A user area dynamic characterization method, apparatus, device, medium and product

Abstract

The application provides a user area dynamic characterization method and device, equipment, medium and product, relates to the technical field of artificial intelligence, and includes: acquiring original signaling data between a user equipment and a base station; processing the original signaling data to obtain user position information data; then constructing a multi-granularity user area dynamic graph, performing representation learning on the multi-granularity user area dynamic graph, obtaining multi-scale user dynamic characterization, performing cross-scale alignment optimization on the multi-scale user dynamic characterization, and obtaining aligned user dynamic characterization; the city-level user dynamic characterization is used to reflect the activity characteristics of a user corresponding to the user equipment in a city, and the city group-level user dynamic characterization is used to reflect the migration characteristics of the user across cities. Therefore, unified modeling and semantic alignment of user behavior under the city group background are achieved, and the expression ability and cross-city generalization performance of user behavior analysis are effectively improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, medium and product for dynamic representation of user areas. Background Technology

[0002] With the deep integration of mobile communication technology and smart city construction, utilizing massive user signaling data for trajectory modeling and behavior analysis has become a routine method for understanding urban operational patterns and optimizing the allocation of social resources. Currently, this field generally extracts user trajectory points (including user identifiers, base station locations, timestamps, etc.) based on base station connection records and uses machine learning or deep learning technologies to achieve specific tasks such as user behavior recognition and regional functional division, providing important data support for urban management.

[0003] Against this technological backdrop, researchers have proposed various advanced models to further explore the value of data. For example, some methods fuse traffic flow and user trajectory data through collaborative attention mechanisms to construct dynamic representations of road networks, serving macro-level traffic analysis; others employ dynamic hypergraph structures to model urban peak-hour travel demand, focusing on characterizing spatiotemporal dependencies and predicting specific scenarios; still others explore frameworks based on pre-training and transfer learning, aiming to solve the problem of traffic flow prediction in target cities with scarce data. These representative works have all promoted the development of intelligent analysis of urban spatiotemporal data.

[0004] However, the aforementioned approaches still have several limitations, restricting their effectiveness in more complex scenarios. First, existing methods are mostly focused on static or quasi-static modeling within cities at a single spatiotemporal scale, lacking the ability to characterize multi-layered and dynamic structural relationships across larger spatial areas such as urban agglomerations. This makes it difficult for the models to support structural generalization of cross-regional behavior. Second, most studies have not fully considered the correlation and differences in user behavior patterns across different spatial scales (e.g., between a single city and an urban agglomeration), lacking effective cross-scale behavior transfer and representation alignment mechanisms, thus limiting their application in predicting user behavior evolution. Third, some studies involving transfer learning rely too heavily on semantic data transfer (e.g., mask reconstruction), neglecting the inherent differences in spatial structures and behavioral patterns across different regions, resulting in insufficient adaptability of the models in transfer scenarios where structural characteristics need to be preserved. Finally, most existing solutions are specifically optimized for single tasks such as traffic flow prediction and peak demand estimation, lacking an integrated modeling framework compatible with multiple downstream tasks such as user profiling, behavior transfer, and policy regulation. The generalization and expansion capabilities of the models need improvement.

[0005] In summary, the main technical problems in urban spatiotemporal data analysis methods are: single modeling scale, lack of cross-scale alignment, insufficient structural transfer, and weak task generalization ability. Summary of the Invention

[0006] This application provides a method, apparatus, device, medium, and product for dynamic representation of user regions, in order to solve the technical problems that exist in urban spatiotemporal data analysis methods, such as single modeling scale, lack of cross-scale alignment, insufficient structural transfer, and weak task generalization ability.

[0007] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a method for dynamically representing user regions, the method comprising: Acquire raw signaling data between user equipment and base stations; The original signaling data is processed to obtain user location information data; A multi-granularity user region dynamic map is constructed based on the user location information data, and representation learning is performed on the multi-granularity user region dynamic map to obtain a multi-scale user dynamic representation, wherein the multi-scale user dynamic representation includes city-level user dynamic representation and city cluster-level user dynamic representation. Cross-scale alignment optimization is performed on the multi-scale user dynamic representation to obtain the aligned user dynamic representation.

[0008] Optionally, after performing cross-scale alignment optimization on the multi-scale user dynamic representation to obtain the aligned user dynamic representation, the method further includes: Obtain the task type of the target downstream task, and process the aligned user dynamic representation according to the task type to obtain task input features; The task input features are input into a task classifier for classification processing to obtain structured prediction results.

[0009] Optionally, the original signaling data includes a user identifier, a base station identifier, and connection time information between the user equipment and the base station; the processing of the original signaling data to obtain user location information data includes: Obtain the corresponding latitude and longitude coordinates of the base station based on the base station identifier, determine the grid area to which the base station belongs based on the base station latitude and longitude coordinates, and determine the city-level area identifier and city cluster-level area identifier of the base station based on the grid area; Based on the user identifier, the city-level region identifier, the city cluster-level region identifier, and the connection time information, an initial location information quadruple is constructed; wherein, the initial location information quadruple is used to describe the user's stay record in the region, and the region includes city-level regions and city cluster regions; The initial location information quadruple is cleaned and then classified into an intra-city trajectory subset and an inter-city trajectory subset. The intra-city trajectory subset and the inter-city trajectory subset are then identified as the user location information data.

[0010] Optionally, the step of constructing a multi-granularity user region dynamic map based on the user location information data, and performing representation learning on the multi-granularity user region dynamic map to obtain a multi-scale user dynamic representation includes: Construct a city-level dynamic map based on the subset of trajectories within the city; Construct a city cluster-level dynamic map based on the aforementioned intercity trajectory subset; The city-level dynamic map and the city cluster-level dynamic map are used as the multi-granularity user area dynamic map. The city-level dynamic graph is subjected to representation learning processing to obtain the city-level user dynamic representation; the city cluster-level dynamic graph is subjected to representation learning processing to obtain the city cluster-level user dynamic representation. The city-level user dynamic representation and the city cluster-level user dynamic representation are used as the multi-scale user dynamic representation.

[0011] Optionally, the step of performing representation learning processing on the city-level dynamic graph to obtain the city-level user dynamic representation includes: For the user nodes and region nodes in the city-level dynamic graph, initialize the user memory matrix and region memory matrix respectively; For each dwell event in the city-level dynamic map, the dwell duration, interaction interval, and regional attribute information of the dwell event are obtained. The dwell duration, interaction interval, and regional attribute information are encoded and fused to obtain an event feature embedding vector. The event feature embedding vector is subjected to context encoding to obtain a context-enhanced representation; Based on the context-enhanced representation, message aggregation is performed on the adjacency events of the user node and each node in the region node to obtain the local dynamic representation of the node, wherein the adjacency events are a set of dwell events associated with the node; Based on the local dynamic representation, the user memory matrix, and the region memory matrix, the memory vector of the corresponding node is updated to obtain the updated user memory vector and the updated region memory vector. The updated user memory vector is mapped to obtain the city-level user dynamic representation.

[0012] Optionally, the step of encoding and fusing the dwell time, the interaction interval, and the regional attribute information to obtain an event feature embedding vector includes: The dwell time is encoded using sine and cosine position coding to obtain the dwell time embedding; The interaction interval is encoded using sine and cosine position encoding to obtain the interaction interval embedding; The dwell time embedding and the interaction interval embedding are concatenated to obtain the time feature embedding; Based on the regional attribute information, feature encoding is performed to obtain the regional static features; The event feature embedding vector is obtained by concatenating the time feature embedding with the regional static feature and then fusing them.

[0013] Optionally, based on the context-enhanced representation, message aggregation is performed on the adjacency events between the user node and each node in the region nodes to obtain a local dynamic representation of the node, including: For the user node, based on the association relationship of the edges in the city-level dynamic graph, the most recent preset number of dwell events associated with the user node are obtained, and the obtained dwell events are used as the adjacency events of the user node to form a user adjacency event set. For the region node, based on the association relationship of the edges in the city-level dynamic graph, the most recent preset number of dwell events associated with the region node are obtained, and the obtained dwell events are used as the adjacency events of the region node to form a region adjacency event set. Obtain the context-enhanced representation corresponding to each adjacency event from the user adjacency event set and the region adjacency event set; The local dynamic representation of the node is obtained by aggregating the context-enhanced representations corresponding to each of the adjacent events.

[0014] Optionally, based on the local dynamic representation, the user memory matrix, and the region memory matrix, the memory vector of the corresponding node is updated to obtain the updated user memory vector and the updated region memory vector, including: For the user node, the user memory vector of the user node at the previous time step is obtained from the user memory matrix, and the local dynamic representation of the user node is updated with the user memory vector at the previous time step to obtain the updated user memory vector. For the region node, the region memory vector of the region node at the previous time step is obtained from the region memory matrix. The local dynamic representation of the region node and the region memory vector at the previous time step are input into the gated loop unit for update processing to obtain the updated region memory vector.

