Long-tail human movement prediction method and system based on adaptive graph marginal learning

Abstract

The invention provides a long-tail human movement prediction method and system based on adaptive graph marginal learning, and the method comprises the steps: building a multi-source semantic graph structure: building a multi-layer semantic graph which comprises a physical address, an address category, an activity mode and a demand type, so as to express the multivariate function attributes of a geographic position; performing adaptive marginal learning based on the multi-source semantic graph: calculating a basic marginal according to an access frequency of a geographic position, and performing propagation and smoothing on the basic marginal through a semantic neighbor relationship by using a topological structure of the multi-source semantic graph to generate an effective marginal used for adjusting a model decision boundary; and carrying out joint model optimization based on the multi-source semantic graph and the effective margins, and predicting a next visited geographic position of the user according to a historical movement track of the user by using a trained model. According to the method, the accuracy, fairness and robustness of human movement prediction in a real complex scene are remarkably improved, and the method is more suitable for being used in multiple scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and urban computing, in particular to a long-tail human movement prediction method and system based on adaptive graph margin learning. BACKGROUND

[0002] Human movement prediction, as a core technology in the field of artificial intelligence and urban computing, aims to predict the next possible geographic location visited by a user based on their historical trajectory data. In recent years, with the popularity of mobile intelligent terminals and the maturity of positioning technology, location-based service platforms have accumulated massive user movement trajectory data, providing a rich data foundation for understanding human movement patterns. Effective human movement prediction technology not only promotes the development of commercial applications such as location-aware recommendation, personalized content distribution, and real-time service delivery, but also provides important technical support for key urban functions such as intelligent transportation planning, dynamic road network management, and public health monitoring in smart city construction.

[0003] However, human movement data in the real world generally exhibits a significant long-tail distribution characteristic. Specifically, a small number of popular locations are frequently visited, constituting the head of the data; while a large number of ordinary locations have very low visit frequency, forming the tail of the data. This unbalanced distribution makes traditional prediction models severely biased towards head locations, with significantly decreased prediction performance for tail locations, thereby restricting the accuracy and fairness of the model in practical applications.

[0004] To capture the complex spatio-temporal patterns in human movement behavior, researchers have proposed various prediction models based on deep learning. Methods based on recurrent neural networks are good at capturing local dependencies in continuous trajectories through their inherent temporal modeling capabilities; models based on graph neural networks model geographic locations and their associated relationships as a graph structure, effectively mining the spatial proximity and functional connectivity between locations; models based on Transformer utilize the advantages of self-attention mechanisms to effectively capture long-distance global context; while methods based on diffusion models focus on modeling and reasoning about uncertainty in human movement behavior through their powerful generative capabilities.

[0005] Despite the significant progress made by these methods in their respective modeling perspectives, they generally have the following fundamental limitations: First, existing methods lack the ability to integrate multiple sources of dynamic semantics for geographic locations. Traditional models mostly use static semantic representation frameworks, which are difficult to support the dynamic integration and adaptation of multiple source heterogeneous semantic features such as physical addresses, functional categories, activity patterns, and demand types. This rigid representation mode results in the model's inability to adaptively adjust its semantic understanding of geographic locations according to specific context environments, limiting the model's deep perception and reasoning capabilities for complex movement intentions.

[0006] Secondly, the prior art lacks a semantic-guided long-tail optimization mechanism. Current long-tail processing techniques mostly rely on simple sample reweighting or data augmentation strategies, which are separated from the semantic understanding process and lack a directional optimization mechanism based on semantic relationships. This fragmented architecture makes it difficult for the model to achieve knowledge transfer from data-rich head locations to sparse tail locations, and cannot fully utilize the guiding role of semantic information in alleviating the long-tail effect.

[0007] In addition, the relationship-driven feature learning mechanism of existing models is imperfect. Most methods lack structured constraints based on semantic relationships in the feature learning process, making it difficult to obtain effective regularization guidance for the representation learning of tail locations. This makes it difficult for tail locations to form a discriminative distribution pattern with other locations in the semantic feature space, restricting the overall representation learning effect and generalization ability of the model.

