Platform user interest recommendation method and system based on artificial intelligence

Abstract

The invention relates to the technical field of information pushing, in particular to a platform user interest recommendation method and system based on artificial intelligence. The method comprises the following steps: acquiring user action data and environment perception data, analyzing the user action data and the environment perception data, and constructing a unified space-time semantic network; according to the unified space-time semantic network, analyzing scene features of the user, and determining a multi-dimensional dynamic scene feature set; and analyzing the multi-dimensional dynamic scene feature set, determining composite scene demand information of the current user, and retrieving and pushing real-time demand information of the user according to the composite scene demand information. According to the method and the device, the perception capability of the pushed information to the user demand scene is improved, the multi-dimensional dynamic adaptation of the pushed information to the platform user is realized, the information pushing accuracy for the platform user is improved, and the composite information demand of the user in a complex scene is met.

Description

Technical Field

[0001] This application relates to the technical field of information push, and in particular to a method and system for recommending platform user interests based on artificial intelligence. Background Art

[0002] As the core carrier of urban digital transformation, the smart city service platform integrates real-time data across fields such as transportation, energy, environmental protection, emergency response, and public services, aiming to improve urban governance efficiency and residents' living experience through intelligent means.

[0003] However, the information push strategy and logic of existing smart city service platforms usually rely on static rules or single-dimensional tags, lacking the ability to accurately perceive user demand scenarios, and not fully considering the information coupling requirements brought by scenario intersections, resulting in a mismatch between the pushed information content and users' real-time needs. Summary of the Invention

[0004] This application provides a method and system for recommending platform user interests based on artificial intelligence to solve the above technical problems.

[0005] In a first aspect, this application provides a method for recommending platform user interests based on artificial intelligence, the method including: Obtaining user action data and environmental perception data, analyzing the user action data and environmental perception data, and constructing a unified spatio-temporal semantic network; According to the unified spatio-temporal semantic network, analyzing the scene characteristics where the user is located to determine a multi-dimensional dynamic scene feature set; Analyzing the multi-dimensional dynamic scene feature set, determining the composite scene demand information of the current user, and retrieving and pushing the real-time demand information of the user according to the composite scene demand information.

[0006] Through this solution, the user action data and environmental perception data are jointly analyzed, the constructed unified spatio-temporal semantic network is used to perform multi-dimensional dynamic scene feature analysis on the scene where the user is located, based on the obtained multi-dimensional dynamic scene feature set, the composite scene demand information of the current user is analyzed, and the real-time demand information of the user is retrieved and pushed according to the composite scene demand information, improving the perception ability of the pushed information for the user demand scene, realizing the multi-dimensional dynamic adaptation of the pushed information to platform users, improving the accuracy of information push for platform users, and meeting the composite information needs of users in complex scenarios.

[0007] Optionally, the user action data includes user stay action information and historical search records; The environmental perception data includes real-time traffic status data, meteorological status data, public facility operation status data, and public event data.

[0008] Through this solution, the specific content of user action data and environmental perception data is clarified and constrained. By integrating user action data (stay actions, search records) and environmental perception data (traffic, meteorology, facilities, events), a "user-environment" two-way interaction model is constructed to align user actions with the real-time environmental state, identify the environmental sensitivity of user needs, and significantly improve the matching degree of the subsequent push information with the user's real-time scenario.

[0009] Optionally, analyzing the user action data and environmental perception data to construct a unified spatio-temporal semantic network includes: Based on the user stay action information, analyze the environmental perception data and screen and determine the local environmental perception data corresponding to the area where the user is located; Perform timestamp alignment processing on the real-time traffic state data, the meteorological state data, the public facility operation state data, and the public event data in the local environmental perception data to determine the local environmental perception synchronization data under the same time axis; Based on the graph neural network algorithm with a fusion attention mechanism, use the user action data as the attention target node and the data items in the corresponding local environmental perception synchronization data as neighbor nodes to construct a basic association graph; According to the basic association graph, use the real-time relative relationship between the entity corresponding to the neighbor node and the entity corresponding to the attention target node as the attention weight between the neighbor node and the corresponding attention target node to determine the attention graph neural network; According to the dynamic attention update strategy, dynamically update the attention graph neural network to construct the unified spatio-temporal semantic network in real time.

[0010] Through this solution, spatio-temporal alignment is performed on the user stay action information and the corresponding environmental perception data to obtain local environmental perception synchronization data, avoiding spatio-temporal misalignment between subsequent push information and user needs. Based on the graph neural network algorithm, use the user action data as the attention target node and the data items in the corresponding local environmental perception synchronization data as neighbor nodes, use the basic association graph with a fusion attention mechanism to represent the association relationship between the user and the environment, and introduce the attention weight representing the association degree between the user and the environment to construct the attention graph neural network. The attention graph neural network is dynamically updated through the dynamic attention update strategy to realize the construction of the unified spatio-temporal semantic network, so that the subsequent analysis and push information analysis process can be quickly and dynamically adapted to the changes in the user's scenario.

[0011] Optionally, dynamically updating the attention graph neural network includes at least the following steps: Extract the influence range of each piece of data in the environmental perception data corresponding to the entity perception device according to the environmental perception data; Draw the user movement trajectory information according to the user stay action information; The dynamic attention update strategy includes a passive update mechanism and an active update mechanism; The passive update mechanism: when the user movement trajectory information changes, according to the relative position relationship between the influence range and the change of the user stay action, dynamically update several neighbor nodes corresponding to the current attention target node in the attention graph neural network; The active update mechanism: based on the long short-term memory collaborative algorithm, predict the subsequent movement trajectory of the user according to the user movement trajectory information and the historical search record, and preload the environmental perception data nodes involved in the subsequent as pre-update nodes corresponding to several neighbor nodes. Through this solution, using the dynamic attention update strategy including the passive update mechanism and the active update mechanism, and using the dual-channel update strategy, the node relationship in the attention graph neural network is automatically and quickly updated with the user's actions. At the same time, by predicting the user's action trajectory, the efficiency of subsequent user information demand analysis is improved, and the data push delay and hardware calculation pressure are reduced.

[0012] Optionally, analyzing the scene characteristics where the user is located according to the unified spatio-temporal semantic network to determine the multi-dimensional dynamic scene feature set includes: Extract the interaction features between the user's actions and the environment where the user is located based on the dynamic attention association relationship between the attention target node and the corresponding neighbor nodes in the unified spatio-temporal semantic network; Extract time-sensitive features according to the time continuity of the interaction features; Extract space-associated features according to the spatial topological relationship of the interaction features; Extract event-response features according to the event trigger conditions of the interaction features; Analyze the spatio-temporal coupling strength between the time-sensitive features, the space-associated features and the event-response features to identify the scene cross-combination patterns; Construct the multi-dimensional dynamic scene feature set according to the scene cross-combination patterns; The cross-combination patterns include time-space superposition scenes, time-event trigger scenes, space-event linkage scenes and multi-dimensional linkage scenes.

[0013] Through this solution, based on the interaction characteristics between the user's actions and the environment they are in, the influence of environmental factors on the user's information needs is divided into three dimensions: time-sensitive, space-related, and event-responsive, which respectively reflect the influence characteristics of the environment on the user's information needs from the three perspectives of time, space, and events. Through the cross-characteristics between different characteristics, the current scenario where the user is located is identified, so as to avoid the situation where the pushed information does not fully match the user's needs caused by single-dimensional feature analysis and preset scenario modes.