[0015] Optionally, cross-scale alignment optimization is performed on the multi-scale user dynamic representation to obtain an aligned user dynamic representation, including: User identifiers are determined from the intercity trajectory subset, and the users corresponding to the user identifiers are identified as target users with cross-city travel behavior, thus forming a target user set; Obtain the city-level user dynamic representation and the city cluster-level user dynamic representation corresponding to each target user in the target user set; Configure a first projection module and a second projection module. Input the city-level user dynamic representation into the first projection module for mapping processing to obtain the projected city-level representation vector; input the city cluster-level user dynamic representation into the second projection module for mapping processing to obtain the projected city cluster-level representation vector. Using each target user in the target user set as an anchor point, the projected city-level representation vector and the projected city cluster-level representation vector of the same target user are constructed as positive sample pairs; based on the user location information data, the city with the highest frequency of occurrence of the target user is determined as the active city of the target user; the projected city cluster-level representation vectors of other users are extracted from the active cities of the target user, and these are constructed as negative sample pairs with the projected city-level representation vector of the target user; Based on the comparison and learning between the positive sample pairs and the negative sample pairs, the parameters of the first projection module and the second projection module are optimized and updated. The first projection module, after optimizing and updating the input parameters of the city-level user dynamic representation, is mapped to obtain an aligned and optimized city-level representation vector. The aligned and optimized city-level representation vector is then used as the aligned user dynamic representation output, wherein the aligned user dynamic representation has both city-level and city cluster-level features.

[0016] Optionally, based on comparative learning between the positive sample pairs and the negative sample pairs, the parameters of the first projection module and the second projection module are optimized and updated, including: For the positive sample pair, calculate the cosine similarity between the projected city-level representation vector and the projected city cluster-level representation vector in the positive sample pair to obtain the positive sample similarity. For the negative sample pair, calculate the cosine similarity between the projected city-level representation vector of the target user and the projected city cluster-level representation vector of the other users in the negative sample pair to obtain the negative sample similarity set; Calculate the contrast loss value based on the positive sample similarity and the negative sample similarity set; The parameters of the first projection module and the second projection module are optimized and updated based on the contrast loss value, wherein the goal of the optimization and update is to maximize the similarity of the positive samples and minimize the similarity in the set of negative samples.

[0017] Secondly, embodiments of this application provide a user area dynamic characterization device, the device comprising: The acquisition module is used to acquire raw signaling data between user equipment and base station; The execution module is used to process the original signaling data to obtain user location information data; A multi-granularity user region dynamic map is constructed based on the user location information data, and representation learning is performed on the multi-granularity user region dynamic map to obtain a multi-scale user dynamic representation, wherein the multi-scale user dynamic representation includes city-level user dynamic representation and city cluster-level user dynamic representation. Cross-scale alignment optimization is performed on the multi-scale user dynamic representation to obtain the aligned user dynamic representation.

[0018] Thirdly, embodiments of this application provide a network device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of a user area dynamic representation method as described in the first aspect.

[0019] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a user region dynamic representation method as described in the first aspect.

[0020] Fifthly, embodiments of this application provide a computer program product, including computer instructions, which, when executed by a processor, implement the steps of a user region dynamic representation method as described in the first aspect.

[0021] In this embodiment, user location information data is obtained by acquiring and processing the raw signaling data between the user equipment and the base station, providing a standardized data foundation for subsequent dynamic graph construction. A multi-granularity user region dynamic graph is constructed based on the user location information data. By constructing dynamic graph structures at both the city and city cluster levels, the problem of existing technologies that only model single-scale graph structures at the city level and struggle to support cross-regional behavioral structure generalization is solved. Furthermore, representation learning is performed on the multi-granularity user region dynamic graph to obtain multi-scale user dynamic representations, including city-level and city cluster-level user dynamic representations. This enables city-level representations to capture fine-grained behavioral characteristics of users within a city, and city cluster-level representations to capture macro-level migration characteristics of users across cities.

[0022] Furthermore, by performing cross-scale alignment optimization on multi-scale user dynamic representations, aligned user dynamic representations are obtained, ensuring that the representations of the same user at different scales remain consistent in the shared semantic space. This solves the problem of representation splitting caused by existing technologies not considering the migration modeling of user behavior between different spatial scales, and improves the model's generalization ability for cross-regional user behavior.

[0023] In summary, by employing techniques such as multi-granularity dynamic graph construction, multi-scale representation learning, and cross-scale alignment optimization, unified modeling and semantic alignment of user behavior in the context of urban agglomerations have been achieved, effectively improving the expressive power and cross-city generalization performance of user behavior analysis. Attached Figure Description

[0024] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a user region dynamic representation method provided in this application embodiment; Figure 2 A flowchart illustrating a user region dynamic representation method provided in this application embodiment; Figure 3 A structural block diagram of a user area dynamic characterization device provided in an embodiment of this application; Figure 4 This is a structural block diagram of a network device provided in an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] The technical terms related to a user region dynamic representation method provided in the embodiments of this application are introduced below.

[0027] Dynamic Graph: A graph structure model that evolves over time to represent the dynamic changes of nodes and their relationships under time-series or event-driven conditions. It is widely used in social network modeling, traffic prediction, communication networks and other fields.

[0028] Representation learning: This involves automatically learning to extract high-quality latent feature representations from raw inputs to improve the model's generalization ability in downstream tasks. It is commonly used in natural language processing, graph neural networks, and computer vision.

[0029] Cross-scale alignment is a technique for establishing correspondences between multi-scale data or features, aiming to integrate information representations of different granularities and improve the model's ability to model structural hierarchy and scale changes.

[0030] Urban agglomeration: refers to a cluster of geographically adjacent, functionally complementary and economically closely linked cities within a specific region. It is an important organizational form for regional spatial development and resource synergy.

[0031] Info Noise Contrastive Estimation (InfoNCE) is a loss function based on contrastive learning. It enhances the discriminativeness of representations by maximizing the similarity between positive sample pairs and minimizing their similarity with negative samples. It is often used in self-supervised learning frameworks to construct effective feature representations.

[0032] Figure 1 This application illustrates a method for dynamically representing user regions according to an embodiment of the present application, such as... Figure 1 As shown, the method includes: Step S101: Obtain the raw signaling data between the user equipment and the base station; Step S102: Process the raw signaling data to obtain user location information data; Step S103: Construct a multi-granularity user region dynamic map based on user location information data, and perform representation learning on the multi-granularity user region dynamic map to obtain a multi-scale user dynamic representation. The multi-scale user dynamic representation includes city-level user dynamic representation and city cluster-level user dynamic representation. The city-level user dynamic representation is used to reflect the activity characteristics of the user corresponding to the user device within the city, and the city cluster-level user dynamic representation is used to reflect the migration characteristics of the user across cities. Step S104: Perform cross-scale alignment optimization on the multi-scale user dynamic representation to obtain the aligned user dynamic representation.

[0033] It should be noted that, Figure 1The user region dynamic representation method shown here processes raw signaling data to obtain user location information data, then constructs a multi-granularity user region dynamic map based on the user location information data and performs representation learning, thereby obtaining a multi-scale user dynamic representation that simultaneously covers the city and city cluster levels, and further performs cross-scale alignment optimization on this basis. Specifically, firstly, raw signaling data is acquired and converted into user location information data that can reflect the spatiotemporal behavior of users, so that subsequent modeling no longer stops at scattered connection records, but can form trajectory and regional association information that can be used for graph structure expression; on this basis, a multi-granularity user region dynamic map is constructed based on the user location information data, and representation learning is performed on the multi-granularity user region dynamic map to obtain a multi-scale user dynamic representation. Among them, the city-level user dynamic representation is used to characterize the dynamic behavior of users at the city-level regional identifier scale, and the city cluster-level user dynamic representation is used to characterize the cross-city behavior characteristics of users at the city cluster-level regional identifier scale. The two together constitute a complementary characterization of the user's spatiotemporal movement pattern. Furthermore, cross-scale alignment optimization is performed on the multi-scale user dynamic representation to obtain aligned user dynamic representation. This aligned user dynamic representation retains fine-grained information at the city level while absorbing cross-city pattern information carried by the regional identifier scale at the city cluster level, thus better adapting to the modeling needs of tasks with different granularities.

[0034] Therefore, by connecting multi-scale user dynamic representation with cross-scale alignment optimization, user dynamic representation can reflect both intra-city details and cross-city patterns, which is more conducive to downstream tasks obtaining stable and transferable task input features.

[0035] In one possible implementation, after performing cross-scale alignment optimization on the multi-scale user dynamic representation to obtain the aligned user dynamic representation, the method further includes: obtaining the task type of the target downstream task, processing the aligned user dynamic representation according to the task type to obtain task input features; and inputting the task input features into a task classifier for classification processing to obtain structured prediction results.

[0036] It should be noted that after performing cross-scale alignment optimization on multi-scale user dynamic representations to obtain aligned user dynamic representations, the method also includes: obtaining the task type of the target downstream task, processing the aligned user dynamic representations according to the task type to obtain task input features, and inputting the task input features into a task classifier to obtain structured prediction results. This allows the unified user dynamic representation capability to be naturally applied to specific tasks such as user commuting type recognition, intercity mobility detection, or regional function classification. This enables the application to flexibly adapt to various smart city analysis tasks such as user commuting type recognition, intercity mobility detection, and regional function classification, solving the problem that existing technologies optimize for single tasks but lack generalization ability.

