Knowledge graph node recommendation method and system based on multi-source data fusion

Abstract

The invention relates to the technical field of knowledge graph node recommendation, and provides a knowledge graph node recommendation method and system based on multi-source data fusion, and the method comprises the steps: collecting multi-source data, representing the multi-source data with vectors, and obtaining a vehicle track vector, a traffic network vector and an environment semantic vector; constructing a vehicle node and a parking node, extracting a vehicle parking event of the vehicle node, constructing a time node and an environment node, and generating a hyperedge of the vehicle parking event; determining hyperedge weights of hyperedges of the vehicle parking event, and generating a hypergraph; and taking the vehicle track vector, the traffic network vector and the environment semantic vector as feature vectors, constructing a hypergraph neural network according to the hypergraph, obtaining representation vectors of vehicle nodes and parking nodes, calculating recommendation scores of the vehicle nodes and the parking nodes, and obtaining a result of recommending the vehicle parking nodes by the knowledge graph according to the recommendation scores. According to the invention, the actual fitting degree of the recommendation result of the vehicle parking node can be improved.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph node recommendation technology, specifically to a knowledge graph node recommendation method and system based on multi-source data fusion. Background Technology

[0002] Multi-source data fusion unifies the processing of heterogeneous and cross-domain data, mapping them onto the node-edge structure of the same knowledge graph (KG). Node recommendation, based on the constructed knowledge graph KG, automatically predicts and returns the "most likely relevant or valuable node for the next step" based on user queries, current tasks, or system states. Knowledge graph node recommendation based on multi-source data fusion can capture node attribute features, topological structure, and cross-modal associations, and combine them with scene semantics to dynamically calculate node similarity, achieving more semantically interpretable intelligent recommendations.

[0003] When using knowledge graph node recommendation to recommend vehicle stopping points, it typically relies on fixed entity, relation, and attribute triples. It only performs simple embedding of dynamic scene features such as vehicle location and stopping point timestamps and geographic locations, failing to effectively integrate spatiotemporal dynamic features and environmental semantic features from vehicle trajectory data. When a vehicle enters an unfamiliar area or encounters unexpected road conditions, the lack of real-time correlation modeling of current time, location, and environment easily leads to a disconnect between the recommended results and the vehicle's actual needs. Therefore, it is necessary to address the problem of the recommendation results for vehicle stopping points being disconnected from real-time requirements due to the failure to effectively integrate diverse dynamic features into the knowledge graph during the knowledge graph node recommendation process. Summary of the Invention

[0004] This invention provides a knowledge graph node recommendation method and system based on multi-source data fusion to solve the problem that the recommendation results for vehicle stop nodes are out of sync with real-time requirements due to the failure to effectively integrate diverse dynamic features into the knowledge graph during the knowledge graph node recommendation process. The specific technical solution adopted is as follows: In a first aspect, one embodiment of the present invention provides a knowledge graph node recommendation method based on multi-source data fusion, the method comprising the following steps: Collect multi-source data, including vehicle trajectory data, traffic network data, and environmental semantic data, and convert the multi-source data into vector representations to obtain vehicle trajectory vectors, traffic network vectors, and environmental semantic vectors. Construct vehicle nodes and dwell nodes, extract vehicle dwell events from vehicle trajectory data, construct time nodes and environment nodes based on vehicle dwell events and corresponding environmental semantic data, and generate a hyperedge for the same vehicle dwell event based on the vehicle nodes, dwell nodes, time nodes and environment nodes corresponding to the same vehicle dwell event. Based on the time interval between the vehicle stay event and the recommended time of the vehicle stay node, the distance between the real-time location of the vehicle node and the location of the stay node, the vehicle trajectory vector, the traffic network vector, and the environmental semantic vector, the hyperedge weight of the hyperedge of the vehicle stay event is determined, and a hypergraph is generated by combining the hyperedge. Using vehicle trajectory vectors, traffic network vectors, and environmental semantic vectors as feature vectors, a hypergraph neural network is constructed based on the hypergraph to obtain the representation vectors of vehicle nodes and stationary nodes. Combining the time decay factor of vehicle stationary events, the distance decay factor of stationary nodes, and the environmental semantic vector, recommendation scores for vehicle nodes and stationary nodes are calculated. Based on the recommendation scores, the knowledge graph's recommendation results for vehicle stationary nodes are obtained.