[0014] Optionally, extracting the time-sensitive features according to the time continuity of the interaction characteristics includes: Based on the user's stay action information and the historical search records, divide the user's action time segments, and mark the action frequencies and durations in different time periods; According to the real-time traffic state data and the meteorological state data in the same time segment of the environmental perception data, analyze the action frequencies and the durations, identify the influence intensity of the external environment on the user's actions within the current time segment, and generate a time-sensitive feature vector as the time-sensitive feature.

[0015] Through this solution, using machine learning algorithms, identify the characteristics of the user's action time cycle, and on this basis, construct the correlation characteristics between the user's actions and environmental factors in different time cycles, so as to analyze the influence of external conditions on the user's information needs in the time dimension, generate the corresponding time-sensitive feature vector as the time-sensitive feature, and complete the leapfrog upgrade of the time dimension analysis process for the user's information needs from "mechanical time period processing" to "intelligent time series perception" by establishing a refined time-sensitive feature analysis system.

[0016] Optionally, extracting the space-related features according to the spatial topological relationship of the interaction characteristics includes: According to the user's stay action information, construct a dynamic spatial grid centered on the user's stay location, and divide the core area and the peripheral radiation area; Analyze the local environmental perception synchronization data, extract the location distribution and operating status of public facilities, and combine the real-time traffic state data to identify the spatial topological relationship between the public facilities and the dynamic spatial grid, and generate a facility accessibility map; According to the geographical location keywords involved in the historical search records, extract the high-frequency interest space nodes, and mark the high-frequency interest space nodes in the facility accessibility map to generate a point-of-interest heat map as the space-related feature.

[0017] Through this solution, a dynamic spatial grid centered on the user's staying location is constructed to adapt the analysis process in the spatial dimension to the user's action scenario. The core area is used to focus on the user's real-time location, and the peripheral radiation area is used to expand the user's potential needs. Combining real-time traffic status data, a facility accessibility map is generated. According to the geographical location keywords reflecting the user's spatial dimension needs, high-frequency interest space nodes are extracted, and the high-frequency interest space nodes are marked in the facility accessibility map to generate a point-of-interest heat map as a spatial association feature, enabling the spatial dimension analysis process for the user's information needs to achieve a leapfrog upgrade from "mechanical location matching" to "scenario-based spatial services".

[0018] Optionally, the extracting of the event response type feature according to the event trigger condition of the interaction feature includes: Analyze the public event data to determine whether the corresponding public event belongs to a strong push demand event; If the public event belongs to the strong push demand event, analyze the spatial overlap degree between the current user movement trajectory information and the predicted movement trajectory and the influence range corresponding to the public event. When the spatial overlap degree is greater than the preset overlap degree threshold, determine the current public event data as the event response type feature; If the public event does not belong to the strong push demand event, analyze the user's event interest degree in the current public event according to the historical search record. When the event interest degree is greater than the preset interest degree threshold, determine the current public event data as the event response type feature.

[0019] Through this solution, public events are divided into strong push demand events and non-strong push demand events. Taking the influence range of the strong push demand event as the main factor, analyze the spatial overlap degree between the current user movement trajectory information and the predicted movement trajectory and the influence range corresponding to the public event to determine whether to use the corresponding public event data as the event response type feature. Taking the user's event interest degree in the non-strong push demand event as the main factor, determine whether to use the corresponding public event data as the event response type feature, enabling the event dimension analysis process for the user's information needs to achieve a leapfrog upgrade from "mechanical location matching" to "scenario-based spatial services".

[0020] Optionally, the analyzing of the multi-dimensional dynamic scenario feature set to determine the composite scenario demand information of the current user includes: Based on the dynamic priority adjustment strategy, according to the scenario cross-combination mode characterized in the multi-dimensional dynamic scenario feature set, identify the primary and secondary driving dimensions in the scenario where the current user is located, and construct a scenario influence weight set; According to the described scenario impact weight set, in combination with the multi-dimensional dynamic scenario feature set, perform multi-modal feature fusion on the time-sensitive feature vector, the space-associated feature, and the event-responsive feature to generate a composite requirement feature vector as the composite scenario requirement information; The dynamic priority adjustment strategy includes: when the scenario cross-combination pattern includes the time-space superposition scenario, dynamically allocate scenario priorities based on the urgency of the time segment where the user is located and the accessibility of core area facilities; When there is an event-responsive feature corresponding to the strong push requirement event in the scenario cross-combination pattern, using the event-responsive feature as the dominant dimension, and according to the urgency of the corresponding strong push requirement event, increase the impact weight of the corresponding event-responsive feature; When the public event corresponding to the event-responsive feature in the scenario cross-combination pattern does not belong to the strong push requirement event, adjust the impact weight of the corresponding event-responsive feature according to the current event interest degree.

[0021] Through this solution, through dynamic priority adjustment and multi-modal fusion, solve the problem of fuzzy recognition of user information requirements under multi-scenario coupling. At the same time, use the dynamic weight mechanism to make the information requirement analysis process quickly respond to the dynamic changes of the scenario, so that the analysis process of user information requirements completes a leapfrog upgrade from "mechanical feature matching" to "scenario-based spatial service", and improves the matching degree between the pushed information and the actual information requirements of users.

[0022] In a second aspect, the present application provides an artificial intelligence-based platform user interest recommendation system, and the system includes: An action perception module, configured to obtain user action data and environmental perception data, analyze the user action data and environmental perception data, and construct a unified spatio-temporal semantic network; A scenario analysis module, configured to analyze the scenario features where the user is located according to the unified spatio-temporal semantic network, and determine a multi-dimensional dynamic scenario feature set; A push module, configured to analyze the multi-dimensional dynamic scenario feature set, determine the composite scenario requirement information of the current user, and retrieve and push the real-time requirement information of the user according to the composite scenario requirement information. Description of the Drawings

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0024] Figure 1 A schematic diagram of an application scenario provided by an embodiment of the present application; Figure 2 A flowchart of a method for recommending user interests on a platform based on artificial intelligence provided by an embodiment of the present application; Figure 3 A schematic structural diagram of a system for recommending user interests on a platform based on artificial intelligence provided by an embodiment of the present application. Detailed implementation manners

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Apparently, the described embodiments are some but not all of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the scope of protection of the present application.

[0026] In addition, the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the associated objects before and after unless otherwise specified.

[0027] The embodiments of the present application will be further described in detail below with reference to the accompanying drawings of the specification.

[0028] The information push strategy and push logic of existing smart city service platforms usually rely on static rules or single-dimensional tags, lacking the ability to accurately perceive user demand scenarios, and not fully considering the information coupling requirements brought by scenario intersections, resulting in a mismatch between the content of the pushed information and the real-time needs of users.

[0029] Based on this, the present application provides a method and system for recommending user interests on a platform based on artificial intelligence. By jointly analyzing user action data and environmental perception data, using the constructed unified spatio-temporal semantic network, performing multi-dimensional dynamic scenario feature analysis on the user's current scenario, based on the obtained multi-dimensional dynamic scenario feature set, analyzing the composite scenario demand information of the current user, and retrieving and pushing the real-time demand information of the user according to the composite scenario demand information, improving the perception ability of the pushed information to the user demand scenario, realizing the multi-dimensional dynamic adaptation of the pushed information to the platform users, improving the accuracy of information push for platform users, and meeting the composite information needs of users in complex scenarios.

[0030] Figure 1A schematic diagram of an application scenario provided for this application. In the process of pushing information to platform users in a smart city service platform, the method provided by this application is applied to improve the perception ability of the pushed information for the user demand scenario and realize the multi-dimensional dynamic adaptation of the pushed information to the platform users.