[0008] Specifically, existing technical solutions, such as the Graph-Flashback method, although construct a spatio-temporal knowledge graph to learn the representation of users and POIs, fail to effectively address the long-tail distribution problem; the STHGCN method uses hypergraphs to capture user historical trajectory information, but has limitations in semantic fusion; the ROTAN method uses a rotation attention mechanism to process time information, but does not sufficiently understand spatial semantics; the Diff-POI method uses a diffusion model for recommendation, but lacks specialized optimization for the long-tail problem; the latest LoTNext and HiMob methods, although they begin to focus on the long-tail problem, still have deficiencies in multi-source semantic fusion and relationship-driven learning.

[0009] In summary, existing technologies have not systematically addressed the core challenges of long-tail human mobility prediction from the three key dimensions of "multi-source dynamic semantic fusion", "semantic-guided optimization mechanism", and "relationship-driven feature learning". Therefore, there is an urgent need for an innovative framework that can represent the polysemy of locations through flexible graph structures and implement marginal adjustment and contrastive learning optimization based on these structures to comprehensively improve the accuracy, fairness, and generalization ability of mobility prediction in real long-tail scenarios. SUMMARY

[0010] The problem solved by the present application is that existing human mobility prediction technologies lack sufficient consideration of the semantic modeling of location multi-functional attributes, lack a semantic-guided long-tail optimization mechanism, and lack targetedness in feature learning for tail locations.

[0011] To solve the above problems, the present application provides a long-tail human mobility prediction method, system, electronic device, and storage medium based on adaptive graph marginal learning.

[0012] In a first aspect, the present application provides a long-tail human mobility prediction method based on adaptive graph marginal learning, comprising: Constructing a multi-source semantic graph structure: a multi-layer semantic graph containing physical addresses, address categories, activity patterns, and demand types is established to express the multi-functional attribute of geographic locations; Adaptive margin learning based on the multi-source semantic graph: the base margin is calculated according to the access frequency of geographic locations, and the effective margin used to adjust the model decision boundary is generated by propagating and smoothing the base margin through semantic neighbor relationships using the topology of the multi-source semantic graph; Joint model optimization based on the multi-source semantic graph and the effective margin: Based on the multi-source semantic graph, a semantic-driven positive and negative sample pair is constructed for contrastive learning, and the effective margin is introduced to adjust the loss function in the main prediction task. The prediction model is trained by jointly optimizing the main prediction task and the contrastive learning task. Using the trained model, the next visited geographic location of the user is predicted according to the user's historical moving trajectory.

[0013] Optionally, the multi-layer semantic graph includes a four-layer graph structure: The physical address layer contains a plurality of specific geographic location nodes; The address category layer includes a plurality of semantic category nodes based on the open platform classification system; The activity pattern layer contains a plurality of standard activity category nodes; The demand type layer is constructed based on the basic demand type node of the demand hierarchy theory; Wherein, the nodes of different layers are connected through many-to-many edges.

[0014] Optionally, in the adaptive graph margin learning, the calculation formula of the base margin value is: ; Wherein, is the category The frequency of occurrence in the training stage, The total frequency of all categories, Indicates the margin scaling factor, , Indicates the distribution sensitivity parameter, , Indicates the smoothing term.

[0015] Optionally, the margin propagation is realized by the following formula: ; Wherein, Indicates the global smoothing coefficient, , Indicates the relative weight between neighbors, , used to measure the neighbor node For the current node The intensity of the impact, For category and co-occurrence frequency, Represents semantic neighbor nodes The basic marginal value.

[0016] Optionally, the loss function of the main prediction task is: ; in, Indicates the feature scaling factor. It is a real category. Represents the effective margin of the real category. It measures the similarity in direction between a feature vector and the weight vector of its class.

[0017] Optionally, the loss function for the contrastive learning is: ; in, Indicates anchor point sample, Represents positive samples and The feature vector representing the negative sample. This represents the similarity calculation function. Indicates temperature parameter, This represents an exponential function.

[0018] Optionally, the positive samples are obtained through multi-hop neighbor sampling, and the number of samples... The negative samples are obtained using a hard sample mining strategy, with a sampling quantity of [number missing]. .

[0019] Optionally, the total loss function of the joint optimization is: ; in, Indicates the weighting coefficient; ; in, Indicates the initial weights. This is the current training round. This indicates the decay rate.