[0037] In one possible implementation, the original signaling data includes a user identifier, a base station identifier, and connection time information between the user equipment and the base station. The original signaling data is processed to obtain user location information data, including: obtaining the corresponding base station latitude and longitude coordinates based on the base station identifier; determining the grid area to which the base station belongs based on the base station latitude and longitude coordinates; determining the city-level area identifier and city cluster-level area identifier of the base station based on the grid area; constructing an initial location information quadruple based on the user identifier, city-level area identifier, city cluster-level area identifier, and connection time information; cleaning the initial location information quadruple, classifying the cleaned initial location information quadruple into intra-city trajectory subsets and inter-city trajectory subsets, and determining the intra-city trajectory subsets and inter-city trajectory subsets as user location information data.

[0038] It should be noted that the city-level area identifier is the identifier information of the grid area obtained by grid coding the geospatial space using the first-granularity grid, and the city cluster-level area identifier is the identifier information of the grid area obtained by coding the geospatial space using the second-granularity grid. The first granularity is smaller than the second granularity. The initial location information quadruple is used to describe the user's stay record in the area, which includes: city-level areas and city cluster areas. The intra-city trajectory subset is the set of initial location information quadruples generated by the user's activities within the same city, and the inter-city trajectory subset is the set of initial location information quadruples generated by the user's activities across different cities.

[0039] It should be noted that the original signaling data includes user identifiers, base station identifiers (numbers), and connection time information between the user equipment and the base station. By introducing base station latitude and longitude coordinates and grid coding, the spatial location corresponding to the base station identifier is uniformly mapped to city-level and city cluster-level regional identifiers, thus ensuring that subsequent user trajectories have a consistent regional positioning reference at both scales. Specifically, after obtaining the corresponding base station latitude and longitude coordinates based on the base station identifier and performing grid coding, the city-level and city cluster-level regional identifiers become spatial indices that can be directly combined with the connection time information. These are then combined with the user identifier to construct an initial location information quadruple, so that each connection record is organized as a unified expression of who is in which city-level regional identifier and when, and which city cluster-level regional identifier. Subsequently, the initial location information quadruple is cleaned and divided into intra-city trajectory subsets and inter-city trajectory subsets according to the trajectory range. The intra-city trajectory subset more centrally reflects continuous activity at the city-level regional identifier scale, while the inter-city trajectory subset more centrally reflects cross-city activity at the city cluster-level regional identifier scale. Both are then used to determine the user location information data.

[0040] This processing method, which transforms raw signaling data into user location information data, ensures the consistency of regional representation scale and provides structured, decomposable trajectory input for the subsequent construction of multi-granularity user region dynamic maps. This gives the user region dynamic representation method a clear modeling foundation at the data entry stage.

[0041] In one possible implementation, a multi-granularity user region dynamic map is constructed based on user location information data, and representation learning is performed on the multi-granularity user region dynamic map to obtain a multi-scale user dynamic representation, including: Construct a city-level dynamic map based on a subset of trajectories within the city; Construct a dynamic map of city clusters based on a subset of intercity trajectories; The city-level dynamic map and the city cluster-level dynamic map are used as multi-granularity user area dynamic maps. Representation learning is performed on city-level dynamic graphs to obtain city-level user dynamic representations; representation learning is also performed on city cluster-level dynamic graphs to obtain city cluster-level user dynamic representations. City-level user dynamic representation and city cluster-level user dynamic representation are used as multi-scale user dynamic representation.

[0042] It should be noted that constructing a city-level dynamic graph based on a subset of city trajectories specifically includes: extracting user identifiers and corresponding city-level region identifiers from the subset of city trajectories; using user identifiers as user nodes and city-level region identifiers as region nodes; determining dwell events based on dwell records in the subset of city trajectories; and constructing edges between user nodes and region nodes based on dwell events to obtain the city-level dynamic graph.

[0043] Constructing a city cluster-level dynamic graph based on a subset of intercity trajectories specifically includes: extracting user identifiers and corresponding city cluster-level regional identifiers from the subset of intercity trajectories; using user identifiers as user nodes and city cluster-level regional identifiers as regional nodes; determining dwell events based on dwell records in the subset of intercity trajectories; and constructing edges between user nodes and regional nodes based on dwell events to obtain a city cluster-level dynamic graph.

[0044] For example, a pre-configured city-level representation learning model can be used to perform representation learning processing on a city-level dynamic graph to obtain a city-level user dynamic representation; a pre-configured city cluster-level representation learning model can be used to perform representation learning processing on a city cluster-level dynamic graph to obtain a city cluster-level user dynamic representation.

[0045] It should be noted that, firstly, user identifiers and corresponding city-level region identifiers are extracted from the intra-city trajectory subset. User identifiers are used as user nodes and city-level region identifiers are used as region nodes. Then, stay events are determined based on the stay records in the intra-city trajectory subset. Finally, edges are constructed between user nodes and region nodes based on the stay events, thus obtaining a city-level dynamic graph that reflects the temporal stay relationships within the city. Simultaneously, user identifiers and corresponding city cluster-level region identifiers are extracted from the inter-city trajectory subset. User identifiers are used as user nodes and city cluster-level region identifiers are used as region nodes. Similarly, stay events are determined based on the stay records in the inter-city trajectory subset, and edges are constructed, thus obtaining a city cluster-level dynamic graph that reflects the cross-city stay relationships. Subsequently, the city-level dynamic graph and the city cluster-level dynamic graph together constitute a multi-granularity user region dynamic graph, which is then input into the pre-configured city-level representation learning model and city cluster-level representation learning model for representation learning processing. The city-level representation learning model includes a first event encoding module, a first message aggregation module, a first memory update module, and a first representation generation module connected in sequence. The city cluster-level representation learning model includes a second event encoding module, a second message aggregation module, a second memory update module, and a second representation generation module connected in sequence. The parameters of the city-level representation learning model and the city cluster-level representation learning model are independent of each other, so that the city-level user dynamic representation and the city cluster-level user dynamic representation can learn their respective scale rules under the same module logic without interfering with each other.

[0046] The resulting multi-scale user dynamic representation retains both the fine-grained activity trajectory information of the city-level dynamic map and the cross-city movement structure information of the city cluster-level dynamic map, providing an alignable and complementary dual-scale representation basis for subsequent cross-scale alignment optimization.

[0047] In one possible implementation, representation learning is performed on the city-level dynamic graph to obtain a city-level user dynamic representation, including: For user nodes and region nodes in the city-level dynamic graph, initialize the user memory matrix and region memory matrix respectively; For each dwell event in the city-level dynamic map, the dwell time, interaction interval and regional attribute information of the dwell event are obtained. The dwell time, interaction interval and regional attribute information are encoded and fused to obtain the event feature embedding vector. The event features are embedded into a vector and then subjected to context encoding to obtain a context-enhanced representation. Based on the context-enhanced representation, the adjacency events of each node in the user node and the region node are aggregated to obtain the local dynamic representation of the node, where the adjacency event is the set of dwell events associated with the node; Based on the local dynamic representation and the user memory matrix and the region memory matrix, the memory vector of the corresponding node is updated to obtain the updated user memory vector and the updated region memory vector. The updated user memory vector is mapped to obtain a city-level dynamic representation of users.

[0048] It should be noted that, for example, a pre-configured city-level representation learning model can be used to perform representation learning processing on the city-level dynamic graph to obtain a city-level user dynamic representation. The key lies in constructing a continuous link around the dwell event, from the first event encoding module to the first message aggregation module, the first memory update module, and then to the first representation generation module, and using the user memory matrix and the region memory matrix as dynamic state carriers, so that the evolution information of user nodes and region nodes can be continuously accumulated and updated as the dwell event progresses.

[0049] Specifically, user memory matrices and regional memory matrices are initialized for user nodes and regional nodes in the city-level dynamic graph, respectively, so that the memory vectors of different nodes have independent temporal carrying capacity. Then, for each dwell event, the dwell duration, interaction interval, and regional attribute information of the dwell event are extracted, and encoded and fused by the first event encoding module to obtain an event feature embedding vector. The event feature embedding vector is then input into the encoder in the first event encoding module to obtain a context-enhanced representation, so that the representation of a single dwell event includes not only its own attributes but also the context of its association with other dwell events. On this basis, the first message aggregation module performs message aggregation on the adjacent events of user nodes and regional nodes according to the context-enhanced representation to obtain a local dynamic representation, so that the node representation can absorb the temporal dynamic information carried by the set of dwell events associated with it. Further, the first memory update module updates the memory vector of the corresponding node by combining the local dynamic representation with the user memory matrix and the regional memory matrix to obtain the updated user memory vector and the updated regional memory vector. Finally, the first representation generation module maps the updated user memory vector to obtain the city-level user dynamic representation.

[0050] Through this representation learning mechanism driven by dwell events and supported by a memory matrix, the dynamic representation of city-level users can more stably depict the dynamic dwell patterns and temporal evolution characteristics of users at the city-level regional identifier scale.