[0005] Furthermore, the vehicle stop event of the vehicle node specifically refers to: The event of a vehicle node moving from one stop node to another stop node.

[0006] Furthermore, the method for constructing the time nodes and environment nodes is as follows: The characteristics of a vehicle stop event, including the time when the vehicle leaves the stop node, the time when the vehicle arrives at the stop node, and the interval between the vehicle leaving and arriving at the two stop nodes, are encapsulated as time nodes. The environmental semantic data within the time range defined by the vehicle's departure time from the stopping node and the vehicle's arrival time from the stopping node are encapsulated as an environment node.

[0007] Furthermore, the method for generating the superedge of the same vehicle stop event is as follows: Connect the vehicle node, dwell node, time node, and environment node corresponding to the same vehicle dwelling event to generate a superedge corresponding to the vehicle dwelling event.

[0008] Furthermore, the method for determining the hyperedge weight of the vehicle stop event based on the time interval between the vehicle stop event and the recommended time of the vehicle stop node, the distance between the real-time location of the vehicle node and the location of the stop node, the vehicle trajectory vector, the traffic network vector, and the environmental semantic vector includes the following specific methods: The time decay factor of a vehicle stop event is determined based on the time interval between the vehicle stop event and the recommended time of the vehicle stop node. The distance between the real-time location of the vehicle node and the location of the stationary node is used to determine the distance attenuation factor of the stationary node; Calculate vehicle dwell events super-edge Hyperedge weight The calculation formula is: in, Indicates a vehicle stopping event Time decay factor; Indicates the node where the pause occurs. Distance attenuation factor; Represents the ReLU activation function; This represents the weight matrix initialized by the Xavier method; Represents the vehicle trajectory vector. Represents a traffic network vector. Represents the context semantic vector. This represents a row vector composed of vehicle trajectory vectors, traffic network vectors, and environmental semantic vectors.

[0009] Furthermore, the method for determining the time decay factor of the vehicle dwell event is as follows: The product of the time interval between the vehicle dwelling event and the recommended time of the vehicle dwelling node and the preset time decay rate is recorded as the time difference feature value of the vehicle dwelling event. The time difference characteristic value of vehicle dwelling events is used as the exponent of an exponential function with the natural constant as the base, and the calculated value of the exponential function is recorded as the time decay factor of vehicle dwelling events.

[0010] Furthermore, the method for determining the distance attenuation factor of the dwell node is as follows: The distance between the real-time location of a vehicle node and the location of a stationary node, which is reached via traffic roads, is recorded as the first distance of the stationary node. The product of the square of the preset distance decay rate and the number 2 is denoted as the distance decay rate characteristic value. The negative of the ratio of the first distance of the dwelling node to the characteristic value of the distance decay rate is taken as the exponent of an exponential function with the natural constant as the base, and the calculated value of the exponential function is denoted as the distance decay factor of the dwelling node.

[0011] Furthermore, the specific method for obtaining the recommendation scores of the vehicle nodes and the stopping nodes is as follows: The cosine similarity between the representation vectors of vehicle nodes and stationary nodes is denoted as the semantic relevance between vehicle nodes and stationary nodes. The product of the time decay factor of the vehicle stay event, the distance decay factor of the stay node, and the environmental semantic vector is denoted as the first product of the hyperedge connecting the vehicle stay event and the stay node; the sum of the degree of the vehicle node and the degree of the stay node is denoted as the first sum of the hyperedge connecting the vehicle stay event and the stay node; the ratio of the first product of the hyperedge connecting the vehicle stay event and the stay node to the first sum is denoted as the first ratio of the hyperedge connecting the vehicle stay event and the stay node; the sum of the first ratios of all the hyperedges connecting the vehicle stay event and the stay node is denoted as the first cumulative sum of the stay node; and the product of the preset spatiotemporal association weight and the first cumulative sum of the stay node is denoted as the recommendation bonus score of the stay node. The sum of the semantic relevance of vehicle nodes and dwell nodes and the recommendation bonus score of dwell nodes is denoted as the recommendation score of vehicle nodes and dwell nodes.