[0020] Secondly, embodiments of the present invention provide a long-tail human movement prediction system based on adaptive graph marginal learning, comprising: The multi-source semantic graph construction module is used to build a multi-layer semantic graph structure that includes physical address, address category, activity mode and demand type to express the diverse functional attributes of geographic location; An adaptive margin learning module is used to calculate basic margins based on the topological structure and access frequency of the multi-source semantic graph and propagate and smooth them through semantic neighbor relationships, generating effective margins for adjusting the decision boundary of the model. The joint optimization module is used to construct semantically driven positive and negative sample pairs based on the multi-source semantic graph for comparative learning, and to introduce the effective margin in the main prediction task to adjust the loss function. The prediction model is trained by jointly optimizing the main prediction task and the comparative learning task. The prediction execution module uses a trained model to predict the geographical location of a user's next visit based on their historical movement trajectory.

[0021] Optionally, the system is configured to be deployed in at least one of the following application scenarios: In the intelligent transportation planning scenario, the system is deployed on a traffic management server to process vehicle trajectory data in real time to optimize traffic light timing and road network planning. Location service recommendation scenario, wherein the system is integrated into the front-end application and is used to provide personalized address recommendations based on the user's real-time trajectory; In urban management scenarios, the system is integrated with a municipal IoT platform to enable public facility utilization efficiency analysis and layout optimization. In a public health monitoring scenario, the system interfaces with the data system of the Center for Disease Control and Prevention to build a dynamic assessment model of epidemic transmission risk based on population movement patterns.

[0022] Thirdly, embodiments of the present invention provide an electronic device, including a processor, a communication interface, a memory, and a bus, wherein the processor, the communication interface, and the memory communicate with each other through the bus, and the processor can call logical instructions in the memory to execute the steps of the method provided in the first aspect.

[0023] Fourthly, embodiments of the present invention provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the long-tail human movement prediction method based on adaptive graph marginal learning as described in the first aspect.

[0024] The beneficial effects of the long-tail human movement prediction method based on adaptive graph marginal learning of the present invention are as follows: This invention constructs a graph structure with polysemous representation capabilities and combines it with an adaptive marginal adjustment mechanism to achieve the dynamic fusion and effective utilization of diverse semantic attributes of geographic locations and complex spatial context information. Adaptive marginal learning, as a forward adjustment mechanism based on category distribution, can explicitly adjust the classification decision boundary, enabling the model to discriminate against long-tail categories in the feature space. This technical characteristic makes it an ideal choice for building a fair prediction system with semantic awareness. Furthermore, by introducing a multi-granularity contrastive learning strategy, this method further enhances the model's ability to learn features related to semantically relevant locations. This invention aims to systematically address the problems commonly found in existing prediction methods, such as rigid semantic representation, insufficient long-tail optimization, and poor scene generalization ability, significantly improving the accuracy, fairness, and robustness of human movement prediction in real-world complex scenarios, thus making it more adaptable to various scenarios. Attached Figure Description

[0025] Figure 1 This is a flowchart of the long-tail human movement prediction method based on adaptive graph marginal learning in an embodiment of the present invention; Figure 2 This is a schematic diagram of a long-tail human movement prediction framework based on adaptive graph marginal learning in an embodiment of the present invention. Figure 3 This is a structural block diagram of the long-tail human movement prediction system based on adaptive graph marginal learning in an embodiment of the present invention; Figure 4 This is a structural block diagram of the electronic device in an embodiment of the present invention. Detailed Implementation

[0026] To better understand the purpose, technical solution, and advantages of this application, the application is described and explained below in conjunction with the accompanying drawings and embodiments.

[0027] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.

[0028] To better understand this technical solution, it is briefly described below: This invention proposes a long-tail human movement prediction method based on adaptive graph marginal learning, aiming to improve the prediction accuracy and generalization ability of the model in long-tailed distribution scenarios. It assumes that the address set contained in the human movement data is... ,in This refers to the number of addresses. Each user's movement trajectory sequence is represented as... , in It is the length of the trajectory. Each spacetime point in the trajectory... A triple containing spatial location, semantic category, and temporal information. The core objective of this invention is based on historical trajectories. Accurately predict the next access location .

[0029] To address the shortcomings of existing human mobility prediction technologies in semantic representation, long-tail optimization, and feature learning, this paper combines... Figure 2 As shown, this invention proposes an overall technical framework and its core modules based on adaptive graph marginal learning, including: Multi-source Semantic Dynamic Fusion Module: This invention designs a semantic representation system based on a graph structure. By constructing a four-layer graph structure including physical address, address category, activity mode, and demand type, it supports many-to-many connections between nodes. This framework breaks through the single-attribute limitation of traditional tree structures and can naturally express the composite functional attributes of geographical location, providing a fundamental support for deep optimization guided by semantics.