[0051] In one possible implementation, the dwell time, interaction interval, and regional attribute information are encoded and fused to obtain an event feature embedding vector. This includes: performing sine and cosine position encoding on the dwell time to obtain a dwell time embedding; performing sine and cosine position encoding on the interaction interval to obtain an interaction interval embedding; concatenating the dwell time embedding and the interaction interval embedding to obtain a time feature embedding; performing feature encoding based on regional attribute information to obtain regional static features; and concatenating and fusing the time feature embedding and the regional static features to obtain an event feature embedding vector.

[0052] Specifically, the time feature embedding can be concatenated with the regional static feature and then input into the feedforward neural network layer in the first event encoding module for fusion processing to obtain the event embedding vector.

[0053] It should be noted that the dwell time is encoded using sine and cosine positions to obtain the dwell time embedding, and the interaction interval is encoded using sine and cosine positions to obtain the interaction interval embedding. The dwell time embedding and the interaction interval embedding are then concatenated to form the temporal feature embedding, giving the temporal characterization of the dwell event a unified encoding structure. Simultaneously, regional static features are obtained through feature encoding based on regional attribute information. The temporal feature embedding and the regional static features are then concatenated and input into the feedforward neural network layer in the first event encoding module for fusion processing, resulting in the event feature embedding vector. This allows the fused temporal and regional attribute information to become event-level inputs that can be used for subsequent context encoding and message aggregation. Through this encoding fusion method, the event feature embedding vector can more fully express the temporal structure and regional attribute background of the dwell event, thus providing a more solid foundation for context-enhanced representation and local dynamic representation.

[0054] In one possible implementation, the event feature embedding vector is subjected to context encoding to obtain a context-enhanced representation, including: inputting the event feature embedding vector into a multi-layer Transformer encoder in the first event encoding module; and capturing the temporal dependencies between different dwell events through a multi-head attention mechanism in the multi-layer Transformer encoder to obtain a context-enhanced representation.

[0055] It should be noted that the event features are embedded into the vector input to the encoder in the first event encoding module and then subjected to context encoding processing by a multi-layer Transformer encoder to obtain a context-enhanced representation. The key is to use the multi-head attention mechanism in the multi-layer Transformer encoder to capture the temporal dependencies between different dwell events, so that the representation of each dwell event is no longer an isolated static vector, but is enhanced by the correlation information of its preceding and following dwell events.

[0056] Because dwell events in city-level dynamic graphs inherently possess differences in temporal order and interaction intervals, multi-head attention mechanisms can assign different attention weights to different dwell events during context encoding. This injects temporal dependencies into the context-enhanced representation in a learnable attention distribution form, enabling the context-enhanced representation to simultaneously reflect the temporal feature embedding and regional static features of the dwell event itself, as well as its relative temporal relationship and dependency strength with other dwell events. The context-enhanced representation formed through multi-layer Transformer encoders provides a more contextual information source for the subsequent first message aggregation module to aggregate at the adjacent event level, thus helping the local dynamic representations of user nodes and regional nodes to more accurately reflect the true dwell event sequence structure.

[0057] In one possible implementation, message aggregation is performed on the adjacency events of each node in the user node and the region node based on the context-enhanced representation to obtain the local dynamic representation of the node. This includes: for the user node, based on the edge association relationships in the city-level dynamic graph, obtaining the most recent preset number of dwell events associated with the user node, and using the obtained dwell events as the adjacency events of the user node to form a user adjacency event set; for the region node, based on the edge association relationships in the city-level dynamic graph, obtaining the most recent preset number of dwell events associated with the region node, and using the obtained dwell events as the adjacency events of the region node to form a region adjacency event set; obtaining the context-enhanced representation corresponding to each adjacency event from the user adjacency event set and the region adjacency event set; and performing an aggregation operation on the context-enhanced representation corresponding to each adjacency event to obtain the local dynamic representation of the node.

[0058] It should be noted that the key to obtaining the local dynamic representation of a node by aggregating the adjacency events of user nodes and region nodes based on the context-enhanced representation is to limit the set of dwell events associated with a node to the most recent preset number of dwell events, and to perform the aggregation operation on the context-enhanced representation within the scope of this set of adjacency events. This keeps the computation and modeling focused on the recent dynamics while avoiding the accumulation of noise caused by the unbounded event set.

[0059] Specifically, for user nodes, the system obtains the most recent preset number of dwell events associated with user nodes based on the edge relationships in the city-level dynamic graph and forms a user adjacency event set. For region nodes, the system similarly obtains the most recent preset number of dwell events associated with region nodes based on the edge relationships and forms a region adjacency event set. Then, the system obtains the context-enhanced representations corresponding to each adjacency event from the user adjacency event set and the region adjacency event set, and performs an aggregation operation on these context-enhanced representations to obtain the local dynamic representation of the node. This local dynamic representation can centrally reflect the dynamic interaction patterns and temporal context changes of the node in the most recent dwell event sequence.

[0060] Through this message aggregation mechanism, the local dynamic representation is closer to the current behavior state under the window constraint of the most recent preset number of dwell events, thus providing a more representative dynamic input for the memory vector update of the first memory update module.

[0061] In one possible implementation, the memory vector of the corresponding node is updated based on the local dynamic representation, the user memory matrix, and the region memory matrix to obtain the updated user memory vector and the updated region memory vector. This includes: for user nodes, obtaining the user memory vector of the user node at the previous time step from the user memory matrix, and inputting the local dynamic representation of the user node and the user memory vector at the previous time step into the gated loop unit in the first memory update module for update processing to obtain the updated user memory vector; for region nodes, obtaining the region memory vector of the region node at the previous time step from the region memory matrix, and inputting the local dynamic representation of the region node and the region memory vector at the previous time step into the gated loop unit for update processing to obtain the updated region memory vector.

[0062] It should be noted that the first memory update module can update the memory vector of the corresponding node according to the local dynamic representation and the user memory matrix and the regional memory matrix to obtain the updated user memory vector and the updated regional memory vector. Its core is to use a gated loop unit to establish a controllable information fusion channel between the "memory vector of the previous moment" and the "current local dynamic representation", so that the node memory can both retain historical accumulation and absorb the influence of newly occurring stay events.

[0063] For user nodes, the user memory vector of the user node at the previous time step is obtained from the user memory matrix, and the local dynamic representation of the user node and the user memory vector at the previous time step are input into the gated loop unit in the first memory update module for update processing to obtain the updated user memory vector; for region nodes, the region memory vector of the region node at the previous time step is obtained from the region memory matrix, and the local dynamic representation of the region node and the region memory vector at the previous time step are input into the gated loop unit for update processing to obtain the updated region memory vector.

[0064] Through this update method driven by gated loop units, the user memory matrix and the regional memory matrix can continuously evolve with the stay events, making the city-level user dynamic representation obtained by the subsequent first representation generation module more time-series continuous and historically consistent, thus better reflecting the dynamic behavioral trajectory characteristics of users in the city-level dynamic map.

[0065] In one possible implementation, cross-scale alignment optimization is performed on the multi-scale user dynamic representation to obtain the aligned user dynamic representation, including: User identifiers are determined from the intercity trajectory subset, and the users corresponding to the user identifiers are identified as target users with cross-city travel behavior, thus forming a target user set. Obtain the city-level and city cluster-level dynamic representations of each target user in the target user set; Configure a first projection module and a second projection module. Input the city-level user dynamic representation into the first projection module for mapping processing to obtain the projected city-level representation vector; input the city cluster-level user dynamic representation into the second projection module for mapping processing to obtain the projected city cluster-level representation vector. Using each target user in the target user set as an anchor point, construct positive sample pairs by projecting the city-level representation vector and the city cluster-level representation vector of the same target user; determine the city with the highest frequency of occurrence of the target user as the active city of the target user based on user location information data, extract the projected city cluster-level representation vector of other users from the active city of the target user, and construct negative sample pairs by projecting the city-level representation vector of the target user. Based on the comparative learning of positive and negative sample pairs, the parameters of the first and second projection modules are optimized and updated. The first projection module, after optimizing and updating the input parameters of the city-level user dynamic representation, is mapped to obtain the aligned and optimized city-level representation vector. The aligned and optimized city-level representation vector is then used as the aligned user dynamic representation output. The aligned user dynamic representation combines city-level features and city cluster-level features.

[0066] It should be noted that the core of cross-scale alignment optimization of multi-scale user dynamic representations to obtain aligned user dynamic representations is to use target users with cross-city travel behavior as alignment anchors. The first projection module and the second projection module map city-level user dynamic representations and city cluster-level user dynamic representations to a comparable vector space. Positive sample pairs and negative sample pairs are constructed for comparative learning optimization, so that the representations of the two scales are more consistent in the same user dimension and more distinguishable in different user dimensions.