[0012] Furthermore, the specific method for obtaining the recommendation results of vehicle dwell nodes based on the recommendation score using the knowledge graph includes: The recommended scores corresponding to the same vehicle node are arranged in descending order, and the stop nodes corresponding to the recommended scores are arranged in the order of the recommended scores to form a recommended list of stop nodes corresponding to the same vehicle node.

[0013] Secondly, embodiments of the present invention also provide a knowledge graph node recommendation system based on multi-source data fusion, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0014] The beneficial effects of this invention are: After acquiring multi-source data, this application constructs time nodes and environment nodes based on vehicle trajectory data and environmental semantic data. Combining vehicle nodes and dwell nodes, it generates hyperedges for vehicle dwelling events, ensuring that the hyperedges contain the vehicle's actual preference for parking locations and the constraints of the parking environment on parking recommendations. Then, considering that vehicle dwelling behavior is dynamically and significantly influenced by time, space, and environment during the dwell node recommendation process, dwelling events closer to the recommended time are often more valuable. Furthermore, the location and environment of the dwell node at the recommended time also affect the choice of dwelling location. The calculation of hyperedge weights takes into account the potential patterns of vehicle dwelling behavior, providing a foundation for improving the spatiotemporal adaptability and scene fit of the recommendation results. A hypergraph is generated based on the hyperedges and their weights. Finally, the representation vectors of vehicle nodes and dwell nodes are obtained. The representation vector of the vehicle node reflects... The system analyzes the current state and historical behavior patterns of vehicles corresponding to vehicle nodes, as well as their interactions with the surrounding environment. The representation vector of a stopping node reflects the attractiveness, utilization rate, and synergistic effect of the possible stopping locations of the vehicle node with respect to other facilities. Furthermore, it calculates recommendation scores for both vehicle nodes and stopping nodes; higher recommendation scores indicate a more reasonable recommendation of the stopping node. Finally, based on the recommendation scores, it obtains the knowledge graph's recommendation results for stopping nodes. This addresses the problem of the knowledge graph node recommendation process failing to effectively integrate diverse dynamic features into the knowledge graph, leading to a disconnect between the recommended results and real-time needs. The system improves the adaptability of the recommended results to spatiotemporal dynamics and environmental constraints, making the recommendations more aligned with the real-time needs of vehicle users. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a knowledge graph node recommendation method based on multi-source data fusion provided in an embodiment of the present invention. Figure 2 This is a flowchart of the superedge acquisition process provided in one embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 The diagram illustrates a flowchart of a knowledge graph node recommendation method based on multi-source data fusion according to an embodiment of the present invention. The method includes the following steps: Step S001: Collect multi-source data, including vehicle trajectory data, traffic network data, and environmental semantic data. Convert the multi-source data into vector representations to obtain vehicle trajectory vectors, traffic network vectors, and environmental semantic vectors.

[0019] When using knowledge graph node recommendation to recommend vehicle stopping nodes, different data sources have significant differences in data structure, semantic expression, and temporal granularity, as well as in their reliability and impact on the results. It is necessary to preprocess the data from different data sources to effectively integrate them into a unified knowledge representation framework, and then accurately capture the spatiotemporal associations and environmental semantic relationships between the nodes of the knowledge graph through the established knowledge graph.

[0020] First, multi-source data needs to be transformed into a unified feature space to provide standardized input for the constructed temporal hypergraph, thereby improving the recommendation model's adaptability to complex scenarios.

[0021] In this embodiment, vehicle trajectory data, traffic network data, and environmental semantic data are selected to form multi-source data.

[0022] The vehicle trajectory data includes the vehicle's spatiotemporal location, vehicle ID, speed, and direction during movement. The spatiotemporal location includes the vehicle's latitude and longitude coordinates and the time. The vehicle trajectory data is collected in real time through the vehicle's built-in or external GPS. Map matching algorithms such as HMM-Map Matching are used to map the vehicle trajectory data onto the road network, and the vehicle trajectory data is converted into a vector representation to obtain the vehicle trajectory vector.