[0030] Adaptive Graph Marginal Learning Module: Addressing the imbalance problem in long-tailed distributions, this invention proposes a graph-based margin adjustment mechanism. This module calculates adaptive margins based on location access frequency and utilizes the topological relationships of the graph structure for margin propagation and smoothing. This enables semantic knowledge transfer from data-rich head locations to sparse tail locations, effectively improving the model's ability to discriminate long-tailed locations.

[0031] Multi-granularity contrastive joint optimization module: This invention designs a joint optimization strategy that combines the main prediction task with contrastive learning. By constructing semantically driven positive and negative sample pairs on the graph structure, this module enhances the representation quality of tail locations in the feature space, enabling them to form an effective clustering structure with semantically similar head locations, thereby comprehensively improving the model's representation learning ability and generalization performance.

[0032] This technical solution, through the collaborative work of the three core modules mentioned above, systematically solves the key problems of existing human mobility prediction technologies in terms of semantic richness, long-tail adaptability, and feature discriminability while maintaining the generality of the architecture. The following section will elaborate on this method.

[0033] Combination Figure 1 and Figure 2 As shown, this embodiment of the invention provides a method for predicting long-tail human movement based on adaptive graph marginal learning, including: S1. Construct a multi-source semantic graph structure: Establish a multi-layer semantic graph that includes physical address, address category, activity mode and demand type to express the diverse functional attributes of geographic location; To address the limitations of existing human mobility prediction methods in semantic representation, this invention designs a multi-source semantic dynamic fusion module (i.e., constructing a multi-source semantic graph structure or a multi-layer semantic graph). By establishing a four-layer heterogeneous semantic network, it achieves accurate characterization of complex functional attributes of geographic locations.

[0034] In this embodiment, the multi-layer semantic graph includes a four-layer graph structure: The physical address layer contains multiple specific geographic location nodes; The address category layer, based on the open platform classification system, includes multiple semantic category nodes; The activity mode layer contains multiple standard activity category nodes; The demand type layer is constructed based on the demand hierarchy theory, which builds basic demand type nodes. Nodes in different layers are connected by many-to-many edges.

[0035] In one optional implementation, the multi-layer semantic graph (multi-source semantic dynamic fusion module) comprises four tightly coupled hierarchical structures. The base layer is the physical address layer, containing... Each is a specific geographic location node. An address category layer is built upon this, based on the classification system of open platforms such as Foursquare, containing... Each node represents a semantic category. The higher-level activity pattern layer includes 10 standard activity categories: arts and entertainment, business services, community and government, catering, healthcare, landmarks and outdoor activities, nightlife, retail, sports and entertainment, and travel and transportation. The top-level demand type layer is based on Maslow's hierarchy of needs, constructing three basic demand types: physiological needs, safety needs, and social needs. .

[0036] Edge connection construction employs a multi-source knowledge fusion strategy. Connections from physical addresses to address categories establish a deterministic mapping based on geocoding services. Connections from address categories to activity patterns generate ambiguous mappings through a large language model, using thought chain prompts to obtain probability distributions, retaining probabilities greater than a threshold. Semantic connections. The connection between activity patterns and demand types is based on a hierarchical association network established by cognitive psychology theory.

[0037] S2. Adaptive Margin Learning Based on Multi-Source Semantic Graph: Calculate basic margins based on the access frequency of geographical locations, and use the topological structure of the multi-source semantic graph to propagate and smooth the basic margins through semantic neighbor relationships to generate effective margins for adjusting the decision boundary of the model. Furthermore, this step is implemented based on the constructed adaptive graph edge learning module. Traditional long-tail learning methods often ignore semantic structural information, resulting in low knowledge transfer efficiency. The adaptive graph edge learning module proposed in this invention achieves semantic-driven fine-grained optimization by deeply coupling edge learning with graph structure.

[0038] In adaptive graph marginal learning, based on category The formula for calculating the basic marginal value of a category based on sample probability is: ; in, It is a category Frequency of occurrence during the training phase The total frequency across all categories. The marginal scaling factor is used to control the overall marginal scale. , The distribution sensitivity parameter is used to adjust the strength of the response to a long-tailed distribution. , The smoothing term is used to ensure numerical stability. The sample statistics use an exponential moving average update mechanism, with a smoothing factor set to 0.9.