[0031] Specifically, the method of this application is applied to any server, which communicates with the platform user terminal device and the Internet of Things environment sensor. The server obtains the user behavior data provided by the platform user terminal device and the environment perception data provided by the Internet of Things environment sensor, jointly analyzes the user action data and the environment perception data, uses the constructed unified spatio-temporal semantic network to perform multi-dimensional dynamic scenario feature analysis on the user's current scenario, and based on the obtained multi-dimensional dynamic scenario feature set, analyzes the composite scenario demand information of the current user, and retrieves and pushes the user's real-time demand information according to the composite scenario demand information, improving the perception ability of the pushed information for the user demand scenario, realizing the multi-dimensional dynamic adaptation of the pushed information to the platform users, improving the accuracy of information push for platform users, and meeting the composite information needs of users in complex scenarios.

[0032] The specific implementation method can refer to the following embodiments.

[0033] Figure 2 A flowchart of a platform user interest recommendation method based on artificial intelligence provided for an embodiment of this application. The method of this embodiment can be applied to the server in the above scenario. As Figure 2 shown, the method includes: S201. Obtain user action data and environment perception data, analyze the user action data and the environment perception data, and construct a unified spatio-temporal semantic network.

[0034] The user action data can be the dynamic interaction information of the user on the smart city service platform, and the user action data comes from the positioning module and the log system of the platform user terminal device.

[0035] The environment perception data can be the real-time state information of the external environment where the user is located, and the environment perception data can be obtained through various Internet of Things environment sensors deployed in the city.

[0036] The unified spatio-temporal semantic network can be a dynamic graph structure based on spatio-temporal correlation, used to represent the dynamic interaction relationship between user actions and the environment.

[0037] Specifically, in the information push technology of existing smart city service platforms, user action data and environmental perception data are processed independently, lacking unified association in the spatio-temporal dimension, resulting in the disconnection between the pushed information and the user's real-time scenario, and it is difficult to accurately reflect the user's composite information needs under the cross-influence of factors such as time, space, and events. By analyzing user action data and environmental perception data, align the user action and environmental perception data within a unified spatio-temporal framework to ensure the spatio-temporal correlation between the user action and the corresponding environmental perception data, and avoid misjudgment of association caused by spatio-temporal misalignment. According to the unified spatio-temporal framework, based on the attention mechanism of the graph neural network, construct a graph structure for characterizing the dynamic interaction relationship between the user action and the surrounding environment, and obtain a unified spatio-temporal semantic network, providing a basic framework for subsequent analysis of the user's composite needs in complex scenarios.

[0038] S202. According to the unified spatio-temporal semantic network, analyze the scene characteristics where the user is located to determine a multi-dimensional dynamic scene feature set.

[0039] The multi-dimensional dynamic scene feature set can be a set of feature information starting from the perspectives of time, space, and events, used to characterize the feature information related to the user's information needs within the scene where the user is located.

[0040] Specifically, user needs are often driven by the intersection of multiple scenarios. For example, when encountering a subway failure (event trigger) during the commute (time-space superposition), it is necessary to push alternative routes and late arrival warnings simultaneously. However, the information push mechanism of existing smart city platforms is difficult to identify such coupled needs. This solution decomposes the complex physical scene where the user is located into basic dimension features such as time, space, and events, thereby characterizing the impact of the complex environment where the user is located on the user's composite information needs at the data level, and providing an information basis for the subsequent user need analysis process through the integrated multi-dimensional dynamic scene feature set.

[0041] S203. Analyze the multi-dimensional dynamic scene feature set, determine the composite scene demand information of the current user, and retrieve and push the user's real-time demand information according to the composite scene demand information.

[0042] The composite scene demand information can be user demand vector feature information generated by weighted fusion of multi-dimensional dynamic scene features.

[0043] The user's real-time demand information can be the pushed content retrieved and matched from the platform database according to the composite scene demand information.

[0044] Specifically, according to the multi-dimensional dynamic scene feature set, the scene features in different dimensions are subjected to feature vector conversion and fusion splicing, and the feature vector information obtained by splicing is used as the composite scene demand information. According to the corresponding feature vectors, a targeted search is performed on the push data graph in the smart city platform to obtain the user's real-time demand information, and the user's real-time demand information is pushed to the corresponding user through the platform user device terminal, such as a smart phone.

[0045] Through this solution, the user's action data and environmental perception data are jointly analyzed, and the constructed unified spatio-temporal semantic network is used to perform multi-dimensional dynamic scene feature analysis on the user's location scene. Based on the obtained multi-dimensional dynamic scene feature set, the composite scene demand information of the current user is analyzed, and the user's real-time demand information is retrieved and pushed according to the composite scene demand information, improving the perception ability of the pushed information to the user's demand scene, realizing the multi-dimensional dynamic adaptation of the pushed information to the platform users, improving the accuracy of information push for platform users, and meeting the composite information needs of users in complex scenarios.

[0046] In some embodiments, the user's action data includes the user's stay action information and historical search records; the environmental perception data includes real-time traffic status data, meteorological status data, public facility operation status data, and public event data.

[0047] The user's stay action information can be the spatio-temporal record of the user staying at a specific location in the urban space, including the stay location coordinates, the start time of the stay, the end time of the stay, and the duration of the stay. This data is collected through the built-in GPS module or base station positioning service of the platform user mobile terminal (such as a smart phone, in-vehicle navigation device), and is obtained in combination with the background positioning permission of the platform application.

[0048] The historical search records can refer to the keywords actively input by the user in the platform through the search box, voice assistant, or quick label, as well as the associated search time, frequency, click results, and other interaction logs. This data is sourced from the persistent storage of the user session records by the platform server and is extracted from the user action log database through a standardized API interface.

[0049] The real-time traffic status data can be a quantitative indicator representing the traffic efficiency of the road network, including the average vehicle speed of the road section, the congestion index, the accident location, etc. The data is sourced from the real-time traffic flow monitoring system of the urban traffic management center and is synchronized to the smart city service platform through the data open platform interface.

[0050] The meteorological status data can be a parameter set describing the current and predicted weather conditions in the corresponding area. This data is sourced from the meteorological monitoring stations and satellite remote sensing systems of the meteorological bureau and is synchronized to the smart city service platform through the data open platform interface.

[0051] The operating status data of public facilities can be dynamic data reflecting the service status of urban infrastructure. This data is collected by Internet of Things sensors, transmitted through Internet of Things communication protocols to the urban Internet of Things platform, and then accessed to the smart city service platform through a data bus.

[0052] Public event data can refer to information on sudden or planned events affecting urban operations, including traffic control announcements, large event arrangements, emergency warning notifications, etc. This data is sourced from information disclosure platforms, social media monitoring systems, and third-party event aggregation services, and is stored in the platform event knowledge base after being structured through natural language processing technology.

[0053] Specifically, when analyzing the information needs of users in the smart city platform, if only relying on users' explicit actions for recommendations without analyzing users' actions in a dynamic environmental context (such as real-time traffic, facility status), it is very likely to lead to a disconnection between the recommended results and the actual scenario needs of users. For example, when a user searches for "indoor parking lot" during a rainstorm, the traditional system only returns a static list of parking lots, ignoring the impact of real-time waterlogged roads on the accessibility of parking lots; by integrating users' action data (staying actions, search records) with environmental perception data (traffic, meteorology, facilities, events), constructing a "user-environment" two-way interaction model, correlating and aligning users' actions with the real-time environmental status, and identifying the environmental sensitivity of users' needs, the matching degree of the subsequent push information obtained from the analysis with the users' real-time scenarios will increase significantly.

[0054] Through this solution, the specific contents of users' action data and environmental perception data are clarified and restricted. By integrating users' action data (staying actions, search records) with environmental perception data (traffic, meteorology, facilities, events), constructing a "user-environment" two-way interaction model, correlating and aligning users' actions with the real-time environmental status, and identifying the environmental sensitivity of users' needs, the matching degree of the subsequent push information obtained from the analysis with the users' real-time scenarios will increase significantly.