[0023] Traffic network data refers to geographic data related to the topology and attributes of road networks. Specifically, it includes road segments, intersections, road grades, speed limits, number of lanes, parking lot locations, and charging station locations. This traffic network data is obtained through the open APIs of Gaode or Baidu Maps service providers. A road network topology map is constructed, and road segment attribute features such as length and capacity are extracted. Simultaneously, the locations and capacities of potential stopping points, such as parking lots and charging stations, are labeled. Finally, the traffic network data is converted into a vector representation to obtain the traffic network vector.

[0024] Environmental semantic data refers to unstructured information about the environment in which the user stays. Specifically, it includes weather conditions, time-period labels such as morning and evening rush hours, event information affecting the road environment such as concerts, and types of surrounding facilities such as shopping malls. Weather conditions in the environmental semantic data are obtained through the meteorological bureau's API. Event information affecting the road environment and time-period labels are obtained from a holiday database and manually labeled. Event information affecting the road environment is crawled from social media and news websites using NLP technology. The types of surrounding facilities in the environmental semantic data are obtained through the open APIs of Gaode or Baidu Maps service providers. The environmental semantic data is then converted into vector representations to obtain environmental semantic vectors.

[0025] At this point, the vehicle trajectory vector, traffic network vector, and environmental semantic vector are obtained.

[0026] Step S002: Construct vehicle nodes and dwell nodes; extract vehicle dwell events from vehicle nodes based on vehicle trajectory data; construct time nodes and environment nodes based on vehicle dwell events and corresponding environmental semantic data; and generate a hyperedge for the same vehicle dwell event based on the vehicle node, dwell node, time node, and environment node corresponding to the same vehicle dwell event.

[0027] The actual driving behavior of vehicles corresponding to vehicle nodes is subject to multiple constraints, including temporal dynamics, spatial correlation, and environmental semantics. Traditional binary relations in graph models alone cannot capture these multiple constraints. Temporal dynamics include differences in parking demand due to time-of-day variations such as rush hour; spatial correlation includes differences in parking demand due to the synergistic effect of nearby parking lots; and environmental semantics includes differences in parking demand due to events affecting road conditions, such as weather conditions influencing parking choices or concerts. Therefore, the entities corresponding to vehicles and their potential parking locations are abstracted as vehicle nodes and parking nodes, respectively.

[0028] Construct vehicle nodes and dwell nodes. Each vehicle node corresponds to a vehicle and is associated with vehicle trajectory data; each dwell node corresponds to a possible location where a vehicle may stay, such as a parking lot, charging station, gas station, shopping mall, office building, and residential community.

[0029] After obtaining vehicle nodes and stopping nodes, a multi-dimensional relationship network can be formed by connecting vehicle nodes and stopping nodes through hyperedges. The influence of vehicle trajectory data time on vehicle stopping location, the distance between the vehicle's current location and the stopping location, and the vehicle's actual preference for parking location are all integrated into the calculation of hyperedge weights. Based on the multi-dimensional relationship network with calculated hyperedge weights, a hypergraph neural network is constructed to realize the dynamic aggregation and propagation of features. Finally, vehicle stopping node recommendation is realized based on the hypergraph neural network.

[0030] First, generate the hyperedge.

[0031] Vehicle stop events are extracted from vehicle trajectory data. A vehicle stop event is an event in which a vehicle node moves from one stop node to another. The characteristics of a vehicle stop event include the stop node that the vehicle node leaves, the stop node that the vehicle node arrives at, the time that the vehicle node leaves the stop node, the time that the vehicle node arrives at the stop node, and the interval between the vehicle node leaving and arriving at the two stop nodes.

[0032] The characteristics of a vehicle stop event, including the time when the vehicle leaves the stop node, the time when the vehicle arrives at the stop node, and the interval between the vehicle leaving and arriving at the two stop nodes, are encapsulated as time nodes.

[0033] The environmental semantic data within the time range defined by the vehicle's departure time from the stopping node and the vehicle's arrival time from the stopping node are encapsulated as an environment node.

[0034] It is understandable that time nodes and environment nodes do not correspond to entities, but are used to describe the temporal and environmental characteristics of the entities corresponding to vehicle nodes and stationary nodes.