[0039] The edge propagation mechanism fully utilizes the topological characteristics of graph structures to achieve coordinated adjustment of edge values ​​through semantic neighbor relationships. Specifically, the effective edge of a node is a weighted combination of its own edge and the edges of its semantic neighbors, calculated as follows: ; in, This represents the global smoothing coefficient, used to control the balance between the influence of a node's own perimeter and its neighboring groups. It is dynamically adjusted based on the node's connection density. For semantically isolated nodes, set To avoid noise transmission. This represents the relative weights among neighbors. Used to measure neighboring nodes For the current node The intensity of the impact, For category and co-occurrence frequency, Represents semantic neighbor nodes The basic marginal value.

[0040] S3. Joint model optimization based on multi-source semantic graph and effective margin: construct semantically driven positive and negative sample pairs based on multi-source semantic graph for comparative learning. At the same time, introduce effective margin in the main prediction task to adjust the loss function. Train the prediction model by jointly optimizing the main prediction task and the comparative learning task. This step is implemented based on the constructed adaptive graph marginal learning module and multi-granularity comparison joint optimization module.

[0041] The loss function design incorporates adjusted marginal information, and the loss function for the main prediction task is calculated as follows: ; in, The feature scaling factor is automatically optimized based on the feature distribution characteristics. , It is a real category. Represents the effective margin of the real category. It measures the similarity in direction between a feature vector and the weight vector of its class. The adaptive graph marginal learning module creates buffer zones (i.e., boundaries) for the model's decision boundary, in order to ensure... Subtract Even if it can still produce high probabilities, the model must learn more discriminative and directional features.

[0042] Furthermore, to improve the feature learning quality of tail locations, this invention designs a multi-granularity contrastive joint optimization module. This module provides high-quality supervision signals for contrastive learning by mining semantic relationships within the graph structure.

[0043] During the sample construction phase, positive samples are sampled using multi-hop neighbor sampling based on the topological relationships of the graph structure, with each anchor point sampled. There are 10 semantically relevant positive samples. Negative samples are selected from the same training batch using a hard sample mining strategy. The most challenging negative samples were selected. The sampling process employed a degree-weighted strategy to effectively balance the selection probabilities of popular and unpopular nodes.

[0044] The contrast loss calculation uses the temperature-regulated InfoNCE loss function: ; in, Indicates anchor point sample, Represents positive samples and The feature vector representing the negative sample. This represents the similarity calculation function, which uses cosine similarity to measure the degree of similarity between feature vectors. This represents a temperature parameter used for fine-tuning the concentration of characteristic distributions. , This represents an exponential function. This loss function effectively improves the model's feature discrimination ability by maximizing the similarity between the anchor point and positive samples and minimizing the similarity between the anchor point and negative samples.

[0045] The joint optimization framework achieves multi-task collaborative learning through dynamic weight adjustment. The total loss function of the joint optimization is: ; in, This represents the weighting coefficient, which employs a gradual decay strategy. ; in, This refers to the current training epoch (counting from 1) and the initial weights. decay rate Total number of training cycles .

[0046] At the engineering implementation level, the system employs a multi-layered optimization architecture to ensure algorithm efficiency. Trajectory sequence processing supports dynamic length adjustment, including sequence length... Sliding window step size Embedded Dimension Adaptively selects batch size based on hardware configuration. This approach combines gradient accumulation techniques to balance training efficiency and stability. The AdamW optimizer is employed, with a weight decay range... Initial learning rate In conjunction with the cosine annealing scheduling strategy. Gradient clipping norm threshold. Effectively maintain training stability.

[0047] S4. Using the trained model, predict the geographical location of the user's next visit based on the user's historical movement trajectory.

[0048] Compared with existing technologies, the long-tail human movement prediction method based on adaptive graph marginal learning proposed in this application effectively solves key problems such as single semantic representation, insufficient long-tail optimization, and weak discriminative power of feature learning in long-tail distribution scenarios, and has the following significant technical advantages and effects: 1. Prediction accuracy has been significantly improved. This method significantly improves the overall accuracy of human movement prediction through polysemous graph structure representation and an adaptive marginal learning mechanism. Experimental results on two standard human movement datasets, JKT and KLP, show that our method (denoted as AGML) achieves significant improvements across all evaluation metrics compared to the baseline method. Specifically, on the JKT dataset, AGML improves upon HiMob's performance in the three evaluation metrics MRR@1, MRR@5, and MRR@10 of the graph-Flashback structure by 8.67%, 5.67%, and 5.41%, respectively; on the KLP dataset, the corresponding improvements are 8.02%, 5.07%, and 4.80%. This improvement fully demonstrates the effectiveness and robustness of our method in complex urban scenarios.

