Passenger multi-modal travel path recommendation system based on knowledge graph and machine learning

Abstract

This invention relates to the fields of intelligent transportation and information processing technology, specifically a multimodal travel route recommendation system for passengers based on knowledge graphs and machine learning. Deployed on a distributed cloud computing architecture, the system comprises a multi-source heterogeneous data acquisition and preprocessing layer, a travel knowledge graph construction and management layer, a dynamic spatiotemporal semantic perception layer, a knowledge-guided deep recommendation decision layer, and an interactive route presentation and feedback layer. These layers are interconnected through a high-speed communication protocol to achieve closed-loop linkage. The system integrates static traffic, real-time operations, passenger attributes, geographical environment, and unstructured text data to construct a semantically rich travel knowledge graph, and combines graph convolution and temporal models to perceive the travel context. The recommendation engine, based on a multi-task neural network, completes route generation, personalized ranking, and interpretive reasoning. The system supports interactive feedback and dynamic optimization, making it suitable for intelligent transportation and personalized travel services.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and information processing technology, and in particular to a passenger multimodal travel route recommendation system based on knowledge graphs and machine learning. Background Technology

[0002] With the rapid development of intelligent transportation and personalized travel services, providing accurate and efficient multimodal travel route recommendations for passengers has become a research hotspot. Existing route recommendation systems typically combine user preferences, historical data, and real-time traffic information for optimization, but they still have shortcomings in deep semantic understanding, multi-source heterogeneous data fusion, and dynamic personalized modeling, making it difficult to fully capture the complex travel intentions and environmental changes of passengers. A search revealed a multimodal travel route recommendation method and apparatus (publication number CN118427448B), which constructs a multimodal traffic hypernetwork graph, combines historical user travel data, uses a random forest model to determine the weights of preference indicators, and employs a Double Q-learning algorithm for route planning. However, this scheme mainly relies on structured traffic station data and statistical preference indicators, lacking the ability to integrate unstructured semantic information and failing to construct a deep entity relationship network, resulting in a relatively shallow understanding of users' complex travel intentions. Furthermore, its machine learning model and reinforcement learning are loosely integrated, failing to achieve end-to-end joint optimization, thus limiting the adaptability and generalization ability of the recommendation system.

[0003] On the other hand, a multimodal travel mode fusion recommendation method based on dynamic transportation networks, with publication number CN114461934B, was disclosed. This method constructs a heterogeneous information network from user spatiotemporal trajectories, uses a meta-path-guided graph neural network to dynamically model users and travel modes, and learns personalized preferences through MLP. Although this method introduces a graph neural network to capture complex interaction relationships, its "heterogeneous information network" is essentially still based on feature aggregation according to predefined meta-paths, lacking the ontological constraints, logical reasoning, and cross-domain knowledge association capabilities of knowledge graphs. Furthermore, this scheme does not explicitly integrate external knowledge, nor does it utilize knowledge graphs to perform unified semantic modeling of multimodal travel entities, resulting in insufficient robustness of recommendations when facing new scenarios or sparse data.

[0004] The aforementioned problems indicate that existing technologies for multimodal travel route recommendation generally suffer from insufficient semantic understanding depth, weak multi-source heterogeneous data fusion capabilities, lack of interpretable knowledge reasoning mechanisms, and limited adaptability to personalized modeling and dynamic environments. Therefore, this invention proposes a multimodal travel route recommendation system for passengers based on knowledge graphs and machine learning. The aim is to construct a travel knowledge graph that integrates multi-source data and combine it with deep learning models to achieve semantic perception, contextual understanding, and personalized reasoning, thereby providing a more intelligent, accurate, and interpretable travel route recommendation service. Summary of the Invention

[0005] To address the shortcomings of existing multimodal travel recommendation systems, such as insufficient semantic understanding depth, weak multi-source heterogeneous data fusion capabilities, lack of interpretable reasoning mechanisms, and limited modeling accuracy in dynamic environments, this invention provides a passenger multimodal travel route recommendation system based on knowledge graphs and machine learning.