[0055] In some embodiments, based on the user's stay action information, the environmental perception data is analyzed to screen and determine the local environmental perception data corresponding to the area where the user is located; the real-time traffic state data, meteorological state data, public facility operation state data, and public event data in the local environmental perception data are subjected to timestamp alignment processing to determine the local environmental perception synchronization data under the same time axis; based on the graph neural network algorithm with a fusion attention mechanism, with the user action data as the attention target node and the various data in the corresponding local environmental perception synchronization data as neighbor nodes, a basic association graph is constructed; according to the basic association graph, taking the real-time relative relationship between the entity corresponding to the neighbor node and the entity corresponding to the attention target node as the attention weight between the neighbor node and the corresponding attention target node, an attention graph neural network is determined; according to the dynamic attention update strategy, the attention graph neural network is dynamically updated to construct a unified spatio-temporal semantic network in real time.

[0056] The local environmental perception data can be a subset of the environmental perception data within a preset geographical range (such as within a radius of 200 meters) around the user's current location.

[0057] The timestamp alignment processing can be a standardization process that maps the time series information of heterogeneous data sources to a reference time axis.

[0058] The local environmental perception synchronization data can be the local environmental perception data after completing the timestamp alignment processing.

[0059] The graph neural network algorithm with a fusion attention mechanism can be a mathematical algorithm that combines graph structure modeling and attention weight calculation, and is used to establish a dynamic association between user actions and environmental perception data.

[0060] The attention target node can be a graph node representing the user action entity.

[0061] The neighbor node can be a graph node representing the environmental perception entity.

[0062] The basic association graph can be a graph structure used to represent the association between the attention target node and the corresponding neighbor nodes.

[0063] The real-time relative relationship can be an index used to represent the degree of correlation between the user and the corresponding environmental perception features, including the real-time spatial distance (Euclidean distance) and semantic correlation (cosine similarity).

[0064] The attention weight can be an index quantitatively obtained according to the real-time relative relationship and used to represent the influence degree of different neighbor nodes on their corresponding attention target nodes.

[0065] The attention graph neural network can be a structured data network used to represent the dynamic association relationship between the attention target node and the corresponding several neighbor nodes.

[0066] The dynamic attention update strategy can be a mechanism that adjusts the corresponding graph neural network structure in real time according to the user's movement trajectory and environmental changes.

[0067] Specifically, user action data and environmental perception data are stored in independent systems (such as traffic data platforms and meteorological data platforms), lacking a unified spatio-temporal alignment mechanism. If the user action data and environmental perception data are processed separately, it is easy to cause spatio-temporal misalignment between the subsequent push information and the user's needs. For example, when recommending a "rain shelter" for the user, due to the lack of spatio-temporal alignment between the associated data, it is easy to fail to synchronously associate the real-time rainfall data with the opening status of surrounding facilities, resulting in the recommended results including closed facilities. Moreover, in the absence of a spatio-temporal alignment mechanism, the push information cannot be quickly and dynamically adapted to the changes in the user's scenario, resulting in the recommended information lagging behind the user's needs and causing information push misalignment. In this solution, a buffer area with a preset radius, such as a 200-meter radius, is generated centered on the user's stay position. Range queries are executed through a spatial database to extract the status data of all traffic, meteorology, public facilities, and public events within this area. Align the user's stay time period (such as 14:00 - 14:30), filter out the environmental perception data that is not within this time window (such as a traffic snapshot at 13:50), and use NTP (Network Time Protocol) to calibrate the clocks of each data source to obtain locally synchronized environmental perception data, achieving spatio-temporal alignment between user action data and environmental perception data. Based on the graph neural network algorithm with a fusion attention mechanism, feature vector encoding is performed on the user action data and the corresponding locally synchronized environmental perception data respectively. Then, taking the user action data as the attention target node and each data within the corresponding locally synchronized environmental perception data as neighbor nodes, a basic association graph is constructed. According to the weighted sum of the real-time spatial distance (Euclidean distance) and semantic relevance (cosine similarity) between the attention target node and the corresponding neighboring nodes, the attention weight between the attention target node and the neighbor nodes is quantified. The attention weights between different nodes are introduced into the basic association graph to construct an attention graph neural network for representing the overall association relationship between the user and the environment. Based on the user's subsequent actions, through the dynamic attention update strategy, the neighbor nodes and the corresponding attention weights in the attention graph neural network are dynamically updated to construct a unified spatio-temporal semantic network, enabling the subsequent analysis and push information analysis process to be quickly and dynamically adapted to the changes in the user's scenario.

[0068] Through this solution, the user's stay action information and the corresponding environmental perception data are aligned in space and time to obtain local environmental perception synchronization data, avoiding spatio-temporal misalignment between subsequent push information and user needs. Based on the graph neural network algorithm, with the user action data as the attention target node and each item of data in the corresponding local environmental perception synchronization data as the neighbor nodes, a basic association graph representing the association relationship between the user and the environment is constructed using a graph structure with a fusion attention mechanism, and an attention weight representing the degree of association between the user and the environment is introduced to construct an attention graph neural network. The attention graph neural network is dynamically updated through a dynamic attention update strategy to realize the construction of a unified spatio-temporal semantic network, so that the subsequent analysis and push information analysis process can be quickly and dynamically adapted to the changes in the user's scene.

[0069] In some embodiments, according to the environmental perception data, the influence range of each entity perception device corresponding to the data in the environmental perception data is extracted; according to the user stay action information, the user movement trajectory information is drawn; the dynamic attention update strategy includes a passive update mechanism and an active update mechanism; the passive update mechanism: when the user movement trajectory information changes, according to the relative position relationship between the influence range and the change of the user stay action, several neighbor nodes corresponding to the current attention target node in the attention graph neural network are dynamically updated; the active update mechanism: based on the long short-term memory collaborative algorithm, according to the user movement trajectory information and the historical search record, the subsequent movement trajectory of the user is predicted, and according to the predicted movement trajectory, the environmental perception data nodes involved in the subsequent process are pre-loaded as the pre-update nodes of the corresponding several neighbor nodes.

[0070] The entity perception device can be an entity Internet of Things device for collecting corresponding environmental perception data.

[0071] The user movement trajectory information can be path data recording the spatial position changes of the user in a continuous time series.

[0072] The influence range can be the effective action range of the environmental perception data collected by the entity perception device.

[0073] The relative position relationship can be the relative spatial position relationship between the user's stay position and the influence range of the corresponding entity perception device.

[0074] The long short-term memory collaborative algorithm can be a mathematical algorithm based on the LSTM (Long Short-Term Memory) model to predict the subsequent movement trajectory of the user according to the user's historical trajectory.

[0075] The predicted movement trajectory can be the subsequent movement trajectory of the user predicted by the long short-term memory collaborative algorithm.

[0076] The pre-updated node can be a data node that pre-loads data of subsequent neighbor nodes.