[0049] 2. Excellent long-tail optimization effect To address the long-tail distribution problem, this method significantly improves the performance of tail address prediction through a semantically guided margin propagation mechanism. Experimental data show that AGML demonstrates further advantages over the existing long-tail processing method HiMob in all metrics. For example, based on the STHGCN model, AGML achieves an MRR@1 of 0.2245 on the JKT dataset, a 7.06% improvement over the original model and superior to HiMob's 0.2097. Particularly noteworthy is the 12.78% improvement in the MRR@1 of the Diff-POI model on the JKT dataset, demonstrating its superior adaptability in extreme long-tail scenarios.

[0050] 3. Strong architectural versatility This method demonstrates excellent architectural compatibility and versatility. Consistent performance improvements were observed when integrating the AGML framework onto different types of base prediction models, including Graph-Flashback (based on recurrent neural networks), STHGCN (based on graph neural networks), MCLP (based on Transformers), and Diff-POI (based on diffusion models). On all four base models and three evaluation metrics of the JKT and KLP datasets, AGML consistently outperformed baseline methods and existing long-tail processing methods, with improvements ranging from 4.65% to 12.78%, demonstrating the method's strong versatility and scalability.

[0051] 4. Excellent semantic perception ability Through the synergistic effect of polysemous graph structures and contrastive learning mechanisms, this method demonstrates outstanding performance in complex semantic scenarios. Experimental results show that in urban scenarios with rich semantic relationships, AGML can effectively capture the diverse functional attributes of geographic locations, thereby making more accurate predictions. The improved stability across different urban datasets proves the method's ability to understand and adapt to diverse semantic patterns, providing reliable technical support for smart city applications.

[0052] Table 1. Experimental results of different models on the JKT and KLP datasets;

[0053] This technical solution systematically addresses key issues in long-tail human movement prediction, such as semantic representation, distribution optimization, and feature learning, by constructing an adaptive graph marginal learning framework. Specifically: 1. Ambiguous Graph Structure Semantic Representation Framework: This solution innovatively designs a graph structure semantic representation system supporting many-to-many connections. By constructing a complex network of associations among four semantic levels—physical address, address category, activity mode, and demand type—it overcomes the single-attribution limitation of traditional tree structures. This framework utilizes a large language model to generate ambiguous mappings and combines prior psychological knowledge to establish cross-layer connections, achieving accurate modeling of the multifunctional attributes of geographic locations. Through this technique, the model can accurately capture the complex characteristics of locations across different semantic dimensions, such as simultaneously associating a coffee shop with multiple activity modes like dining and socializing. This significantly improves the completeness and accuracy of semantic representation, laying a solid foundation for subsequent semantic guidance optimization.

[0054] 2. Semantic-Guided Adaptive Marginal Learning Mechanism: To address the model bias problem caused by long-tail distribution, this solution proposes a graph-based adaptive marginal learning method. This method calculates marginal baseline values ​​related to class frequencies and utilizes graph topology for marginal propagation and smoothing, achieving semantically driven refined optimization. Specifically, by establishing a marginal influence mechanism based on semantic neighbor relationships, the optimization process of tail locations can fully leverage the knowledge representation of their semantically related head locations. This technical solution effectively alleviates the learning imbalance problem caused by long-tail distribution. Experimental results on multiple public datasets show that the recall rate of tail locations is improved by 15%-25%, while maintaining the predictive performance of head locations, significantly improving the practical value of the model in real-world scenarios.

[0055] 3. Relationship-Driven Contrastive Joint Optimization Strategy: To improve the feature learning quality of tail locations, this scheme designs a joint optimization framework integrating multi-granularity contrastive learning and the main task. This strategy defines semantic similarity based on graph structures, constructs discriminative positive and negative sample pairs, and enhances the clustering characteristics of the feature space through a temperature-adjusted contrastive loss function. Combined with a dynamic weight scheduling mechanism, a balanced optimization of the main prediction task and the contrastive learning task is achieved during training. This technical solution significantly improves the representation quality of tail locations in the feature space, enabling them to form a compact cluster distribution with semantically similar head locations, thereby comprehensively improving the model's discriminative ability and generalization performance. Cross-domain experimental verification shows that this method reduces the generalization error by 10%-15% on multiple city datasets, demonstrating good transferability and adaptability.