[0035] Connect the vehicle node, dwell node, time node, and environment node corresponding to the same vehicle dwell event to generate a hyperedge corresponding to the vehicle dwell event. The flowchart for obtaining the hyperedge is as follows: Figure 2 As shown.

[0036] At this point, the hyperedge for each vehicle stop event is obtained.

[0037] Step S003: Based on the time interval between the vehicle stay event and the recommended time of the vehicle stay node, the distance between the real-time location of the vehicle node and the location of the stay node, the vehicle trajectory vector, the traffic network vector, and the environmental semantic vector, determine the hyperedge weight of the hyperedge of the vehicle stay event, and generate a hypergraph by combining the hyperedge.

[0038] The weight of a hyperedge is a learnable scalar value used to measure the importance or influence of each hyperedge in the information aggregation process, and is generally defined based on time or spatial distance. However, considering that vehicle stopping behavior is dynamically and significantly affected by time, space, and environment during the vehicle stopping node recommendation process, vehicle stopping events closer to the recommended time are often more valuable. Furthermore, the location and environment of the recommended time also influence the choice of vehicle stopping location. Therefore, it is necessary to determine the weight values ​​of hyperedges using vehicle trajectory vectors, traffic network vectors, and environmental semantic vectors. This provides a foundation for the constructed hypergraph neural network to accurately capture the potential patterns of vehicle stopping behavior and improve the spatiotemporal adaptability and scene fit of the recommendation results.

[0039] First, the behavior patterns of vehicle users are time-dependent. By measuring the time interval between vehicle dwelling events and the recommended time of vehicle dwelling nodes, we can capture the time sensitivity of the behavior of vehicle users corresponding to vehicle nodes.

[0040] The product of the time interval between the vehicle stop event and the recommended time of the vehicle stop node and the preset time decay rate is recorded as the time difference feature value of the vehicle stop event. The time difference feature value of the vehicle stop event is used as the exponent of an exponential function with the natural constant as the base. The calculated value of the exponential function is recorded as the time decay factor of the vehicle stop event.

[0041] The time decay rate should be greater than or equal to 0.1 and less than or equal to 10. In this embodiment, the time decay rate is set to 0.3. In the process of calculating the time decay factor of vehicle stay events, the nonlinear decay characteristics of the exponential function are used to assign a larger time decay factor value to vehicle stay events that are closer to the recommended time of the vehicle stay node, so as to provide a stronger influence of vehicle stay events on the recommended time of the vehicle stay node in the future. At the same time, the influence of vehicle stay events that are farther away from the recommended time of the vehicle stay node is retained.

[0042] The smaller the time decay factor, the greater the interval between the vehicle dwelling event and the recommended time of the vehicle dwelling node corresponding to the time decay factor, and the less reference value it has for the vehicle dwelling event and the recommended time of the vehicle dwelling node. At this time, the weight of the hyperedge corresponding to the vehicle dwelling event is smaller.

[0043] Due to limitations imposed by road networks and geographical distances, vehicle users generally prefer to park at locations closer to the target vehicle. Therefore, a distance attenuation factor for the parking node is determined based on the distance between the vehicle node's location and the parking node's location.

[0044] The distance between the real-time location of the vehicle node and the location of the stationary node, reached via traffic roads, is denoted as the first distance of the stationary node. The product of the square of the preset distance decay rate and the number 2 is denoted as the distance decay rate characteristic value. The negative of the ratio of the first distance of the stationary node to the distance decay rate characteristic value is taken as the exponent of an exponential function with the natural constant as the base. The calculated value of the exponential function is denoted as the distance decay factor of the stationary node.

[0045] The distance decay rate should be greater than or equal to 0.5 and less than or equal to 5. In this embodiment, the time decay rate is set to 1.5. The real-time vehicle location of the vehicle node is obtained by the GPS installed on the vehicle node.

[0046] The smaller the distance decay factor, the farther the distance between the dwell node corresponding to the distance decay factor and the real-time location of the vehicle is reached through traffic roads. The less reliable the dwell node corresponding to the distance decay factor is for vehicle users as a reference for the vehicle's dwell location, and the smaller the weight of the hyperedge corresponding to the dwell node.

[0047] The superedge weights of vehicle stop events are determined based on vehicle trajectory vectors, traffic network vectors, environmental semantic vectors, time decay factors of vehicle stop events, and distance decay factors of stop nodes.