[0006] The technical solution adopted in this invention is as follows:

[0007] A multimodal travel route recommendation system for passengers based on knowledge graphs and machine learning is characterized in that the system is deployed on a distributed cloud computing cluster in terms of physical architecture, including a multi-source heterogeneous data acquisition and preprocessing layer, a travel knowledge graph construction and management layer, a dynamic spatiotemporal semantic perception layer, a knowledge-guided deep recommendation decision layer, and an interactive route presentation and feedback layer; each layer realizes the closed-loop flow of data and instructions through a high-speed internal bus and communication protocol.

[0008] The multi-source heterogeneous data acquisition and preprocessing layer is configured to access and integrate static traffic facility data, dynamic operational status data, passenger individual attribute data, geographic environment information, and open text data from the internet in real time via standard interfaces. The static traffic facility data includes rail transit station coordinates, bus route directions, road topology connections, and the rated capacity of parking lots and charging piles. The dynamic operational status data is obtained through real-time sensor acquisition and includes real-time pedestrian flow, vehicle punctuality rate, current average speed, and congestion index at each traffic node. The passenger individual attribute data is uploaded by the user terminal after anonymization and includes historical travel frequency, sensitivity scores for different modes of transportation, real-time location coordinates, and preset preference tendencies. The geographic environment information includes the operating attributes, operating hours, and rating levels of points of interest (POIs), as well as real-time regional meteorological parameters, including precipitation, visibility, and wind speed. The open text data from the internet is obtained through web crawlers or public APIs and covers emergency traffic control announcements, notices of major social events, and user reviews of the service quality of transportation hubs.

[0009] The multi-source heterogeneous data acquisition and preprocessing layer includes a structured processing unit. This unit utilizes natural language processing (NLP) technology to extract entities and relations from the aforementioned open internet text data. The entity extraction process employs a pre-trained language model to perform sequence labeling on the text, identifying four core elements: location, time, event type, and impact level. The relation extraction process, based on dependency parsing, establishes semantic associations between events and geographic coordinates and transportation routes, transforming unstructured text into structured triples that conform to the ontology definition, providing standardized input for subsequent dynamic updates of the knowledge graph.

[0010] The travel knowledge graph construction and management layer is configured to build and maintain a high-dimensional travel knowledge network based on a pre-defined transportation domain ontology model. The ontology model defines multiple entity classes, including "passengers," "vehicles," "transportation stations," "route segments," "points of interest," "emergencies," and "environmental factors," as well as the attribute constraints between these entity classes. These attribute constraints include the attribution of stations to their respective routes, the adjacency relationships between route segments, the spatial reachability between passengers and stations, and the constraints imposed by emergencies on specific transportation nodes.

[0011] The travel knowledge graph construction and management layer includes a knowledge fusion unit for aligning attributes from different sources that point to the same physical entity. This knowledge fusion unit determines the uniqueness of entities by calculating spatial distance similarity, name character edit distance, and semantic vector cosine similarity, eliminating redundancy and completing missing attributes. Furthermore, the travel knowledge graph management layer uses a distributed graph database for storage, employing an attribute graph structure to represent the complex topological relationships between nodes. This graph database supports deep reasoning based on graph traversal, capable of calculating the path connectivity between any two points and their reachability strength in a specific semantic context.

[0012] The dynamic spatiotemporal semantic perception layer is configured to map high-dimensional discrete nodes in the aforementioned travel knowledge graph to a low-dimensional continuous dense vector space using embedding representation techniques. This layer includes a feature encoding module that extracts structured features from the knowledge graph using a graph convolutional network. In each convolutional iteration, a specific node aggregates the attribute information of its first-order and multi-order neighbor nodes and assigns different weight coefficients based on the type of connecting edges. This process captures the spatial local dependencies and semantic relationships within the transportation network.