[0067] First, target users with cross-city travel behavior are identified and a target user set is formed based on the number of city areas traversed by each user's trajectory in the user location information data. Then, city-level and city cluster-level user dynamic representations corresponding to each target user in the target user set are obtained and input into the first and second projection modules respectively to obtain projected city-level and city cluster-level representation vectors. Subsequently, positive sample pairs are constructed using the two types of projection vectors of the same target user, and negative sample pairs are extracted using the active cities of the target user as the extraction range. Projected city cluster-level representation vectors of other users are extracted from the active cities and projected city-level representation vectors of the target user to construct negative sample pairs. This ensures that the alignment optimization not only emphasizes the cross-scale consistency of the same user, but also explicitly suppresses confusion with other users. After the comparative learning is completed and the parameters of the first and second projection modules are optimized and updated, the city-level user dynamic representation is input into the first projection module with optimized parameters to obtain the aligned and optimized city-level representation vector, which is then output as the aligned user dynamic representation. This achieves that the aligned user dynamic representation has both city-level and city cluster-level features.

[0068] Through the cross-scale alignment optimization mechanism, the aligned user dynamic representation can absorb urban cluster-level structural information while maintaining city-level fine-grained expression, making the input features of subsequent tasks more consistent in cross-scale semantics and more conducive to generalization.

[0069] In one possible implementation, the parameters of the first and second projection modules are optimized and updated based on comparative learning using positive and negative sample pairs. This includes: for positive sample pairs, calculating the cosine similarity between the projected city-level representation vector and the projected city cluster-level representation vector in the positive sample pair to obtain the positive sample similarity; for negative sample pairs, calculating the cosine similarity between the projected city-level representation vector of the target user and the projected city cluster-level representation vector of other users in the negative sample pair to obtain the negative sample similarity set; calculating a contrastive loss value (e.g., based on the InfoNCE loss function) based on the positive and negative sample similarity sets; and optimizing and updating the parameters of the first and second projection modules based on the contrastive loss value, wherein the optimization and update objective is to maximize the positive sample similarity and minimize the similarity in the negative sample similarity set.

[0070] It should be noted that when comparing and learning based on positive and negative sample pairs and optimizing and updating the parameters of the first and second projection modules, a directly optimizable comparative loss value is formed by using cosine similarity and the InfoNCE loss function. This makes the goal of "the cross-scale representation of the same target user should be similar, and the representation of different users should be separated" a clear direction for numerical optimization.

[0071] Specifically, for positive sample pairs, the cosine similarity between the projected city-level representation vector and the projected city cluster-level representation vector is calculated to obtain the positive sample similarity. For negative sample pairs, the cosine similarity between the projected city-level representation vector of the target user and the projected city cluster-level representation vector of other users is calculated to obtain the negative sample similarity set. Then, based on the positive sample similarity and the negative sample similarity set, the contrast loss value is calculated using the InfoNCE loss function. The optimization update objective is to maximize the positive sample similarity and minimize the similarity in the negative sample similarity set, thereby driving the first projection module and the second projection module to gradually bring the cross-scale projection vectors of the same user closer and push the projection vectors of different users further apart during the training process.

[0072] By using this comparative learning approach based on cosine similarity and the InfoNCE loss function, cross-scale alignment optimization has quantifiable and repeatable optimization criteria, making the aligned user dynamic representations more consistent and discriminative within the similarity space.

[0073] In one possible implementation, the task type of the target downstream task is obtained, and the aligned user dynamic representation is processed according to the task type to obtain task input features; the task input features are then input into a task classifier for classification processing to obtain structured prediction results, including: The dynamic representations of city-level users, the dynamic representations of city clusters, and the aligned dynamic representations of users are structured and organized to obtain a multi-granularity representation set. Obtain the task type of the target downstream task, and select the corresponding scale representation from the multi-granularity representation set as the task input feature according to the requirements of the task type, or splice and fuse the representations of multiple scales in the multi-granularity representation set as the task input feature. The task input features are input into the task classifier for classification processing to obtain structured prediction results. The task classifier is a multilayer perceptron structure that is independent of the city-level representation learning model and the city cluster-level representation learning model. Among them, the target downstream tasks include user commuting type identification, intercity mobility detection, or regional function classification.

[0074] It should be noted that after obtaining the task type of the target downstream task, the city-level user dynamic representation, the city cluster-level user dynamic representation, and the aligned user dynamic representation are structured and organized to obtain a multi-granularity representation set. According to the requirements of the task type, the representation of the corresponding scale is selected from the multi-granularity representation set as the task input feature, or the representations of multiple scales in the multi-granularity representation set are spliced ​​and fused as the task input feature. The task input feature is then input into the task classifier to obtain the structured prediction result. The task classifier is a multilayer perceptron structure independent of the city-level representation learning model and the city cluster-level representation learning model.

[0075] The key to this approach lies in unifying multi-scale user dynamic representations and aligned user dynamic representations into a multi-granularity representation set. This allows different task types to access the most suitable scale information as needed. For example, user commuting type identification may rely more on the intra-city patterns reflected by city-level user dynamic representations, while inter-city mobility detection may rely more on the cross-city patterns reflected by city cluster-level user dynamic representations. In certain scenarios, regional function classification may be more suitable for using aligned user dynamic representations to simultaneously utilize both city-level and city cluster-level features. Furthermore, since the task classifier is an independent multilayer perceptron structure, its input interface can naturally accommodate two task input feature construction methods: "selecting a single-scale representation" or "splitting and fusing multiple-scale representations." This transforms the multi-granularity representation set into structured prediction results that can be directly used for classification inference.

[0076] Through this task-type driven representation selection and fusion mechanism, multi-scale user dynamic representations can more flexibly serve different downstream tasks and output structured prediction results in a consistent form under a unified framework.

[0077] In one possible implementation, before inputting the task input features into a task classifier for classification and obtaining structured prediction results, the method also includes fine-tuning the task classifier: Obtain labeled sample data corresponding to the target downstream task based on the task type. The labeled sample data includes sample features and sample labels. While keeping the parameters of the city-level representation learning model and the city cluster-level representation learning model unchanged, the task input features and the sample features in the labeled sample data are input into the task classifier to obtain the classification output. Calculate the supervised loss based on the classification output and the sample labels in the labeled sample data; The parameters of the task classifier are updated based on the supervised loss to adapt to the target downstream task.

[0078] It should be noted that before inputting the task input features into the task classifier to obtain structured prediction results, the task classifier is further fine-tuned to make it more closely match the label distribution and discrimination boundary of the target downstream task, while keeping the parameters of the city-level representation learning model and the city cluster-level representation learning model unchanged, thereby avoiding unnecessary changes in the multi-scale user dynamic representation during the fine-tuning process.

[0079] This method obtains labeled sample data corresponding to the target downstream task based on the task type. The labeled sample data includes sample features and sample labels. While keeping the parameters of the city-level representation learning model and the city cluster-level representation learning model unchanged, the task input features and the sample features in the labeled sample data are input into the task classifier to obtain the classification output. Then, the supervision loss is calculated based on the classification output and the sample labels, and the parameters of the task classifier are updated according to the supervision loss, so that the task classifier learns a classification mapping relationship that is more in line with the target downstream task within a fixed representation space.

[0080] By using this fine-tuning strategy of "fixing multi-scale representations and updating only the task classifier", we can improve the adaptability of the task classifier to specific task types, and maintain the stability of the aligned user dynamic representations when reused across different tasks, making the structured prediction results more in line with the actual needs of the target downstream tasks.

[0081] In this embodiment, user location information data is obtained by acquiring and processing raw signaling data, providing a standardized data foundation for subsequent dynamic graph construction. A multi-granularity user region dynamic graph is constructed based on the user location information data. By constructing dynamic graph structures at both the city and city cluster levels, the problem of existing technologies that only model single-scale graph structures at the city level and struggle to support cross-regional behavioral structure generalization is solved. Furthermore, representation learning is performed on the multi-granularity user region dynamic graph to obtain multi-scale user dynamic representations, including city-level and city cluster-level user dynamic representations. This enables city-level representations to capture fine-grained behavioral characteristics of users within a city, and city cluster-level representations to capture macro-level migration characteristics of users across cities.

[0082] Furthermore, by performing cross-scale alignment optimization on multi-scale user dynamic representations to obtain aligned user dynamic representations, the representations of the same user at different scales can remain consistent in a shared semantic space. This solves the representation split problem caused by existing technologies not considering the migration modeling of user behavior across different spatial scales, and improves the model's generalization ability for cross-regional user behavior. Finally, by obtaining the task type of the target downstream task and processing the aligned user dynamic representations according to the task type to obtain task input features, and then inputting the task input features into the task classifier for classification processing to obtain structured prediction results, this application can flexibly adapt to various smart city analysis tasks such as user commuting type recognition, intercity mobility detection, and regional function classification, solving the problem that existing technologies optimize for single tasks but lack generalization ability.

[0083] In summary, by employing techniques such as multi-granularity dynamic graph construction, multi-scale representation learning, and cross-scale alignment optimization, we have achieved unified modeling and semantic alignment of user behavior in the context of urban agglomerations. This effectively enhances the expressive power and cross-city generalization performance of user behavior analysis and supports flexible adaptation to various downstream tasks.

[0084] This paper introduces a method for dynamic representation of user regions from the perspective of specific application scenarios. This application provides a method for dynamic representation of user regions in urban agglomerations based on dynamic graph learning and cross-scale alignment. It aims to achieve unified modeling and semantic alignment of user dynamic behavior in multi-urban agglomeration, large-scale user signaling data environments, thereby improving the expressive power and generalization performance of user behavior analysis. This technical solution is a methodological approach that combines key modules such as signaling data construction, dynamic graph modeling, representation learning, and comparative optimization to form a complete processing flow. The following are the specific implementation steps and detailed explanations of each part (for reference). Figure 2 ).