[0077] Specifically, as the user moves, the environment they are in is in a dynamic change process, which in turn brings about dynamic changes in the user's demand information. This dynamic change is mapped by adjusting the relationships between the nodes in the attention graph neural network. By establishing a spatial influence model for each environmental perception device (such as a traffic camera), converting its coverage area into GeoJSON polygon data, and storing it in a spatial database, the data quantization processing of the influence range of each entity perception device is realized. When the user's movement trajectory information changes, the relative spatial position between the user and the corresponding environmental influence factors changes. Since the degree of association between the attention target node representing the user and the neighbor node representing the environmental influence factors is determined by the real-time spatial distance (Euclidean distance) and semantic relevance (cosine similarity) between the nodes, when the user's movement trajectory information changes, it is necessary to adjust the attention weight according to the relative spatial position between the user and the corresponding environmental influence factors. When the relative distance reflected by the relative spatial position exceeds the preset geographical range centered on the user, the influence of the corresponding neighbor node on the user's information needs is reduced to the lowest level. At this time, the corresponding neighbor node is removed from the corresponding graph structure, the new neighbor node currently within the preset geographical range is added to the corresponding graph structure, and the attention weight between the new neighbor node and the target attention node is calculated to achieve the dynamic local update of the attention graph neural network; the user's historical action trajectory is input into the LSTM model, and the output is the predicted path probability distribution map within a preset update period (such as 10 minutes). Several high-probability waypoints are extracted from the predicted path probability distribution map to construct a predicted movement trajectory. According to this predicted movement trajectory, the environmental perception data located around the predicted movement trajectory and within the preset geographical range centered on the user's position is extracted as the pre-updated node of the subsequent neighbor node, improving the efficiency of subsequent user information needs analysis and reducing data push latency and hardware computing pressure.

[0078] Through this solution, using a dynamic attention update strategy that includes a passive update mechanism and an active update mechanism, and using a dual-channel update strategy, the node relationships in the attention graph neural network are automatically and quickly updated as the user acts. At the same time, by predicting the user's action trajectory, the efficiency of subsequent user information needs analysis is improved, and data push latency and hardware computing pressure are reduced.

[0079] In some embodiments, based on the dynamic attention association relationship between the attention target node and the corresponding neighbor nodes in the unified spatio-temporal semantic network, the interaction features between the user's actions and the environment where the user is located are extracted; according to the time continuity of the interaction features, time-sensitive features are extracted; according to the spatial topological relationship of the interaction features, space-associated features are extracted; according to the event trigger conditions of the interaction features, event response features are extracted; the spatio-temporal coupling strength between the time-sensitive features, space-associated features, and event response features is analyzed to identify the scene cross-combination patterns; according to the scene cross-combination patterns, a multi-dimensional dynamic scene feature set is constructed; the cross-combination patterns include time-space superposition scenes, time-event trigger scenes, space-event linkage scenes, and multi-dimensional linkage scenes.

[0080] The dynamic attention association relationship can be the attention weight that changes with the user's actions.

[0081] The interaction features can be feature data used to characterize the dynamic association relationship between the user's actions and environmental perception data, that is, the attention weight during the change process.

[0082] The time continuity can be the change timestamp corresponding to the attention weight.

[0083] The time-sensitive features can be feature information that affects the user's information needs in the time dimension.

[0084] The spatial topological relationship can be a topological relationship map describing the spatial position of the user and the positions of surrounding facilities.

[0085] The space-associated features can be feature information that affects the user's information needs in the space dimension.

[0086] The event trigger condition can be the trigger condition for the impact of public events on the user's information needs.

[0087] The event response features can be feature information that affects the user's information needs in the event dimension.

[0088] The spatio-temporal coupling strength can be an index used to characterize the degree of interaction between time, space, and event features.

[0089] The scene cross-combination pattern can be a classification of typical user demand scenarios formed by the multi-dimensional interweaving of time, space, and event features.

[0090] The time-space superposition scene can be a scene where time-sensitive features and space-associated features intersect, such as the route information recommendation triggered by the superposition of the morning rush hour on weekdays (time) and the traffic congestion around the company (space).

[0091] The time-event trigger scenario can be a scenario where time-sensitive features and event-responsive features intersect, such as route information recommendation triggered by the morning rush hour on weekdays (time) combined with a sudden traffic accident (event).

[0092] The space-event linkage scenario can be a scenario where space-related features and event-responsive features intersect, such as the emergency channel navigation push triggered by a sudden traffic event (event) when the user approaches a hospital (space).

[0093] The multi-dimensional linkage scenario can be a scenario where time-sensitive features, space-related features, and event-responsive features intersect, such as the recommendation of multi-modal transportation transfer information triggered by the suspension of subway operations (space) during the evening rush hour (time) due to heavy rain (event).

[0094] Specifically, the impact of environmental factors on users' information needs is divided into three dimensions: time-sensitive, space-related, and event-responsive, which respectively reflect the impact characteristics of the environment on users' information needs from the three perspectives of time, space, and event. By the cross-features between different features, the scenario where the user is currently located is identified, so as to avoid the situation that the pushed information does not fully match the user's needs caused by single-dimensional feature analysis and preset scenario modes. Based on the time continuity of interaction features, time segment clustering analysis is performed on the user's actions. According to the user's action characteristics under the environmental perception data corresponding to different time segments, the impact of environmental perception data on users' information needs in different time segments is identified, and time-sensitive features are obtained; centered on the user's real-time location, a spatial grid with a preset geographical radius is generated, and the accessibility of each facility in the spatial grid by the user is analyzed to obtain space-related features; through the event knowledge graph matching mechanism, according to the specific type of public events and the potential association between events and users' actions, event-responsive features are generated; furthermore, a three-dimensional feature tensor of time-space-event is constructed, and the potential association strength between each dimension is extracted through tensor decomposition to obtain the spatio-temporal coupling strength between different features. The time-sensitive features, space-related features, and event-responsive features with the current spatio-temporal coupling strength are matched with historical scenario data in terms of similarity to identify the corresponding scenario cross-combination mode.

[0095] Through this solution, based on the interaction features between the user's actions and the environment where the user is located, the impact of environmental factors on users' information needs is divided into three dimensions: time-sensitive, space-related, and event-responsive, which respectively reflect the impact characteristics of the environment on users' information needs from the three perspectives of time, space, and event. By the cross-features between different features, the scenario where the user is currently located is identified, so as to avoid the situation that the pushed information does not fully match the user's needs caused by single-dimensional feature analysis and preset scenario modes.

[0096] In some embodiments, based on the user's stay action information and historical search records, the user's action time segments are divided, and the action frequency and duration in different time periods are marked; according to the real-time traffic status data and meteorological status data in the same time segment in the environmental perception data, the action frequency and duration are analyzed, the influence of the internal and external environment on the user's action in the current time segment is identified, and a time-sensitive feature vector is generated as a time-sensitive feature.

[0097] The user action time segment may be a time interval into which the user activity cycle is divided.

[0098] Action frequency and duration can be two-dimensional indicators used to quantify the intensity of user activities during a period of time.

[0099] The time-sensitive feature vector may be vector information that represents the characteristics of user information needs changing over time.

[0100] Specifically, the user stay data is cleaned, and the positioning drift points (such as stays with a duration of <30 seconds) are removed. A time series clustering algorithm (such as K-Shape) is used to aggregate behaviors with similar time distributions into typical time periods to complete the division of user action time segments. A search analysis engine, such as Elasticsearch, is used to establish a time period-behavior type aggregation index, and the frequency distribution of each new time period is calculated in real time to obtain the action frequency in different time segments. At the same time, a streaming computing framework (such as Flink) is used to perform sliding window statistics on the duration to obtain the duration of actions in different time segments. The gray correlation analysis method is used to quantify the correlation between real-time traffic status data and meteorological status data and the corresponding action frequency and duration, so as to map the impact of the internal and external environment on user actions in the current time segment, and vector splicing is performed on the action characteristics, environmental characteristics and the impact intensity of the corresponding time segment to generate a time-sensitive feature vector as a time-sensitive feature.