[0013] Furthermore, the feature encoding module combines long short-term memory networks to process time-series data, encoding the periodic fluctuations in traffic flow and the attenuation process of the impact of sudden events into the dynamic feature vectors of nodes. The final generated node embedding vectors contain features across four dimensions: static attributes, spatial topology, temporal dynamics, and semantic environment. The dynamic spatiotemporal semantic perception layer also includes a context constraint module, which dynamically adjusts the activation thresholds of different edge weights in the knowledge graph based on the current time, weather conditions, and regional congestion levels, thereby achieving a keen perception of the real-time travel environment.

[0014] The knowledge-guided deep recommendation decision layer is configured to implement end-to-end path recommendation logic using a multi-task neural network framework. This layer receives node embedding vectors from the perception layer and real-time passenger request parameters as input. Internally, it integrates a path candidate set generation module, a personalized ranking module, and an interpretive reasoning module.

[0015] The path candidate set generation module, based on an improved graph search strategy, retrieves a set of candidate paths that meet basic spatiotemporal constraints within a multimodal transportation network. This improved graph search strategy not only considers physical distance but also automatically avoids areas affected by unforeseen events or harsh environments by introducing a semantic cost factor from the knowledge graph. The semantic cost factor is calculated in real-time based on the correlation strength between "unforeseen event" entities and "path segment" entities in the knowledge graph.

[0016] The personalized ranking module employs a deep interest evolution network to provide refined scoring for candidate paths. This module captures the correlation between a traveler's current travel intention and historical behavioral patterns through an attention mechanism. During computation, the system performs fully connected layer operations on the feature vectors of candidate paths, the traveler's historical trajectory features, and real-time preference vectors. The loss function of the personalized ranking module combines four dimensions: path time, economic cost, transfer convenience, and environmental comfort. Through training and optimization, the recommendation results closely approximate the traveler's actual choices.

[0017] The interpretive reasoning module is configured to extract key reasoning chains from the knowledge graph while generating recommended routes. These reasoning chains demonstrate the logical connections between the recommended routes and passenger preferences, real-time events, and POI attributes. For example, when the system suggests that passengers choose rail transit instead of surface public transport due to rainfall, the interpretive reasoning module traces and extracts the logical connections between "rainfall event - leading to - increased surface traffic congestion level" and "rail transit - possessing - all-weather operation attributes" from the knowledge graph, and transforms them into a semantic description understandable to the user, thereby improving the interpretability and credibility of the recommendation decision.

[0018] The interactive path presentation and feedback layer is configured to display recommended solutions in a multi-dimensional visualization on the user terminal and monitor passengers' clicks, acceptance, and actual implementation of the recommended results in real time. This layer includes a real-time feedback loop that feeds back passenger preference corrections and abnormal offset data to the deep recommendation decision layer. This feedback loop uses an online learning mechanism and an incremental gradient update method to fine-tune the connection weights in the deep learning model and knowledge graph, achieving adaptive evolution of system performance.

[0019] In a preferred embodiment of the present invention, the system further includes a global consistency verification unit. This unit is deployed in the travel knowledge graph construction and management layer to detect and correct logical contradictions in the knowledge graph by executing preset ontology logic rules when multi-source data is updated frequently. These logical contradictions include, but are not limited to, coordinate offsets of the same station exceeding a threshold under different traffic modes, or unreasonable spatiotemporal jumps of the same mode of transportation within the same time period.

[0020] Furthermore, as a preferred embodiment of the present invention, the multimodal travel route recommendation system adopts a hierarchical parallel processing architecture. In the data preprocessing stage, the system utilizes a stream processing engine to perform real-time aggregation and noise reduction on high-frequency sensor data; in the knowledge reasoning stage, the system uses distributed parallel traversal operators of a graph database to complete semantic search of multi-hop paths within milliseconds. This architecture ensures that the system maintains extremely low response latency even under large-scale road networks and tens of millions of concurrent user requests.

[0021] Furthermore, as a preferred embodiment of the present invention, the system specifically includes a hub-within-a-hub micro-path modeling subsystem for multimodal transition scenarios. This subsystem establishes refined station internal triples in the knowledge graph, detailing the walking distance, floor span, vertical elevator location, and average transfer time required for transfers between different lines. When calculating the total path cost, the deep recommendation decision layer automatically calls upon the refined data from this subsystem, addressing the problem of overly coarse descriptions of multimodal connection points in existing technologies and improving the accuracy of recommended paths in actual execution.