[0048] in, Indicates a vehicle stopping event super-edge The weight of the superedge; Indicates a vehicle stopping event Time decay factor; Indicates the node where the pause occurs. Distance attenuation factor; This represents the ReLU activation function, which can filter out negative associations; This represents the weight matrix, which is initialized using the Xavier method in this embodiment. Represents the vehicle trajectory vector. Represents a traffic network vector. Represents the context semantic vector. This represents a row vector composed of vehicle trajectory vectors, traffic network vectors, and environmental semantic vectors.

[0049] The hypergraph is generated by taking the weight of the hyperedge of the vehicle stop event as the value of the weight of the hyperedge of the vehicle stop event.

[0050] Hypergraphs can reflect the dynamic dwell patterns of vehicles at each vehicle node, providing a structured knowledge base for recommending vehicle dwell nodes.

[0051] At this point, the supermap has been obtained.

[0052] Step S004: Using vehicle trajectory vector, traffic network vector, and environmental semantic vector as feature vectors, construct a hypergraph neural network based on the hypergraph to obtain the representation vectors of vehicle nodes and stationary nodes. Combine the time decay factor of vehicle stationary events, the distance decay factor of stationary nodes, and the environmental semantic vector to calculate the recommendation scores of vehicle nodes and stationary nodes. Based on the recommendation scores, obtain the knowledge graph's recommendation results for vehicle stationary nodes.

[0053] Vehicle trajectory vectors, traffic network vectors, and environmental semantic vectors are used as extracted feature vectors. Based on the hypergraph, a multi-layer architecture of the HGNN hypergraph neural network is used for feature propagation, enabling deep embedding representations of vehicle nodes, station nodes, time nodes, and environmental nodes that integrate neighborhood information, thereby obtaining the representation vectors of vehicle nodes and station nodes.

[0054] The construction process of the hypergraph neural network is a well-known technology and will not be elaborated further. Specifically, the deep embedding representation of vehicle nodes, stationary nodes, time nodes, and environment nodes, which integrates neighborhood information, includes the node's own attribute features, structural position, and relational information within the hypergraph. In each layer of the hypergraph neural network, vehicle nodes, stationary nodes, time nodes, and environment nodes collect information about their respective hyperedges, calculate the importance weights of different hyperedges through an attention mechanism, and pass information from the source node to the target node according to the message passing function, the node type, and the relation type contained in the hyperedge. All incoming messages are aggregated, the node's embedding representation is updated, and through multi-layer propagation, the node captures the semantic information of higher-order neighborhoods.

[0055] It is understandable that the representation vector of a vehicle node can reflect the current state and historical behavior patterns of the vehicle corresponding to the vehicle node, as well as the interaction between the vehicle node and the surrounding environment; the representation vector of a stationary node can reflect the attractiveness of the possible stationary location of the vehicle corresponding to the stationary node to the vehicle node, its utilization rate, and its synergistic effect with other facilities.

[0056] Recommendation scores for vehicle nodes and resting nodes are calculated based on the similarity of their representation vectors, the time decay factor of vehicle resting events, the distance decay factor of resting nodes, and the environmental semantic vector.

[0057] The cosine similarity of the representation vectors of vehicle nodes and stationary nodes is denoted as the semantic relevance between them. The product of the time decay factor of the stationary event, the distance decay factor of the stationary node, and the environmental semantic vector is denoted as the first product of the hyperedges connecting the stationary event and the stationary node. The sum of the degrees of the vehicle node and the stationary node is denoted as the first sum of the hyperedges connecting them. The ratio of the first product of the hyperedges connecting them to the first sum is denoted as the first ratio of the hyperedges connecting them. The cumulative sum of the first ratios of all the hyperedges connecting them is denoted as the first cumulative sum of the stationary nodes. The product of the preset spatiotemporal association weight and the first cumulative sum of the stationary nodes is denoted as the recommendation bonus score for the stationary node. The sum of the semantic relevance between the vehicle node and the stationary node and the recommendation bonus score for the stationary node is denoted as the recommendation score for the vehicle node and the stationary node.