[0101] Through this solution, machine learning algorithms are used to identify the time period characteristics of user actions, and on this basis, the correlation characteristics between user actions and environmental factors in different time periods are constructed, so as to analyze the impact of external conditions on user information needs in the time dimension, generate corresponding time-sensitive feature vectors as time-sensitive features, and establish a refined time-sensitive feature analysis system to enable the time dimension analysis process for user information needs to complete a leapfrog upgrade from "mechanical time period processing" to "intelligent time series perception".

[0102] In some embodiments, according to the user's stay action information, a dynamic spatial grid centered on the user's stay location is constructed, dividing the core area and the peripheral radiation area; analyzing the local environment perception synchronization data, extracting the location distribution and operating status of public facilities, and combining with the real-time traffic status data, identifying the spatial topological relationship between the public facilities and the dynamic spatial grid, generating a facility accessibility map; according to the geographical location keywords involved in the historical search records, extracting the high-frequency interest spatial nodes, and labeling the high-frequency interest spatial nodes into the facility accessibility map, generating a point of interest heat map as a spatial association type feature.

[0103] The dynamic spatial grid can be an adaptive geographical partition grid centered on the user's real-time location.

[0104] The core area can be an area covering the user's immediate activity range.

[0105] The peripheral radiation area can be an area covering the user's potential movement range.

[0106] The location distribution and operating status can be the distribution location and operating status of public facilities.

[0107] The facility accessibility map can be a spatial topological relationship network quantifying the service capacity of public facilities.

[0108] The geographical location keywords can be keywords highly relevant to geographical locations in the historical search records.

[0109] The high-frequency interest spatial nodes can be geographical location tags that repeatedly appear in the historical search records.

[0110] The point of interest heat map can be a graph structure data reflecting the distribution of user spatial demand information.

[0111] Specifically, the Kalman filtering algorithm is used to fuse GPS, WiFi fingerprint, and base station positioning data to determine the user's precise coordinates. Taking the user's location as the origin, a quadtree structure is adopted to divide the space grid, and the grid accuracy of the core area (such as 10 meters × 10 meters) and the radiation area (such as 20 meters × 20 meters) is set. Semantic analysis is performed on the synchronized data of the local environment perception to extract the location distribution and operating status of public facilities (such as the occupancy rate of charging piles, the congestion degree of subways, etc.). Combining with the navigation data mapped by the real-time traffic status data, a multi-dimensional scoring model is constructed (such as traffic weight 40% + real-time facility status weight 30% + user distance weight 30%) to generate the comprehensive accessibility index of each facility. Combining with the facility location, a facility accessibility map is generated. Through a pre-trained language representation model, such as the BERT model, implicit geographical location feature analysis is performed on the historical search records to determine geographical location keywords, and according to the occurrence frequency and timeliness of different keywords, high-frequency interest space nodes are extracted. By annotating the high-frequency interest space nodes into the facility accessibility map, a point-of-interest heat map is generated as a spatial association type feature.

[0112] Through this solution, a dynamic space grid centered on the user's stay location is constructed, enabling the analysis process in the space dimension to adapt to the user's action scenario. The core area is used to focus on the user's real-time location, and the peripheral radiation area is used to expand the user's potential needs. Combining with the real-time traffic status data, a facility accessibility map is generated. According to the geographical location keywords reflecting the user's space dimension needs, high-frequency interest space nodes are extracted, and the high-frequency interest space nodes are annotated into the facility accessibility map to generate a point-of-interest heat map as a spatial association type feature, enabling the space dimension analysis process for the user's information needs to achieve a leapfrog upgrade from "mechanical location matching" to "scenario-based space service".

[0113] In some embodiments, public event data is analyzed to determine whether the corresponding public event belongs to a strong push demand event; if the public event belongs to a strong push demand event, the spatial overlap degree between the current user's movement trajectory information and the predicted movement trajectory and the influence range corresponding to the public event is analyzed. When the spatial overlap degree is greater than the preset overlap degree threshold, the current public event data is determined as an event response type feature; if the public event does not belong to a strong push demand event, according to the historical search records, the user's event interest degree in the current public event is analyzed. When the event interest degree is greater than the preset interest degree threshold, the current public event data is determined as an event response type feature.

[0114] A strong push demand event can be a high-priority public event that needs to be forced to reach the user.

[0115] The spatial overlap degree can be a quantitative index of the geographical intersection of the user's activity trajectory and the event impact area.

[0116] The preset overlap threshold can be the spatial overlap critical value for determining whether event information needs to be pushed.

[0117] The event interest degree can be the quantification value of the user's attention degree to the event types with non-strong push requirements.

[0118] The preset interest degree threshold can be the lowest interest benchmark value for triggering the push of events with non-strong push requirements.

[0119] Specifically, whether to push the information related to public events to users mainly depends on the event type and the user's interest in the event. When the event type is an event that will directly affect users within the influence range, such as sudden disasters and traffic control, it is necessary to actively push the relevant event information to the platform users within its influence range. This type of event is classified as an event with strong push requirements, without considering the user's interest, but focusing on the event influence range. Through a range overlap quantification algorithm, such as the Jaccard coefficient formula, analyze the spatial overlap degree between the current user's movement trajectory information and the predicted movement trajectory and the influence range corresponding to the public event. When the spatial overlap degree is greater than the preset overlap threshold, it indicates that the user is within the influence range corresponding to the current event with strong push requirements, and determine the public event data corresponding to the current event with strong push requirements as the event response type feature; if the relevant public event is not an event with strong push requirements, then dominated by the user's interest (for example, if the user has a high interest in public welfare events, then push the public welfare event information within the push range), construct a TF-IDF weight matrix based on the historical search records, match the event keyword similarity, and quantify the user's event interest degree in the public events within the current range. When the event interest degree is greater than the preset interest degree threshold, it indicates that the user is interested in the current event with non-strong push requirements, and then determine the current public event data as the event response type feature.

[0120] Through this solution, public events are divided into events with strong push requirements and events with non-strong push requirements. Taking the influence range of events with strong push requirements as the main factor, analyze the spatial overlap degree between the current user's movement trajectory information and the predicted movement trajectory and the influence range corresponding to the public event, and judge whether to use the corresponding public event data as the event response type feature. Taking the user's event interest degree in events with non-strong push requirements as the main factor, judge whether to use the corresponding public event data as the event response type feature, so that the event dimension analysis process for the user's information needs completes a leapfrog upgrade from "mechanical position matching" to "scenario-based spatial service".

[0121] In some embodiments, based on a dynamic priority adjustment strategy, according to the scene cross-combination patterns represented in the multi-dimensional dynamic scene feature set, the primary and secondary driving dimensions in the current user's scene are identified, and a scene influence weight set is constructed; according to the scene influence weight set, combined with the multi-dimensional dynamic scene feature set, multi-modal feature fusion is performed on the time-sensitive feature vector, the space-correlation type feature, and the event-response type feature to generate a composite demand feature vector as composite scene demand information; the dynamic priority adjustment strategy includes: when the scene cross-combination pattern includes a time-space superposition scene, the scene priority is dynamically allocated based on the urgency of the time segment where the user is located and the accessibility of the core area facilities; when there is an event-response type feature corresponding to a strong push demand event in the scene cross-combination pattern, with the event-response type feature as the leading dimension, according to the urgency of the corresponding strong push demand event, the influence weight of the corresponding event-response type feature is increased; when the public event corresponding to the event-response type feature in the scene cross-combination pattern does not belong to a strong push demand event, according to the current event interest, the influence weight of the corresponding event-response type feature is adjusted.

[0122] The dynamic priority adjustment strategy can be a decision-making mechanism for dynamically allocating the feature weights of each dimension according to the real-time scene combination features.