[0022] Furthermore, as a preferred embodiment of the present invention, the personalized ranking module introduces a pre-training mechanism based on contrastive learning. By comparing and modeling massive amounts of unlabeled historical travel trajectories, the system learns general spatiotemporal representation features, enhancing its robustness in the face of new users or cold start scenarios. When the system encounters a user with no prior travel history, it can generate reasonable recommendation results based on the user's basic attributes in the knowledge graph and the group profile of their group through knowledge association transfer.

[0023] Furthermore, in a preferred embodiment of the present invention, the system further includes a sudden event simulation and evolution prediction unit. This unit utilizes a recurrent neural network to learn the evolution patterns of historical sudden events recorded in the knowledge graph. Upon initial detection of a new sudden event, it can predict its diffusion impact on surrounding traffic flow and node connectivity within the next few hours. The prediction results are input as dynamic constraints to the deep recommendation decision layer in real time, enabling the recommendation system to possess forward-looking obstacle avoidance capabilities.

[0024] In this invention, the multi-source data acquisition and preprocessing layer also performs a feature normalization process. For continuous physical values, such as vehicle speed, temperature, and pedestrian flow, the system uses range normalization or standard deviation normalization to scale them to a uniform numerical range. For discrete semantic features, such as weather phenomenon classifications and traffic control types, the system uses one-hot encoding or word embedding methods to convert them into fixed-length numerical vectors. This processing method eliminates the negative impact of different units on the convergence speed and prediction accuracy of subsequent machine learning models.

[0025] As a core technical detail of this invention, the entity relationships in the travel knowledge graph not only include deterministic logical relationships but also probability-weighted association relationships. These association relationships are obtained through statistical learning from massive historical samples and are used to describe, for example, "the correlation strength between the passenger flow at a shopping mall on a Saturday afternoon and the entry volume of nearby subway stations." This semi-deterministic relationship network enables the knowledge graph to possess statistical inference capabilities for handling fuzzy traffic phenomena, in addition to rigorous logical reasoning.

[0026] In its implementation, the system distributes passenger requests to multiple recommendation engine instances via a load balancer. Each recommendation engine instance, while performing path search, concurrently executes knowledge graph-based logical constraint filtering and neural network-based ranking and scoring. The filtering process rapidly narrows the search space based on preset hard constraints, such as maximum number of transfers and longest walking distance. The scoring process then performs deep semantic matching and preference alignment within the narrowed candidate set. Through this two-stage collaborative work, an optimal balance between recommendation efficiency and recommendation quality is achieved.

[0027] The interactive path presentation and feedback layer also supports dynamic adjustment of multiple criteria. Passengers can adjust the weight sliders for time priority, cost priority, or comfort priority in real time through the terminal interface. When a change in the weight slider is detected, the system does not need to re-retrieve all the data, but updates the recommended path within milliseconds by fine-tuning the output fusion ratio of the multi-task head in the deep recommendation decision layer.

[0028] Furthermore, the system features robust engineering implementation for data security and passenger privacy protection. At the data acquisition and preprocessing layer, all sensitive information involving individual identity is de-identified using hash algorithms. At the knowledge graph storage layer, the system employs a role-based access control mechanism to ensure that only authorized modules can access specific attribute data. Simultaneously, the training process for all models is conducted in an encrypted sandbox environment to prevent reverse engineering and leakage of training data.

[0029] The system described in this invention fundamentally solves the "black box" dilemma of traditional route recommendation systems when handling complex intentions and their vulnerability to sudden disturbances by constructing a global travel knowledge cockpit that integrates static topology, dynamic flow field, and deep semantics. It not only improves the accuracy of recommendations but also achieves transparency in the recommendation logic through semantic association.