[0085] Step S1: Using the raw signaling data as input, complete the regional gridding encoding and data cleaning process, constructing a quadruple dataset containing user identifier, region identifier, and timestamp, and outputting intra-city and inter-city trajectory subsets. The Google S2 spatial partitioning system is used to grid the region, with Level 16 for the city level and Level 12 for the city cluster level. The signaling data is cleaned by using temporal interpolation to fill in missing fields, employing a sliding window mechanism to identify and merge pseudo-stop points caused by the "ping-pong effect," and filtering abnormal trajectory jump points based on the principle of spatiotemporal continuity. Three types of intra-city subsets and one type of inter-city subset are constructed, outputting standardized user spatiotemporal quadruple datasets.

[0086] Google S2 is an open-source spatial indexing system developed by Google for efficiently dividing the Earth's surface (not just flat maps) into grids of varying precision. In the Google S2 system, Level represents the precision of the grid division.

[0087] Step S2: Construct a two-tiered (city-level and city cluster-level) user-region bipartite dynamic graph with city-level and city cluster-level granularity. , where nodes Represents a user or region, edge This represents the user's stay events within a region. A representation learning model is deployed on the graph structure, and instantiated as a city-level representation learning model and a city-group-level representation learning model (parameter-independent) on the city-level graph structure and the city-group-level graph structure, respectively. The model consists of event encoding, message aggregation, memory updating, and representation generation modules. First, the method initializes memory matrices for both the user and the region, embedding them with temporal features, stay duration, interaction interval, and region attributes. Then, based on a framework of Transformer encoder and GRU (Gated Recurrent Unit) memory units, it captures the temporal dependencies of user behavior and the dynamic evolution of region states. Predicting the user's next hop region is used as a self-supervised task to optimize the model, outputting dynamic representations of users and regions at different levels.

[0088] Step S3: Input the user dynamic representations obtained from the city-level and city cluster-level dynamic maps, and map them to a shared semantic space using a Multi-Layer Perceptron (MLP). Construct positive sample pairs for the same user at different scales, and simultaneously construct negative sample pairs for other users. Positive sample pairs are defined as the two dynamic representations learned by the same user in the city-level and city cluster-level dynamic maps, while negative samples are the user representations of other users within the same geographical area (such as the same city) in the city cluster-level dynamic map. A contrastive learning mechanism is employed, and the representation space is optimized based on the InfoNCE loss function to output cross-scale aligned user dynamic representations.

[0089] Step S4: Using the aligned and optimized user and region representations, along with different representations, as input, adapt them to downstream tasks such as commuting identification and travel analysis, outputting structured prediction results. This solution employs a flexible and multi-task-supporting adaptive framework. Specifically, it can dynamically select or fuse different scales (city-level, city cluster-level) and aligned user or region representations according to the specific needs of downstream tasks.

[0090] In one embodiment, step S1 includes: constructing a cleaned multi-level user location information quadruple dataset suitable for dynamic graph modeling based on the original signaling data, specifically including:

[0091] Step S11: Base station mapping and area coding. In this embodiment, the original signaling data is represented in the form of the following four-tuple: Where u represents the user ID and c is the base station number. and These represent the start and end times of the connection. First, obtain the latitude and longitude corresponding to base station c. The data was then grid-coded based on the Google S2 spatial partitioning system, dividing the data into regions according to granularity. Urban areas used Level 16 grid coding, while urban cluster areas used Level 12 grid coding. This yielded the grid identifier for the base station's region, and the user data was organized into initial location information quadruples. , where r represents the grid ID after dividing the region.

[0092] Step S12: Perform data cleaning on the quadruple set, mainly by completing the following operations: First, use temporal interpolation and sliding aggregation strategies to complete missing fields; then, use a sliding window mechanism to identify and process "ping-pong effect" behavior and extract stable dwell areas; at the same time, combine the time and space information of the trajectory to detect and remove jump points and abnormal trajectory segments; finally, filter abnormal records with drastic displacement or large-scale switching in a short period of time to improve the accuracy and timeliness of trajectory data.

[0093] Step S13: Constructing intra-city and inter-city signaling data subsets: Based on the processed trajectory data, classify according to city area grid identifiers: If the user trajectory is continuous, stable, and has sufficient dwell time within the same city grid (i.e., the time interval between trajectory points is short, the spatial offset is small, and there is a long dwell time in the area), it is classified as an intra-city trajectory; if the user trajectory exhibits jumping behavior across different city areas and has obvious dwell records in multiple cities, it is classified as an inter-city trajectory. Each type of trajectory data is ultimately uniformly formatted into the following set of structured quadruples.

[0094] .

[0095] Where u is the user ID, and r is the region ID mapped by the Google S2 spatial partitioning system. and These represent the user's entry and exit times in the area, respectively, and N is the total number of events in which the user interacts with the area.

[0096] In one embodiment, step S2 above, which involves constructing dynamic graphs at the city-level and city cluster-level granularities and performing dynamic representation learning, specifically includes the following sub-steps.

[0097] Step S21: Graph Structure Construction and Task Partitioning. This step aims to construct the basic structure required for dynamic graph representation based on the user-region interaction quadruple data constructed in Step S1. Independent dynamic graph structures are constructed for each city Ci and the entire city group Cgroup. This is used to capture the evolution of user interactions with geographic regions over time. Where: time t represents the current time step; V represents the user node. uWith regional nodes r The set; E(t) represents the user's dwell time event in the region, i.e., the user's stay time event in the region. u With the region r The interaction events that occur between them are in the form of...

[0098] .

[0099] Step S22: To support the dynamic state maintenance mechanism of the dynamic graph, the method initializes memory matrices for users and regions respectively on the city-level graph structure and the city cluster-level graph structure. User memory unit initialization matrix: ,in Indicates the total number of users. This represents the dimension of the user memory vector. Region memory unit initialization matrix: Where Nr represents the total number of region nodes, and dr is the dimension of the region memory vector. The initialization process of the memory matrix does not depend on the graph structure construction or encoding results in the preceding steps, but is independently initialized based on the set information of users and region nodes, for continuous state updates during subsequent dynamic evolution.

[0100] Step S23: Event Encoding and Context Feature Extraction. For each user-area interaction event... That is, the edge set in step S21 For each individual event edge in the event encoding module, its event feature representation is constructed. The input to the event encoding module includes three types of information: temporal feature embedding, including dwell time. The last time interval between the user and this area Regional static characteristics: including static coding of region type, administrative affiliation, latitude and longitude coordinates, etc. ; Time context embedding.

[0101] .

[0102] in, and Sine and cosine position coding is used to encode continuous-time features using learnable or preset time-frequency parameters. and Mapped to vector space, concatenated with the region's static features, and then input into the FFN (feedforward neural network layer) to obtain... .

[0103] The above information is fused to generate an embedding vector. And input it into a stacked L-layer Transformer encoder to obtain a context-enhanced representation.

[0104] .

[0105] Among them, the Transformer Module default number of layers L = 8 Each layer contains H = 16 Each attention head, with the hidden dimension set to... The above parameter configuration comprehensively references the general experience of existing multi-head attention mechanisms in temporal modeling tasks, and has been verified in practice to have good performance and convergence effects in the application scenarios of this application, and is therefore used as the default setting. This module can capture the temporal dependencies and contextual semantic relationships between different events. Users and region nodes will respectively collect their related event representations for use by the subsequent message aggregation module.

[0106] Step S24: Message Aggregation and Representation Generation. At the current time step t, the message aggregation module performs message aggregation on each node. (Including user nodes and region nodes) perform adjacency event information aggregation operations to capture the evolution trend of the local dynamic graph structure. Specifically, the set of adjacency events for node v is first determined. This set represents all relevant events that have occurred with node v before time t: .

[0107] For user node u, its most recently accessed region interaction events (in one embodiment, this number is set to 10) are used to form its adjacency event set. For a region node r, the interaction events of several (in one embodiment, this number is set to 10) different users who recently visited the region are taken to form its adjacency event set. Then, the context representation of each adjacent event. The data is input into the message aggregation function to generate a local dynamic representation of the current node. : .

[0108] in, The aggregate operation function is when User nodes are denoted as ,when When it is a region node, it is denoted as This application employs, but is not limited to, mean pooling as the message aggregation function. This module aims to integrate the influence of neighboring nodes on the current node's state, providing time-driven input features for the subsequent memory update module.

[0109] Step S25: Memory Update Module. The memory update module uses a gated recurrent unit (GRU) to update the memory vector of each node. The update formula for user node u is: .

[0110] The initial value of the user node's memory state at the initial moment is the value described in step S22. GRU stands for Gated Cyclic Unit.

[0111] Regional nodes The update formula is: .

[0112] The initial value of the region node memory state at the initial moment is the value described in step S22. GRU stands for Gated Cyclic Unit.

[0113] All update operations only apply to nodes that actually interact in the current time step; nodes that do not interact retain their previous state.