[0056] This technical solution combines the above three aspects to construct a complete long-tail human mobility prediction solution. The solution is implemented in key smart city scenarios in the following ways: In the location services field, the system analyzes users' real-time trajectories and historical behavioral patterns, accurately understanding users' complex needs based on a polysemous graph structure, providing accurate next location predictions for navigation software and business recommendation systems, significantly improving the personalization and accuracy of location services; In the transportation planning field, through a prediction engine deployed in the traffic management center, it processes vehicle GPS trajectories and other mobile data in real time, predicting traffic flow distribution and crowd gathering trends in different time periods and areas, providing decision support for intelligent traffic light control, road resource allocation, and public transportation scheduling, effectively optimizing urban traffic operation efficiency.

[0057] Specifically, this solution can be applied to a variety of scenarios, as described in the following system embodiments.

[0058] Furthermore, based on the optimal technical solution, by appropriately transforming and replacing its core modules, several extended technical solutions with similar functions but different implementation methods can be derived. These solutions, while maintaining the core idea of ​​solving the long-tail human mobility prediction problem, provide different technical path options.

[0059] One feasible extension is to simplify the graph structure construction module. In resource-constrained applications, a dynamic graph construction method based on attention mechanisms can replace the predefined semantic hierarchy. This method dynamically generates graph connections by calculating the co-occurrence probability and semantic similarity between locations, without relying on a predefined hierarchical labeling system. Specifically, it can utilize the co-occurrence statistics of location access sequences and pre-trained semantic embeddings to automatically learn the semantic association strength between locations through a two-layer attention network, thus constructing an adaptive graph structure. While this approach sacrifices some semantic interpretability, it achieves better domain adaptability and lower construction costs.

[0060] Another direction for expansion is the alternative implementation of the marginal learning module. A dynamic marginal adjustment strategy based on curriculum learning could be considered to replace static marginal calculation based on frequency statistics. This method divides the training process into multiple stages, dynamically adjusting the learning difficulty of each category according to the model's training state at the current stage. Specifically, by monitoring the classification accuracy trends of each category, the learning difficulty can be predicted using an exponential moving average algorithm, thereby generating a phased marginal adjustment scheme. This approach can better adapt to the dynamic changes of the model during training, but it requires the design of more complex training scheduling strategies.

[0061] This invention also provides a long-tail human movement prediction system based on adaptive graph marginal learning. This system is used to implement the above-described method embodiments, and details already described will not be repeated. The terms "module," "unit," "subunit," etc., used below refer to combinations of software and / or hardware that perform a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation or a combination of software and hardware is also possible and contemplated.

[0062] like Figure 3 As shown, Figure 3 This is a block diagram of the long-tail human movement prediction system based on adaptive graph marginal learning in this invention. The system includes: The multi-source semantic graph construction module 101 is used to establish a multi-layer semantic graph structure that includes physical address, address category, activity mode and demand type to express the multi-functional attributes of geographic location; The adaptive margin learning module 102 is used to calculate the basic margin based on the topological structure and access frequency of the geographic location of the multi-source semantic graph, and to propagate and smooth it through semantic neighbor relationships to generate an effective margin for adjusting the decision boundary of the model. The joint optimization module 103 is used to construct semantically driven positive and negative sample pairs based on the multi-source semantic graph for comparative learning, and to introduce the effective margin in the main prediction task to adjust the loss function. The prediction model is trained by jointly optimizing the main prediction task and the comparative learning task. The prediction execution module 104 is used to predict the geographical location of a user's next visit based on the user's historical movement trajectory using a trained model.

[0063] Furthermore, the system is configured to be deployed in at least one of the following application scenarios: In the intelligent transportation planning scenario, the system is deployed on a traffic management server to process vehicle trajectory data in real time to optimize traffic light timing and road network planning. Location service recommendation scenario, wherein the system is integrated into the front-end application and is used to provide personalized address recommendations based on the user's real-time trajectory; In urban management scenarios, the system is integrated with a municipal IoT platform to enable public facility utilization efficiency analysis and layout optimization. In a public health monitoring scenario, the system interfaces with the data system of the Center for Disease Control and Prevention to build a dynamic assessment model of epidemic transmission risk based on population movement patterns.

[0064] like Figure 4 As shown in the figure, an electronic device provided in this embodiment of the invention includes: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute the method provided in the above embodiment.