[0030] The advantages of this invention compared to the prior art are as follows:

[0031] First, by constructing a travel knowledge graph that integrates multi-source heterogeneous data, this invention breaks through the limitations of traditional recommendation systems that rely solely on structured sites. By introducing unstructured semantic information such as POI attributes, weather, emergencies, and user reviews, it achieves a comprehensive perception of the travel environment. This deep interconnection of entities enables the system to discern the complex semantic relationships hidden behind the physical road network, significantly enhancing the depth of understanding of travelers' travel intentions.

[0032] Second, this invention achieves deep coupling between knowledge graph reasoning logic and deep learning prediction capabilities. Unlike traditional loose combinations, this scheme directly injects the logical constraints of the knowledge graph into the input of the neural network through graph embedding technology, while simultaneously utilizing attention mechanisms to achieve knowledge-guided feature selection within the model. This end-to-end joint optimization mode leverages the powerful nonlinear fitting capabilities of deep learning while preserving the rigorous logical consistency of the knowledge graph, effectively improving the adaptability of the recommendation system.

[0033] Third, this invention introduces an interpretable reasoning mechanism based on knowledge graphs. By attaching reasoning chains extracted from knowledge graphs to the recommendation results, the system can clearly show travelers the logical basis for "why this route is recommended." This transparent decision-making process not only enhances user trust but also provides system administrators with direct tools for fault tracing and logic optimization, greatly improving system maintainability.

[0034] Fourth, this invention possesses strong adaptability and robustness to dynamic environments. By updating event nodes and dynamic edge weights in the knowledge graph in real time, and combining this with a sudden event evolution prediction unit, the system can predict the impact of traffic fluctuations on path utility in advance. In scenarios with sparse data or cold start, the system can still provide logically reasonable recommendation schemes through ontology constraints and association transfer in the knowledge graph, demonstrating excellent generalization performance.

[0035] Fifth, the system's high-performance distributed architecture and hierarchical parallel processing mechanism ensure real-time response to massive concurrent requests in intelligent transportation applications. The refined hub-based modeling and multi-criteria interaction mechanism greatly improve the efficiency of route recommendation in real-world travel scenarios, laying a solid technical foundation for achieving truly personalized, intelligent, and integrated multimodal travel services.

[0036] In the specific operation of the system, when a passenger initiates a travel request, the system first captures the spatiotemporal coordinates and semantic context of the request in the multi-source heterogeneous data acquisition and preprocessing layer. Subsequently, this information is converted into feature vectors and mapped to the current travel knowledge graph in the dynamic spatiotemporal semantic perception layer. The perception layer extracts local subgraphs affected by the current environment in real time and encodes them into high-dimensional embedding features.

[0037] Based on the embedded features output by the perception layer, the deep recommendation decision layer first generates a series of alternative routes in the multimodal transportation network using a knowledge-guided search algorithm. During the generation process, these routes have already eliminated infeasible options due to traffic control, malfunctions, or weather conditions using the logical rules of the knowledge graph. Subsequently, the personalized ranking module calculates multi-dimensional utility scores for these alternative routes, combining passengers' long-term preference features and short-term intent vectors.

[0038] The top-rated paths and their corresponding reasoning chains will be sent to the interactive path presentation and feedback layer. Travelers can view path details, estimated arrival time, estimated cost, and reasons for recommendation on their terminals. If a traveler selects and executes a path, the system will track their actual travel trajectory throughout. If a path deviation occurs, the feedback layer will promptly capture the deviation point and its surrounding environmental parameters, analyze the cause of the deviation (e.g., unforeseen road construction not being reported in time), and feed this information back as new knowledge tuples to the knowledge graph construction and management layer, triggering near real-time updates to the knowledge graph.

[0039] As a further refinement of this invention, the knowledge graph construction and management layer possesses the ability to automatically resolve conflicts. When discrepancies arise in the same attribute from different data sources, such as inconsistent congestion assessments of the same road segment by different navigation APIs, the system determines the final attribute value based on the data source's reliability level, update timestamp, and cross-validation results of surrounding nodes using a weighted voting mechanism. This mechanism ensures the authority and accuracy of the knowledge graph as the system's "single source of fact."