[0123] The primary and secondary driving dimensions can be the main dimension features and secondary dimension features that affect the user's information needs in the current scene.

[0124] The scene influence weight set can be a parameter set that quantifies the influence degree of each scene dimension on the user's needs.

[0125] Multi-modal feature fusion can be a deep learning process for integrating heterogeneous scene features.

[0126] The composite demand feature vector can be a multi-dimensional vector feature information that represents the user's comprehensive information needs.

[0127] Specifically, initialize the weights of each dimension according to the scene type (for example, the initial weights of the time-space superposition scene: 45% for time, 40% for space, and 15% for events). When multiple scene modes are in effect simultaneously, adopt a weight superposition mechanism, but set a weight upper limit (such as 75%) to prevent a single dimension from monopolizing. When the scene cross-combination mode includes a time-space superposition scene, adjust the benchmark in units of 1% respectively based on the urgency of the time segment where the user is located and the accessibility of the core area facilities, and adjust the weights of the time and space dimensions proportionally and specifically. For the scene containing events with strong push requirements, exponentially increase the weights of event response features based on the event urgency (such as the remaining effective time of traffic control), and at the same time compress the weights of other features. For the scene corresponding to events with strong push requirements, increase the weights of event response features proportionally based on the user's event interest. Concatenate and fuse the time-sensitive feature vectors, space-correlated features, and event response features with dynamic impact weights to generate a composite demand feature vector as the composite scene demand information.

[0128] Through this solution, by means of dynamic priority adjustment and multi-modal fusion, the problem of fuzzy recognition of user information needs under multi-scene coupling is solved. At the same time, the dynamic weight mechanism is used to make the information needs analysis process quickly respond to the dynamic changes of the scene, enabling the analysis process of the user's information needs to complete a leapfrog upgrade from "mechanical feature matching" to "scenario-based spatial service", and improving the matching degree between the pushed information and the user's actual information needs.

[0129] Figure 3 The following is a schematic structural diagram of a platform user interest recommendation system based on artificial intelligence provided by an embodiment of the present application, as Figure 3 shown. A platform user interest recommendation system 300 based on artificial intelligence in this embodiment includes: an action perception module 301, a scene analysis module 302, and a push module 303.

[0130] The action perception module 301 is used to obtain user action data and environmental perception data, analyze the user action data and environmental perception data, and construct a unified spatio-temporal semantic network; The scene analysis module 302 is used to analyze the scene characteristics where the user is located according to the unified spatio-temporal semantic network, and determine a multi-dimensional dynamic scene feature set; The push module 303 is used to analyze the multi-dimensional dynamic scene feature set, determine the composite scene demand information of the current user, and retrieve and push the user's real-time demand information according to the composite scene demand information.

[0131] Optionally, in the action perception module 301, the user action data includes user stay action information and historical search records; The environmental perception data includes real-time traffic status data, meteorological status data, public facility operation status data, and public event data.

[0132] Optionally, the action perception module 301 is specifically configured to: Based on the user's stay action information, analyze the environmental perception data, and screen and determine the local environmental perception data corresponding to the area where the user is located; Perform timestamp alignment processing on the real-time traffic status data, the meteorological status data, the public facility operation status data, and the public event data in the local environmental perception data to determine the local environmental perception synchronization data under the same time axis; Based on the graph neural network algorithm with a fusion attention mechanism, use the user action data as the attention target node, and use the various data in the corresponding local environmental perception synchronization data as neighbor nodes to construct a basic association graph; According to the basic association graph, use the real-time relative relationship between the entity corresponding to the neighbor node and the entity corresponding to the attention target node as the attention weight between the neighbor node and the corresponding attention target node to determine the attention graph neural network; According to the dynamic attention update strategy, dynamically update the attention graph neural network, and construct the unified spatio-temporal semantic network in real time.

[0133] Optionally, when the action perception module 301 dynamically updates the attention graph neural network, it is specifically configured to: According to the environmental perception data, extract the influence range of the entity perception device corresponding to each data item in the environmental perception data; According to the user's stay action information, draw the user's movement trajectory information; The dynamic attention update strategy includes a passive update mechanism and an active update mechanism; The passive update mechanism: When the user's movement trajectory information changes, according to the relative position relationship between the influence range and the change in the user's stay action, dynamically update several neighbor nodes corresponding to the current attention target node in the attention graph neural network; The active update mechanism: Based on the long short-term memory collaboration algorithm, according to the user's movement trajectory information and the historical search record, predict the user's subsequent movement trajectory, and preload the environmental perception data nodes involved in the future as the pre-update nodes corresponding to several neighbor nodes.

[0134] Optionally, the scenario analysis module 302 is specifically configured to: Extract the interaction features between the user's actions and the environment based on the dynamic attention association relationship between the attention target nodes and the corresponding neighbor nodes in the unified spatio-temporal semantic network; Extract time-sensitive features according to the time continuity of the interaction features; Extract space-associated features according to the spatial topological relationship of the interaction features; Extract event-response features according to the event trigger conditions of the interaction features; Analyze the spatio-temporal coupling strength among the time-sensitive features, the space-associated features, and the event-response features to identify the scene cross-combination patterns; Construct the multi-dimensional dynamic scene feature set according to the scene cross-combination patterns; The cross-combination patterns include time-space superposition scenes, time-event trigger scenes, space-event linkage scenes, and multi-dimensional linkage scenes.

[0135] Optionally, when extracting time-sensitive features according to the time continuity of the interaction features, the scene analysis module 302 is specifically configured to: Based on the user stay action information and the historical search records, divide the user action time segments, and mark the action frequencies and durations in different time periods; According to the real-time traffic state data and the meteorological state data in the same time segment of the environment perception data, analyze the action frequencies and the durations, identify the influence intensity of the external environment on the user actions in the current time segment, and generate a time-sensitive feature vector as the time-sensitive feature.

[0136] Optionally, when extracting space-associated features according to the spatial topological relationship of the interaction features, the scene analysis module 302 is specifically configured to: According to the user stay action information, construct a dynamic space grid centered on the user stay location, and divide the core area and the peripheral radiation area; Analyze the local environment perception synchronization data, extract the location distribution and operation status of public facilities, and combine the real-time traffic state data to identify the spatial topological relationship between the public facilities and the dynamic space grid, and generate a facility accessibility map; According to the geographical location keywords involved in the historical search records, extract the high-frequency interest space nodes, and label the high-frequency interest space nodes in the facility accessibility map to generate an interest point heat distribution map as the space-associated feature.

[0137] Optionally, when extracting event-response features according to the event trigger conditions of the interaction features, the scene analysis module 302 is specifically configured to: Analyze the public event data to determine whether the corresponding public event belongs to a strong push demand event; If the public event belongs to the strong push demand event, analyze the spatial overlap degree between the current user movement trajectory information and the predicted movement trajectory and the influence range corresponding to the public event. When the spatial overlap degree is greater than the preset overlap degree threshold, determine the current public event data as the event response type feature; If the public event does not belong to the strong push demand event, analyze the user's event interest degree in the current public event according to the historical search record. When the event interest degree is greater than the preset interest degree threshold, determine the current public event data as the event response type feature.