[0114] Step S26: Self-supervised learning and representation learning model optimization. This embodiment introduces a self-supervised task to drive the representation learning process. The target task is to predict the user's next-hop region at the current time. For each user node u, predict its next access region. : .

[0115] in, This represents the predicted probability that, given the current memory state of user node u, its next hop candidate region is . The standard normalization function ensures that the sum of the access probabilities of all candidate regions is 1; This represents the inner product similarity matrix between the user's memory vector and all region memory vectors, and the model is optimized using cross-entropy loss.

[0116] .

[0117] in, This represents the self-monitored loss value. This represents the predicted probability that the next hop of a user node is a real region. The goal of minimizing this loss function is to improve the model's ability to distinguish the user's actual travel trajectory, thereby optimizing the dynamic representation of users and regions.

[0118] Step S27: Task-based training and incremental adaptation. The method described in this application supports independent training of city-level representation learning models and city cluster-level representation learning models. Each level of the model can continuously receive new event batches and achieve incremental memory updates through online GRU iteration. During training, the generalization and adaptability of the representation are improved by fusing long-term historical states and short-term neighbor interactions. Finally, with the support of the city-level and city cluster-level representation learning models, user and region representation vectors with multi-scale spatiotemporal dynamic features are generated respectively. Specifically, both the city-level and city cluster-level representation learning models are based on the user memory vector obtained after memory update in step S25. and Through a lightweight perceptron mapping function Generate the final representation. Specifically, formally, on a city-level graph structure, the output of the city-level representation learning model is: a set of city-level user representations: City-level regional representation set: On the city cluster graph structure, the output of the city cluster representation learning model is: a set of user representations at the city cluster level. Set of regional representations at the city cluster level: .

[0119] The aforementioned multi-scale spatiotemporal representation set will serve as the input basis for subsequent steps, supporting the construction of a unified semantic space and various downstream analysis tasks.

[0120] In one embodiment, step S3 aims to implement a semantic alignment mechanism between multi-granularity user dynamic representations, especially for users with cross-city migration behavior, to construct a consistent representation space between city-level and city cluster-level dynamic representations, so as to improve the model's ability to understand cross-regional behavior and its migration generalization ability.

[0121] Step S31: Multi-granularity dynamic representation extraction. In this step, the set of users with cross-city travel behavior is first identified, and their dynamic representations in the city-level dynamic graph are obtained from the output of step S27. Dynamic representation in city cluster-level dynamic maps .

[0122] Step S32: Representation mapping to a shared semantic space. To achieve scale alignment, the method designs independent projection modules, which act on... and It adopts a multilayer perceptron (MLP) structure.

[0123] .

[0124] in, and These are the projected city-level and city cluster-level representation vectors, respectively.

[0125] Step S33: Constructing positive and negative samples. Using user u as the anchor point, construct positive sample pairs as follows: This represents two representations of the user at different scales. Additionally, for each anchor user, several other users are randomly selected from their active cities. (In one embodiment, the number is set to 10), constructing negative samples This strategy ensures that negative samples are spatially similar, but have no correlation in terms of behavioral paths and identities.

[0126] Step S34: Similarity Calculation and Comparison Loss Optimization. Calculate the cosine similarity for all positive and negative sample pairs. This represents the pre-defined similarity between two vectors, and the projection network parameters are optimized using the InfoNCE contrastive loss function. For each anchor user, the goal is to maximize the similarity between their positive sample pairs and minimize the similarity between them and their negative sample pairs. The loss function is defined as follows: .

[0127] in, This is the temperature coefficient, with a default value of 0.07. This value is consistent with contrastive learning works such as OpenAI CLIP (Contrastive Language–Image Pre-training), which can effectively balance the smoothness of the similarity distribution between positive and negative samples and performs robustly in practice.

[0128] Step S35: Align the target and optimize the output representation. By training the contrastive learning framework described above, the parameters in the projection module are optimized so that the dynamic representation of the same user in the city-level and city cluster-level dynamic maps are highly consistent in the shared semantic space, while irrelevant representations of different users or at different scales remain distinguishable. Finally, the representation in the city-level dynamic map is... After training the MLP projection within the city, that is: This yields the spatially aligned user dynamic representation vector. It possesses both city-level fine features and city cluster-level macro-migration dynamics, serving as one of the input representations for subsequent downstream tasks.

[0129] In one embodiment, step S4 above involves outputting aligned user dynamic representations and region representations, and applying them to multiple downstream smart city analysis tasks. This step aims to structure the user and region dynamic representations learned through dynamic graph modeling and scale alignment mechanisms, and apply them to various analysis and prediction tasks in smart city scenarios.

[0130] Step S41: Representation Output and Organization. In this embodiment, the city-level, city cluster-level, and aligned user dynamic representation vectors obtained in the preceding steps are structurally organized, and the output format is uniformly set as a three-dimensional tensor or dictionary. The output representation set is as follows: City-level user dynamic representation set: A set of dynamic user representations at the city cluster level: Aligned set of user dynamic representations: Set of region representations: including and Each vector encodes the temporal evolution behavior and spatial context features of a user / region.

[0131] Step S42: Downstream task adaptation and model transfer. This application embodiment supports using the various dynamic representation sets output by S41 as input features to flexibly adapt to various smart city analysis tasks, including but not limited to: (1) User commuting type identification: distinguishing user behavior as commuting, staying or high-frequency migration; (2) Intercity mobility detection: identifying travel paths, frequencies and time period characteristics between cities; (3) Regional function classification: inferring functional attributes based on regional representations; (4) Group structure analysis: mining group distribution and behavioral differences based on similarity in the representation space.

[0132] Step S43: Downstream Task Fine-tuning. For various possible downstream tasks, this application uniformly adopts a lightweight multilayer perceptron structure as the task classifier. Based on the multi-granularity dynamic representation vector described in step S41, supervised learning is used to update the parameters of the task classifier to achieve fine-tuning. Fine-tuning can be manifested as updating only the parameters of the task classifier while keeping the parameters of the main dynamic graph encoding model unchanged, or performing small-scale joint updates to adapt to the specific data distribution and target requirements of different tasks. The input is the dynamic representation vector of the user or region, and the output is the classification label of the corresponding task (such as commuting type, region functional category, etc.).

[0133] In summary, the user region dynamic representation method shown in the embodiments of this application has the following technical effects.

[0134] 1. Possesses multi-scale regional modeling capabilities, solving the problem of single granularity in existing technologies: By introducing dynamic graph structures at both the city and city cluster levels, it enables hierarchical modeling of user behavior, making it more suitable for complex migration and cross-regional behavior analysis in the context of city clusters.

[0135] 2. Achieving consistent alignment of user behavior representations and improving the model's cross-city generalization ability: This application's embodiments introduce a representation projection and contrastive learning mechanism to align user behaviors in city-level and city cluster-level dynamic maps within a unified semantic space, solving the representation splitting problem caused by scale inconsistency. This technique significantly improves the model's stability and accuracy in cross-city tasks, while existing methods often focus on a single city or region, making it difficult to generalize the learned representations.

[0136] 3. Enhanced temporal modeling capabilities and improved expression accuracy in dynamically changing scenarios: By extracting contextual semantics through the Transformer encoder and updating memory states in combination with GRU, the embodiments of this application can capture the dynamic changing trends of user behavior; compared with the static graph or single-step modeling methods in existing solutions, the embodiments of this application have stronger temporal response capabilities.

[0137] 4. Possesses self-supervised learning capabilities, reducing reliance on manually labeled data: By predicting the user's next-hop region to construct a self-supervised task, the model can be optimized under conditions of no or weakly labeled data; compared with the training methods in existing technologies that rely on large-scale labeled samples, the embodiments of this application are more easily implemented in real large-scale smart city systems.

[0138] 5. Supports dynamic model updates to adapt to streaming data application scenarios. The embodiments of this application retain node state memory and an event-driven update mechanism in their design, enabling incremental learning of the dynamic graph's evolution over time. This mechanism supports model updates as data arrives in real time, making it suitable for high real-time tasks such as traffic monitoring and crowd flow analysis, thus solving the shortcomings of existing static modeling schemes that cannot adapt to real-time streaming data.

[0139] Figure 3 This application illustrates a user region dynamic characterization device according to an embodiment of the present application, such as... Figure 3 As shown, the device 30 includes: The acquisition module 301 is used to acquire the raw signaling data between the user equipment and the base station; The execution module 302 is used to process the raw signaling data to obtain user location information data; A multi-granularity user region dynamic map is constructed based on user location information data, and representation learning is performed on the multi-granularity user region dynamic map to obtain a multi-scale user dynamic representation, which includes city-level user dynamic representation and city cluster-level user dynamic representation. Cross-scale alignment optimization is performed on the multi-scale user dynamic representation to obtain an aligned user dynamic representation.

[0140] In summary, the method described in this application discloses multi-scale (city and city cluster) dynamic graph construction and cross-scale user behavior alignment, pays more attention to user-level behavior evolution and representation transfer capabilities, and has a wider range of applications; it has stronger advantages in unified modeling of user behavior across regions and downstream multi-task generalization; it can more effectively preserve the dynamic differences of user behavior structure at different scales, has real-time and dynamic update capabilities, and performs better in user behavior modeling and generalization capabilities.