[0058] The degree of a node is the number of edges directly connected to that node. The concept and calculation of node degree are well-known techniques and will not be elaborated further. The sum of the degrees of vehicle nodes and the degrees of resting nodes is the sum of the number of edges directly connected to the degree of vehicle nodes and the number of edges directly connected to resting nodes. The purpose of the sum of the degrees of vehicle nodes and resting nodes is to normalize the data and mitigate the bias problem of popular nodes. For example, it assigns lower weights to full parking lots to avoid giving excessively high recommendation scores to vehicles associated with full parking lots, preventing vehicles from arriving at full parking lots from being unable to park properly, and avoiding over-recommendation of overly popular resting nodes.

[0059] The spatiotemporal correlation weight should be greater than 0 and less than or equal to 2. In this embodiment, the spatiotemporal correlation weight is set to 1.5.

[0060] The greater the semantic relevance, the more significant the similarity between the vehicle nodes and the stopping nodes in terms of semantics, function, or user preferences. The higher the recommendation score of both the vehicle node and the stopping node, the more reasonable it is to recommend the stopping node corresponding to the recommended vehicle node.

[0061] The recommended scores corresponding to the same vehicle node are arranged in descending order, and the stop nodes corresponding to the recommended scores are arranged in the order of the recommended scores to form a recommended list of stop nodes corresponding to the same vehicle node.

[0062] It is understandable that the recommended list of stopping nodes is the result of vehicle stopping node recommendations, and the recommendation level of each stopping node decreases from top to bottom according to the order of the recommended list of stopping nodes.

[0063] At this point, the knowledge graph node recommendation results have been obtained.

[0064] Based on the same inventive concept as the above methods, embodiments of the present invention also provide a knowledge graph node recommendation system based on multi-source data fusion, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described knowledge graph node recommendation methods based on multi-source data fusion.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge graph node recommendation method based on multi-source data fusion, characterized in that, The method includes the following steps: Collect multi-source data, including vehicle trajectory data, traffic network data, and environmental semantic data, and convert the multi-source data into vector representations to obtain vehicle trajectory vectors, traffic network vectors, and environmental semantic vectors. Construct vehicle nodes and dwell nodes, extract vehicle dwell events from vehicle trajectory data, construct time nodes and environment nodes based on vehicle dwell events and corresponding environmental semantic data, and generate a hyperedge for the same vehicle dwell event based on the vehicle nodes, dwell nodes, time nodes and environment nodes corresponding to the same vehicle dwell event. Based on the time interval between the vehicle stay event and the recommended time of the vehicle stay node, the distance between the real-time location of the vehicle node and the location of the stay node, the vehicle trajectory vector, the traffic network vector, and the environmental semantic vector, the hyperedge weight of the hyperedge of the vehicle stay event is determined, and a hypergraph is generated by combining the hyperedge. Using vehicle trajectory vectors, traffic network vectors, and environmental semantic vectors as feature vectors, a hypergraph neural network is constructed based on the hypergraph to obtain the representation vectors of vehicle nodes and stationary nodes. Combining the time decay factor of vehicle stationary events, the distance decay factor of stationary nodes, and the environmental semantic vector, recommendation scores for vehicle nodes and stationary nodes are calculated. Based on the recommendation scores, the knowledge graph's recommendation results for vehicle stationary nodes are obtained.

2. The knowledge graph node recommendation method based on multi-source data fusion according to claim 1, characterized in that, The specific vehicle dwell events of the vehicle node are as follows: The event of a vehicle node moving from one stop node to another stop node.

3. The knowledge graph node recommendation method based on multi-source data fusion according to claim 1, characterized in that, The method for constructing the time nodes and environment nodes is as follows: The characteristics of a vehicle stop event, including the time when the vehicle leaves the stop node, the time when the vehicle arrives at the stop node, and the interval between the vehicle leaving and arriving at the two stop nodes, are encapsulated as time nodes. The environmental semantic data within the time range defined by the vehicle's departure time from the stopping node and the vehicle's arrival time from the stopping node are encapsulated as an environment node.

4. The knowledge graph node recommendation method based on multi-source data fusion according to claim 1, characterized in that, The method for generating the superedge of the same vehicle stop event is as follows: Connect the vehicle node, dwell node, time node, and environment node corresponding to the same vehicle dwelling event to generate a superedge corresponding to the vehicle dwelling event.