[0138] Optionally, the push module 303 is specifically configured to: Based on the dynamic priority adjustment strategy, identify the primary and secondary driving dimensions in the current user's scenario according to the scenario cross-combination pattern characterized in the multi-dimensional dynamic scenario feature set, and construct a scenario influence weight set; According to the scenario influence weight set, combine with the multi-dimensional dynamic scenario feature set, perform multi-modal feature fusion on the time-sensitive feature vector, the spatial association type feature, and the event response type feature, and generate a composite demand feature vector as the composite scenario demand information; The dynamic priority adjustment strategy includes: when the scenario cross-combination pattern includes the time-space superposition scenario, dynamically allocate the scenario priority based on the urgency of the time segment where the user is located and the accessibility of the core area facilities; When there is the event response type feature corresponding to the strong push demand event in the scenario cross-combination pattern, use the event response type feature as the leading dimension, and increase the influence weight of the corresponding event response type feature according to the urgency of the corresponding strong push demand event; When the public event corresponding to the event response type feature in the scenario cross-combination pattern does not belong to the strong push demand event, adjust the influence weight of the corresponding event response type feature according to the current event interest degree.

[0139] The system of this embodiment can be used to execute the method of any of the above embodiments, and its implementation principle and technical effects are similar, which will not be elaborated here.

Claims

1. A method for recommending a platform user's interests based on artificial intelligence, characterized in that, Including: Obtain user action data and environmental perception data, analyze the user action data and environmental perception data, and construct a unified spatio-temporal semantic network; According to the unified spatio-temporal semantic network, analyze the scene characteristics where the user is located, and determine a multi-dimensional dynamic scene feature set; Analyze the multi-dimensional dynamic scene feature set, determine the composite scene demand information of the current user, and retrieve and push the real-time demand information of the user according to the composite scene demand information.

2. The method according to claim 1, wherein The user action data includes user stay action information and historical search records; The environmental perception data includes real-time traffic status data, meteorological status data, public facility operation status data, and public event data.

3. The method according to claim 2, wherein The analyzing the user action data and environmental perception data to construct a unified spatio-temporal semantic network includes: Based on the user stay action information, analyze the environmental perception data, and screen and determine the local environmental perception data corresponding to the area where the user is located; Perform timestamp alignment processing on the real-time traffic status data, the meteorological status data, the public facility operation status data, and the public event data in the local environmental perception data to determine the local environmental perception synchronization data under the same time axis; Based on the graph neural network algorithm with a fusion attention mechanism, use the user action data as the attention target node, and use the data items in the corresponding local environmental perception synchronization data as neighbor nodes to construct a basic association graph; According to the basic association graph, use the real-time relative relationship between the entity corresponding to the neighbor node and the entity corresponding to the attention target node as the attention weight between the neighbor node and the corresponding attention target node, and determine the attention graph neural network; According to the dynamic attention update strategy, dynamically update the attention graph neural network, and construct the unified spatio-temporal semantic network in real time.

4. The method according to claim 3, characterized in that, The dynamically updating the attention graph neural network includes at least the following steps: According to the environmental perception data, extract the influence range of the entity perception device corresponding to each data item in the environmental perception data; Draw the user movement trajectory information according to the user stay action information; The dynamic attention update strategy includes a passive update mechanism and an active update mechanism; The passive update mechanism: when the user movement trajectory information changes, dynamically update several neighbor nodes corresponding to the current attention target node in the attention graph neural network according to the relative position relationship between the influence range and the change of the user stay action; The active update mechanism: based on the long short-term memory collaborative algorithm, predict the subsequent movement trajectory of the user according to the user movement trajectory information and the historical search records, and preload the environmental perception data nodes involved in the future as the pre-update nodes of the corresponding several neighbor nodes.

5. The method according to claim 4, characterized in that, The analyzing the scene characteristics where the user is located according to the unified spatio-temporal semantic network and determining the multi-dimensional dynamic scene feature set includes: Based on the dynamic attention association relationship between the attention target node and the corresponding neighbor node in the unified spatio-temporal semantic network, extract the interaction characteristics between the user's action and the environment where he is located; Extract time-sensitive features according to the temporal continuity of the interaction features; Extract spatially associated features according to the spatial topological relationship of the interaction features; Extract event-responsive features according to the event trigger conditions of the interaction features; Analyze the spatio-temporal coupling strength among the time-sensitive features, the spatially associated features, and the event-responsive features to identify scene cross-combination patterns; Construct the multi-dimensional dynamic scene feature set according to the scene cross-combination patterns; The cross-combination patterns include time-space superposition scenes, time-event trigger scenes, space-event linkage scenes, and multi-dimensional linkage scenes.

6. The method according to claim 5, wherein The extraction of time-sensitive features according to the temporal continuity of the interaction features includes: Based on the user stay action information and the historical search records, divide the user action time segments and mark the action frequencies and durations in different time periods; According to the real-time traffic state data and the meteorological state data in the same time segment of the environmental perception data, analyze the action frequencies and the durations, identify the influence intensity of the external environment on the user actions in the current time segment, and generate a time-sensitive feature vector as the time-sensitive feature.

7. The method according to claim 6, wherein The extraction of spatially associated features according to the spatial topological relationship of the interaction features includes: According to the user stay action information, construct a dynamic spatial grid centered on the user stay location and divide the core area and the peripheral radiation area; Analyze the local environmental perception synchronization data, extract the location distribution and operation status of public facilities, and combine the real-time traffic state data to identify the spatial topological relationship between the public facilities and the dynamic spatial grid, and generate a facility accessibility map; According to the geographical location keywords involved in the historical search records, extract high-frequency interest space nodes and label the high-frequency interest space nodes into the facility accessibility map to generate a point-of-interest heat map as the spatially associated feature.

8. The method according to claim 5, wherein The extraction of event-responsive features according to the event trigger conditions of the interaction features includes: Analyze the public event data to determine whether the corresponding public event belongs to a strong push demand event; If the public event belongs to the strong push demand event, analyze the spatial overlap degree between the current user movement trajectory information and the predicted movement trajectory and the influence range corresponding to the public event. When the spatial overlap degree is greater than a preset overlap degree threshold, determine the current public event data as the event-responsive feature; If the public event does not belong to the strong push demand event, analyze the user's event interest degree in the current public event according to the historical search records. When the event interest degree is greater than a preset interest degree threshold, determine the current public event data as the event-responsive feature.

9. The method according to claim 8, wherein The analysis of the multi-dimensional dynamic scene feature set to determine the composite scene demand information of the current user includes: Based on the dynamic priority adjustment strategy, according to the scene cross-combination patterns represented in the multi-dimensional dynamic scene feature set, identify the primary and secondary driving dimensions in the scene where the current user is located, and construct a scene influence weight set; Based on the described scenario impact weight set and in combination with the multi-dimensional dynamic scenario feature set, perform multi-modal feature fusion on the time-sensitive feature vector, the spatial association feature, and the event response feature to generate a composite demand feature vector as the composite scenario demand information; The dynamic priority adjustment strategy includes: when the time-space superposition scenario is included in the scenario cross-combination pattern, dynamically allocate scenario priorities based on the urgency of the time segment where the user is located and the accessibility of core area facilities; When there is an event response feature corresponding to the strong push demand event in the scenario cross-combination pattern, taking the event response feature as the dominant dimension, and according to the urgency of the corresponding strong push demand event, increase the impact weight of the corresponding event response feature; When the public event corresponding to the event response feature in the scenario cross-combination pattern does not belong to the strong push demand event, adjust the impact weight of the corresponding event response feature according to the current event interest degree.

10. An AI-based platform user interest recommendation system, characterized in that, It includes: An action perception module, which is used to obtain user action data and environmental perception data, analyze the user action data and environmental perception data, and construct a unified spatio-temporal semantic network; A scenario analysis module, which is used to analyze the scenario features where the user is located according to the unified spatio-temporal semantic network and determine a multi-dimensional dynamic scenario feature set; A push module, which is used to analyze the multi-dimensional dynamic scenario feature set, determine the composite scenario demand information of the current user, and retrieve and push the user's real-time demand information according to the composite scenario demand information.

Images