[0141] This application provides a network device 40, such as... Figure 4 As shown, the network device 40 includes a processor 401, a memory 402, and a program stored in the memory 402 and executable on the processor 401. When the program is executed by the processor 401, it implements the steps of a user area dynamic representation method as shown in the above embodiment.

[0142] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of a user area dynamic representation method as shown in the above embodiments, achieving the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0143] This application also provides a computer program product, including computer instructions. When executed by a processor, the computer instructions implement the steps of the user area dynamic representation method shown in the above embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0144] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0145] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, 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 is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0146] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for dynamically representing user regions, characterized in that, The method includes: Acquire raw signaling data between user equipment and base stations; The original signaling data is processed to obtain user location information data; A multi-granularity user region dynamic map is constructed based on the user location information data, and representation learning is performed on the multi-granularity user region dynamic map to obtain a multi-scale user dynamic representation, wherein the multi-scale user dynamic representation includes city-level user dynamic representation and city cluster-level user dynamic representation. Cross-scale alignment optimization is performed on the multi-scale user dynamic representation to obtain the aligned user dynamic representation.

2. The method according to claim 1, characterized in that, After performing cross-scale alignment optimization on the multi-scale user dynamic representation to obtain the aligned user dynamic representation, the method further includes: Obtain the task type of the target downstream task, and process the aligned user dynamic representation according to the task type to obtain task input features; The task input features are input into a task classifier for classification processing to obtain structured prediction results.

3. The method according to claim 1, characterized in that, The original signaling data includes user identifier, base station identifier, and connection time information between user equipment and base station; The process of processing the original signaling data to obtain user location information data includes: Obtain the corresponding latitude and longitude coordinates of the base station based on the base station identifier, determine the grid area to which the base station belongs based on the base station latitude and longitude coordinates, and determine the city-level area identifier and city cluster-level area identifier of the base station based on the grid area; Based on the user identifier, the city-level region identifier, the city cluster-level region identifier, and the connection time information, an initial location information quadruple is constructed; wherein, the initial location information quadruple is used to describe the user's stay record in the region, and the region includes city-level regions and city cluster regions; The initial location information quadruple is cleaned and then classified into an intra-city trajectory subset and an inter-city trajectory subset. The intra-city trajectory subset and the inter-city trajectory subset are then identified as the user location information data.

4. The method according to claim 3, characterized in that, The step of constructing a multi-granularity user region dynamic map based on the user location information data, and performing representation learning on the multi-granularity user region dynamic map to obtain a multi-scale user dynamic representation includes: Construct a city-level dynamic map based on the subset of trajectories within the city; Construct a city cluster-level dynamic map based on the aforementioned intercity trajectory subset; The city-level dynamic map and the city cluster-level dynamic map are used as the multi-granularity user area dynamic map. The city-level dynamic graph is subjected to representation learning processing to obtain the city-level user dynamic representation; the city cluster-level dynamic graph is subjected to representation learning processing to obtain the city cluster-level user dynamic representation. The city-level user dynamic representation and the city cluster-level user dynamic representation are used as the multi-scale user dynamic representation.

5. The method according to claim 4, characterized in that, The process of performing representation learning on the city-level dynamic graph to obtain the city-level user dynamic representation includes: For the user nodes and region nodes in the city-level dynamic graph, initialize the user memory matrix and region memory matrix respectively; For each dwell event in the city-level dynamic map, the dwell duration, interaction interval, and regional attribute information of the dwell event are obtained. The dwell duration, interaction interval, and regional attribute information are encoded and fused to obtain an event feature embedding vector. The event feature embedding vector is subjected to context encoding to obtain a context-enhanced representation; Based on the context-enhanced representation, message aggregation is performed on the adjacency events of the user node and each node in the region node to obtain the local dynamic representation of the node, wherein the adjacency events are a set of dwell events associated with the node; Based on the local dynamic representation, the user memory matrix, and the region memory matrix, the memory vector of the corresponding node is updated to obtain the updated user memory vector and the updated region memory vector. The updated user memory vector is mapped to obtain the city-level user dynamic representation.

6. The method according to claim 5, characterized in that, The process of encoding and fusing the dwell time, the interaction interval, and the regional attribute information to obtain an event feature embedding vector includes: The dwell time is encoded using sine and cosine position coding to obtain the dwell time embedding; The interaction interval is encoded using sine and cosine position encoding to obtain the interaction interval embedding; The dwell time embedding and the interaction interval embedding are concatenated to obtain the time feature embedding; Based on the regional attribute information, feature encoding is performed to obtain the regional static features; The event feature embedding vector is obtained by concatenating the time feature embedding with the regional static feature and then fusing them.

7. The method according to claim 5, characterized in that, Based on the context-enhanced representation, message aggregation is performed on the adjacency events between the user node and each node in the region nodes to obtain a local dynamic representation of the node, including: For the user node, based on the association relationship of the edges in the city-level dynamic graph, the most recent preset number of dwell events associated with the user node are obtained, and the obtained dwell events are used as the adjacency events of the user node to form a user adjacency event set. For the region node, based on the association relationship of the edges in the city-level dynamic graph, the most recent preset number of dwell events associated with the region node are obtained, and the obtained dwell events are used as the adjacency events of the region node to form a region adjacency event set. Obtain the context-enhanced representation corresponding to each adjacency event from the user adjacency event set and the region adjacency event set; The local dynamic representation of the node is obtained by aggregating the context-enhanced representations corresponding to each of the adjacent events.

8. The method according to claim 5, characterized in that, Based on the local dynamic representation, the user memory matrix, and the region memory matrix, the memory vector of the corresponding node is updated to obtain the updated user memory vector and the updated region memory vector, including: For the user node, the user memory vector of the user node at the previous time step is obtained from the user memory matrix, and the local dynamic representation of the user node is updated with the user memory vector at the previous time step to obtain the updated user memory vector. For the region node, the region memory vector of the region node at the previous time step is obtained from the region memory matrix. The local dynamic representation of the region node and the region memory vector at the previous time step are input into the gated loop unit for update processing to obtain the updated region memory vector.

9. The method according to claim 5, characterized in that, The multi-scale user dynamic representation is subjected to cross-scale alignment optimization to obtain an aligned user dynamic representation, including: User identifiers are determined from the intercity trajectory subset, and users corresponding to the user identifiers are identified as target users with cross-city travel behavior, thus forming a target user set. Obtain the city-level user dynamic representation and the city cluster-level user dynamic representation corresponding to each target user in the target user set; Configure a first projection module and a second projection module. Input the city-level user dynamic representation into the first projection module for mapping processing to obtain the projected city-level representation vector; input the city cluster-level user dynamic representation into the second projection module for mapping processing to obtain the projected city cluster-level representation vector. Using each target user in the target user set as an anchor point, the projected city-level representation vector and the projected city cluster-level representation vector of the same target user are constructed as positive sample pairs; based on the user location information data, the city with the highest frequency of occurrence of the target user is determined as the active city of the target user; the projected city cluster-level representation vectors of other users are extracted from the active cities of the target user, and these are constructed as negative sample pairs with the projected city-level representation vector of the target user; Based on the comparison and learning between the positive sample pairs and the negative sample pairs, the parameters of the first projection module and the second projection module are optimized and updated; The first projection module, after optimizing and updating the input parameters of the city-level user dynamic representation, is mapped to obtain an aligned and optimized city-level representation vector. The aligned and optimized city-level representation vector is then used as the aligned user dynamic representation output, wherein the aligned user dynamic representation has both city-level and city cluster-level features.

10. The method according to claim 9, characterized in that, Based on the comparison and learning between the positive sample pairs and the negative sample pairs, the parameters of the first projection module and the second projection module are optimized and updated, including: For the positive sample pair, calculate the cosine similarity between the projected city-level representation vector and the projected city cluster-level representation vector in the positive sample pair to obtain the positive sample similarity. For the negative sample pair, calculate the cosine similarity between the projected city-level representation vector of the target user and the projected city cluster-level representation vector of the other users in the negative sample pair to obtain the negative sample similarity set; Calculate the contrast loss value based on the positive sample similarity and the negative sample similarity set; The parameters of the first projection module and the second projection module are optimized and updated based on the contrast loss value, wherein the goal of the optimization and update is to maximize the similarity of the positive samples and minimize the similarity in the set of negative samples.

11. A user area dynamic representation device, characterized in that, The device includes: The acquisition module is used to acquire raw signaling data between user equipment and base station; The execution module is used to process the original signaling data to obtain user location information data; A multi-granularity user region dynamic map is constructed based on the user location information data, and representation learning is performed on the multi-granularity user region dynamic map to obtain a multi-scale user dynamic representation, wherein the multi-scale user dynamic representation includes city-level user dynamic representation and city cluster-level user dynamic representation. Cross-scale alignment optimization is performed on the multi-scale user dynamic representation to obtain the aligned user dynamic representation.

12. A network device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of a user region dynamic representation method as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of a user region dynamic representation method as described in any one of claims 1 to 10.

14. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of a user region dynamic representation method as described in any one of claims 1 to 10.

Images