[0065] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0067] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0068] 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A long-tail human movement prediction method based on adaptive graph margin learning, characterized in that, The method comprises the following steps: constructing a multi-source semantic graph structure: a multi-layer semantic graph containing physical addresses, address categories, activity patterns, and demand types is established to express the multi-functional attributes of geographic locations; adaptive marginal learning based on the multi-source semantic graph: the base margin is calculated according to the access frequency of geographic locations, and the effective margin used to adjust the decision boundary of the model is generated by propagating and smoothing the base margin through semantic neighbor relationships based on the topological structure of the multi-source semantic graph; joint model optimization based on the multi-source semantic graph and the effective margin: semantic-driven positive and negative sample pairs are constructed based on the multi-source semantic graph for contrastive learning, and the effective margin is introduced into the main prediction task to adjust the loss function, and the prediction model is trained by jointly optimizing the main prediction task and the contrastive learning task; using the trained model, predicting the next visited geographic location of a user according to the user's historical moving trajectory.

2. The method of claim 1, wherein, The multi-layer semantic graph comprises four layers of graph structures: a physical address layer containing a plurality of specific geographic location nodes; an address category layer comprising a plurality of semantic category nodes based on an open platform classification system; an activity pattern layer containing a plurality of standard activity category nodes; a demand type layer comprising basic demand type nodes based on the demand hierarchy theory; wherein the nodes of different layers are connected through many-to-many edges.

3. The method of claim 1, wherein, In the adaptive graph marginal learning, the calculation formula of the base margin value is: ； where, is the class the frequency of occurrence in the training phase, the total frequency of all classes, denotes the marginal scaling factor, , denotes the distribution sensitivity parameter, , denotes the smoothing term.

4. The method of claim 3, wherein, The marginal propagation is realized through the following formula: ； wherein, denotes a global smoothing coefficient, , denotes a relative weight between neighbors, , for measuring the influence strength of a neighbor node on the current node , is the co-occurrence frequency of the categories and , denotes the base marginal value of a semantic neighbor node .

5. The method of claim 1, wherein, The loss function of the main prediction task is: ； wherein, denotes a characteristic scaling factor, is a true class, denotes an effective margin of the true class, measures the similarity of the feature vector to the weight vector of the class it belongs to in direction.

6. The method of claim 5, wherein, The loss function of the contrastive learning is: ； wherein, denotes an anchor sample, denotes a positive sample and denotes a feature vector of a negative sample, denotes a similarity computation function, denotes a temperature parameter, denotes an exponential function.

7. The method of claim 6, wherein, The positive samples are obtained by multi-hop neighbor sampling, and the sampling number The negative samples adopt a difficult sample mining strategy, and the sampling number .

8. The method of claim 6, wherein, The total loss function of the joint optimization is: ； wherein denotes a weight coefficient; ； wherein, denotes the initial weight, is the current training epoch, denotes the decay rate.

9. A long-tail human movement prediction system based on adaptive graph margin learning, characterized in that, The method comprises the following steps: a multi-source semantic graph construction module for establishing a multi-layer semantic graph structure containing physical addresses, address categories, activity patterns, and demand types to express the multi-functional attributes of geographic locations; an adaptive marginal learning module for calculating a base margin and propagating and smoothing it through semantic neighbor relationships based on the topological structure of the multi-source semantic graph and the access frequency of geographic locations to generate an effective margin used to adjust the decision boundary of the model; a joint optimization module for constructing semantic-driven positive and negative sample pairs for contrastive learning based on the multi-source semantic graph, and introducing the effective margin into the main prediction task to adjust the loss function, and training the prediction model by jointly optimizing the main prediction task and the contrastive learning task; a prediction execution module for using the trained model to predict the next visited geographic location of a user according to the user's historical moving trajectory.

10. The system of claim 9, wherein, The system is configured to be deployed in at least one of the following application scenarios: an intelligent traffic planning scenario, in which the system is deployed on a traffic management server to process vehicle trajectory data in real time to optimize signal timing and road network planning; a location service recommendation scenario, in which the system is integrated in a front-end application to provide personalized address recommendations based on real-time user trajectories; a city management scenario, in which the system is interfaced with a municipal Internet of Things platform to analyze the use efficiency of public facilities and optimize their layout; a public health monitoring scenario, in which the system is interfaced with a disease control center data system to construct a dynamic risk assessment model for epidemic transmission based on human movement patterns.

Images