an anchor point: the semantic layer, as shown in Equation 4, focuses on capturing the user's semantic intent by concatenating category embeddings with temporal embeddings; the event layer, as shown in Equation 5, focuses on the user's specific entity selection behavior by concatenating POI entity embeddings with temporal embeddings. Through this concatenated representation, the model learns the user's semantic intent and specific access behavior at the corresponding time point. Official 4; Official 5; in, This represents a vector concatenation operation; Initial node features for constructing a full graph of nodes within the same embedding space: for any graph node Its corresponding POI identifier and category identifier are The initial characteristics of the nodes are: Official 6.

3. The method according to claim 1, characterized in that, The cutoff frequency is adaptively determined through a spectral energy accumulation mechanism. Low-frequency components below or equal to the cutoff frequency are modeled as long-term preferences, and high-frequency components above the cutoff frequency are modeled as short-term preferences. Specifically, this includes: Using Fast Fourier Transform to transform semantic sequences and event sequence Transform from the time domain to the frequency domain: Official 7; in, For sequence length, It is the hidden layer dimension; Before calculation The cumulative energy percentage of each frequency component, when it exceeds a certain proportion When the frequency is equal to the cutoff frequency, the corresponding frequency is called the cutoff frequency. Set the energy of frequency 0 to 0: Official 8; in, Indicates the first The and the first Fourier coefficients at discrete frequency positions; Will Normalization to The cutoff frequency is obtained within the interval. Cutoff frequency calculated from semantic sequence The global threshold is also applied to the event sequence. After determining the cutoff frequency threshold, a Butterworth low-pass filter is used to filter the signal. The Butterworth low-pass filter controls the frequency roll-off characteristic through its order, achieving a good balance between cutoff characteristics and smoothness. Official 9; Where n represents the order of the Butterworth low-pass filter; Indicates the filter at the 1st Amplitude-frequency response coefficient at each frequency component; Construct a filter mask using the Butterworth low-pass filter. Extracting low-frequency long-term components using complementary masks Extracting high-frequency short-term components: Official 10; in, This represents the long / short-term preference of semantic sequences in the frequency domain. This indicates the long / short-term preference of the event sequence in the frequency domain. This indicates element-wise multiplication; The decoupled long-term and short-term components are restored back to the time domain using IFFT: Equation 11; in, This indicates the long / short-term preference of semantic sequences in the time domain. This indicates the long / short-term preference of the event sequence in the time domain; After obtaining the decomposed representations of the semantic sequence and the event sequence, the event component is used as the query, and the corresponding semantic component is used as the key and value. The original sequence features are preserved through residual connections. Official 12; in, It captured long-term preferences driven by users' lifestyle habits. It captures short-term intents triggered by users' immediate needs, where Q represents the query vector, K represents the key vector, and V represents the value vector.

4. The method according to claim 1, characterized in that, By combining energy-controlled gating fusion mechanisms, the long-term preferences and short-term preferences are dynamically fused to obtain a user preference representation, specifically including: A gating mechanism controlled by signal energy is employed, which generates gating coefficients by calculating the logarithmic energy difference between low-frequency and high-frequency signals at the current moment. : Official 13; in, and They represent The energy intensity of low-frequency and high-frequency components at any given time. For a given energy ratio, These are learnable parameters; Final User Preference Representation Represented by long-term preferences and short-term preference representation The weighted fusion yields: Official 14.

5. The method according to claim 1, characterized in that, Based on users' historical check-in trajectories and the geographical adjacency relationships of global POIs, a user intent flow graph and a global spatial context graph are constructed. Combining node features and edge features, edge weights are decomposed into spatiotemporal structure items and semantic interaction items. By dynamically aggregating node neighborhood information through semantic layers, user intent representations and global POI representations are obtained, specifically including: The spatiotemporal attributes are normalized, and the normalized time interval is: Official 15 in, This represents the minimum value in the sequence of user check-in time intervals. It is a very small constant; User records The spherical distance between two adjacent POI visits is the spatial interval, and the latitude and longitude of each POI are recorded as follows: The spherical distance is calculated using the Haversine distance formula, and the normalized spatial interval is: Official 16; in, This represents the minimum non-zero value in the sequence of spatial distances between adjacent points of interest visited by a user. Based on graph node features Define two graph structures: SeqDSE-GCN( ): A node is a unique POI visited in the user's history. A directed edge is established between two consecutively visited POIs to capture the trend of the user's trajectory. The edge feature includes a normalized time interval. and spatial distance , Represents the user intent flow diagram; GeoDSE-GCN ( ): Nodes are all POIs in the dataset. If the distance between two POIs is less than a preset threshold... Then, construct undirected edges, with the edge characteristic being the normalized spatial distance. , Represents a global spatial context graph; for To express the prior influence of spatiotemporal proximity on transition probabilities, we define... For nodes and nodes The spatiotemporal structural edge weights between them are specifically a joint exponential decay function of distance and time: Official 17; in, This is a learnable attenuation coefficient used to balance the relative importance of time and space factors. Represents a node and nodes The time interval, Represents a node and nodes Spatial interval; definition For the first Layer nodes and nodes Dynamic semantic edge weights between: Official 18; in, For learnable parameters, For the sigmoid function, The heat coefficient, Indicates the first Layer nodes and nodes Node characteristics; The final edge weight is the product of two terms. ; for Similarly, the construction uses structural terms plus semantic terms. The structural edge weights are initialized with a Gaussian kernel distance decay function. This structure is used to express the spatial prior that the closer the elements are, the stronger the relationship. Official 19; in, Represents a node and nodes The spherical distance representing the POI; Dynamic semantic edge weights are constructed using a Gaussian kernel function in the feature space: Official 20; in, The projection matrix is ​​learnable. Softplus is used to ensure positivity; the closer the endpoints are in the semantic space, the closer the semantic term is to 1, and the final edge weight is... For user intent flow graphs and global spatial context diagram Different flow characteristics for directed graphs Calculate the independent normalization coefficients: Official 21; in, and They are nodes In the The in-degree normalization factor and out-degree normalization factor of the layer, and Representing the in-degree neighbor set and the out-degree neighbor set, the graph convolution aggregation process is represented as: Official 22; in, and These are the original node features before and after the update. It is a feature of neighboring nodes. It is an activation function. It is a learnable weight matrix, which is aggregated through graph convolution to ultimately obtain a representation of user intent; For undirected graphs , No. The weighting degree of the layer is The aggregation intensity is adjusted using a symmetric normalization operator. The graph convolution aggregation process is represented as follows: Official 23; in, The weight matrix is ​​a learnable matrix, after... After layer propagation, a global POI representation is obtained.