5. The knowledge graph node recommendation method based on multi-source data fusion according to claim 1, characterized in that, The method for determining the hyperedge weight of a vehicle stop event based on the time interval between the vehicle stop event and the recommended time of the vehicle stop node, the distance between the real-time location of the vehicle node and the location of the stop node, the vehicle trajectory vector, the traffic network vector, and the environmental semantic vector includes the following specific methods: The time decay factor of a vehicle stop event is determined based on the time interval between the vehicle stop event and the recommended time of the vehicle stop node. The distance between the real-time location of the vehicle node and the location of the stationary node is used to determine the distance attenuation factor of the stationary node; Calculate vehicle dwell events super-edge Hyperedge weight The calculation formula is: in, Indicates a vehicle stopping event Time decay factor; Indicates the node where the pause occurs. Distance attenuation factor; Represents the ReLU activation function; This represents the weight matrix initialized by the Xavier method; Represents the vehicle trajectory vector. Represents a traffic network vector. Represents the context semantic vector. This represents a row vector composed of vehicle trajectory vector, traffic network vector, and environmental semantic vector.

6. The knowledge graph node recommendation method based on multi-source data fusion according to claim 5, characterized in that, The method for determining the time decay factor of the vehicle dwell event is as follows: The product of the time interval between the vehicle dwelling event and the recommended time of the vehicle dwelling node and the preset time decay rate is recorded as the time difference feature value of the vehicle dwelling event. The time difference characteristic value of vehicle dwelling events is used as the exponent of an exponential function with the natural constant as the base, and the calculated value of the exponential function is recorded as the time decay factor of vehicle dwelling events.

7. The knowledge graph node recommendation method based on multi-source data fusion according to claim 5, characterized in that, The method for determining the distance attenuation factor of the dwell node is as follows: The distance between the real-time location of a vehicle node and the location of a stationary node, which is reached via traffic roads, is recorded as the first distance of the stationary node. The product of the square of the preset distance decay rate and the number 2 is denoted as the distance decay rate characteristic value. The negative of the ratio of the first distance of the dwelling node to the characteristic value of the distance decay rate is taken as the exponent of an exponential function with the natural constant as the base, and the calculated value of the exponential function is denoted as the distance decay factor of the dwelling node.

8. The knowledge graph node recommendation method based on multi-source data fusion according to claim 1, characterized in that, The specific method for obtaining the recommendation scores of the vehicle node and the stop node is as follows: The cosine similarity between the representation vectors of vehicle nodes and stationary nodes is denoted as the semantic relevance between vehicle nodes and stationary nodes. The product of the time decay factor of the vehicle stay event, the distance decay factor of the stay node, and the environmental semantic vector is denoted as the first product of the hyperedge connecting the vehicle stay event and the stay node; the sum of the degree of the vehicle node and the degree of the stay node is denoted as the first sum of the hyperedge connecting the vehicle stay event and the stay node; the ratio of the first product of the hyperedge connecting the vehicle stay event and the stay node to the first sum is denoted as the first ratio of the hyperedge connecting the vehicle stay event and the stay node; the sum of the first ratios of all the hyperedges connecting the vehicle stay event and the stay node is denoted as the first cumulative sum of the stay node; and the product of the preset spatiotemporal association weight and the first cumulative sum of the stay node is denoted as the recommendation bonus score of the stay node. The sum of the semantic relevance of vehicle nodes and dwell nodes and the recommendation bonus score of dwell nodes is denoted as the recommendation score of vehicle nodes and dwell nodes.

9. The knowledge graph node recommendation method based on multi-source data fusion according to claim 1, characterized in that, The specific method for obtaining the recommendation results of vehicle dwell nodes based on the recommendation score using the knowledge graph includes: The recommended scores corresponding to the same vehicle node are arranged in descending order, and the stop nodes corresponding to the recommended scores are arranged in the order of the recommended scores to form a recommended list of stop nodes corresponding to the same vehicle node.

10. A knowledge graph node recommendation system based on multi-source data fusion, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the knowledge graph node recommendation method based on multi-source data fusion as described in any one of claims 1-9.

Images