6. The method according to claim 1, characterized in that, Mapping user preference representations and user intent representations onto a shared latent space allows these two heterogeneous features to jointly represent user behavior from a complementary perspective, resulting in a fused user representation. The user representation and the global POI representation are then used to calculate the user's recommendation score for candidate points of interest, and the next point of interest recommendation is obtained through probability ranking. Specifically, this includes: In the prediction stage, a nonlinear fusion strategy is first used to map heterogeneous features to a shared latent space: Official 24; in, Representation layer normalization, This represents the activation function. For learnable parameters, To represent user preferences, To represent user intent, As a complete user representation, it includes both short-term and long-term user preferences and trends in user intent. Through user representation Representation of Global POI The predicted score is calculated to obtain the recommendation result, where, Provided by the global spatial context map encoder, it contains objective geographical constraint information. Predicted score Defined as: Official 25; in, yes The geographic embedding representation, and the final access probability distribution Obtained through the Softmax function Indicates transpose; To address the sparsity of check-in data and enhance model interpretability, a loss function that incorporates both the primary recommendation task and dual auxiliary constraints is employed. The primary recommendation task utilizes labeled, smoothed cross-entropy loss to suppress overfitting on sparse data and improve robustness to noisy labels. Official 26; in, To recommend losses, To predict scores, It is a model that predicts POI The probability of access. It is a smoothing factor. It is the number of candidate POIs; To prevent the model from focusing too much on POI IDs and ignoring the underlying semantic intent, a semantic auxiliary loss is introduced: Official 27; in, It is the total number of POI categories. It is the one-hot tag of the category to which the real POI belongs. These are learnable hyperparameters; To ensure the long-term components obtained from frequency domain decomposition and short-term components In accordance with physical principles, energy constraints are introduced: Official 28; in, Indicates the sequence at time step The long-term and short-term eigenvectors, These are the balance coefficients. The first term in Formula 28 is a smoothness constraint, ensuring that long-term preferences remain stable over time. The second term in Formula 28 is a recency constraint. The mask vector marks historical time steps far from the current time as 1, which penalizes the energy residue of short-term high-frequency signals in distant history. The model's final objective loss function is a weighted sum of the above components: Official 29; in, It is a weighting factor used to balance semantic guidance and energy constraint sharing.

7. A device for recommending the next point of interest, characterized in that, include: The acquisition module is used to fuse point of interest category information and point of interest entity information based on continuous time encoding, acquire user semantic hierarchical sequence and event hierarchical sequence, establish the correspondence between time-anchored user semantic intent and specific POI access behavior, and construct a joint representation of intent-POI entity; The cutoff frequency module is used to adaptively determine the cutoff frequency through a spectral energy accumulation mechanism, modeling low-frequency components below or equal to the cutoff frequency as long-term preferences, and high-frequency components above the cutoff frequency as short-term preferences. The enhanced fusion module is used to guide the feature update method of entity behavior through semantic features, enhance the coupling relationship between user semantic intent and specific POI access behavior in the intent-POI entity joint representation, and combine the energy-controlled gating fusion mechanism to dynamically fuse the long-term preferences and the short-term preferences to obtain the user preference representation. The calculation update module is used to construct a user intent flow graph and a global spatial context graph based on the user's historical check-in trajectory and the geographical adjacency relationship of global POIs. Combining node features and edge features, the edge weights of the user preference representation are decomposed into spatiotemporal structure items and semantic interaction items. The node neighborhood information is aggregated through dynamic edge weights between semantic layers to obtain the user intent representation and the global POI representation. The mapping module is used to map user preference representations and user intent representations to a shared latent space, enabling the two types of heterogeneous features to jointly represent user behavior from a complementary perspective, resulting in a fused user representation. The scoring module is used to calculate the user's recommendation score for the candidate points of interest using the user representation and the global POI representation, and to obtain the next point of interest recommendation result through probability sorting.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the next point of interest recommendation method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an implementation program for information transmission, which, when executed by a processor, implements the steps of the next point of interest recommendation method as described in any one of claims 1 to 6.

Images