[0040] The system's personalized modeling also considers the heterogeneity of passenger groups. The system defines multiple "passenger profile" nodes in the knowledge graph, each aggregating a group of users with similar behavioral characteristics. For new users with extremely sparse behavioral data, the system maps their static attributes to the nearest profile node, thereby inheriting the group's prior knowledge and achieving high-quality recommendations in a cold-start state.

[0041] For refined processing of multimodal switching, the system has established a transfer cost model specifically for transition scenarios such as "walking-bus," "bus-rail," and "rail-ride-hailing." This model no longer simply treats transfers as a single point, but calculates a highly reliable transfer time distribution based on the geometric topology of the hub within the knowledge graph, real-time congestion levels, and vertical floor distribution. This enables the recommended routes to achieve minute-level accuracy in time prediction.

[0042] Furthermore, the system described in this invention supports a cloud-edge collaborative deployment architecture. For perception and search tasks requiring low latency, execution can be offloaded to edge computing nodes; while large-scale knowledge graph updates and complex deep learning model training are completed in high-performance cloud computing clusters. This architecture further optimizes the system's resource consumption and response speed, enabling it to operate smoothly in a city-level, ultra-large-scale transportation network environment.

[0043] In summary, this invention injects a logical core into a multimodal travel recommendation system by leveraging the powerful semantic representation tool of knowledge graphs. Combined with the perceptual depth of machine learning, it constructs a smart travel decision-making brain with self-learning, self-evolution, and interpretability, which has significant engineering and theoretical value for promoting technological progress in the field of smart transportation. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall physical architecture of the system of the present invention;

[0045] Figure 2 A schematic diagram of the internal logical structure of the multi-source heterogeneous data acquisition and preprocessing layer;

[0046] Figure 3A diagram showing the composition and module connections of the management layer for the travel knowledge graph construction;

[0047] Figure 4 A feature encoding and constraint logic diagram for the dynamic spatiotemporal semantic perception layer;

[0048] Figure 5 A schematic diagram of the module composition of the knowledge-guided deep recommendation decision layer;

[0049] Figure 6 This is a schematic diagram of the interactive path presentation and feedback layer and its overall operation process. Detailed Implementation

[0050] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0051] like Figure 1 As shown, the passenger multimodal travel route recommendation system based on knowledge graphs and machine learning is physically built on a high-performance distributed cloud computing cluster. This cluster consists of multiple computing nodes, storage nodes, and network switching equipment, ensuring high concurrency and fault tolerance when processing large-scale traffic data. The system's core computing resources employ a combination of multi-core processors and large-capacity memory to support complex graph operations and large-scale parallel training of deep learning models. In the underlying storage architecture, the system adopts a combination of a distributed file system and a distributed database to handle unstructured raw data and structured business data, respectively.

[0052] Specifically, the knowledge graph and machine learning-based multimodal travel route recommendation system comprises five core layers: a multi-source heterogeneous data acquisition and preprocessing layer, a travel knowledge graph construction and management layer, a dynamic spatiotemporal semantic perception layer, a knowledge-guided deep recommendation decision-making layer, and an interactive route presentation and feedback layer. These layers are interconnected via 10 Gigabit Ethernet and a high-speed internal bus, and use standardized message middleware for asynchronous communication, thereby ensuring efficient data flow and decoupling between system modules.

[0053] like Figure 2As shown, in the specific engineering implementation of the multi-source heterogeneous data acquisition and preprocessing layer, the system accesses multi-dimensional traffic-related data in real time through a series of preset standard data interfaces. Static traffic facility data, serving as the basic map of the road network, includes the geographical coordinates of rail transit stations (using the WGS-84 coordinate system with an accuracy of six decimal places), the physical routes of bus lines, the connectivity of road topology, and their corresponding physical attributes. The road topology data details the length, number of lanes, gradient, and pavement type of each road segment. For facilities such as parking lots and charging stations, the system collects their rated capacity, current available slots, and charging power specifications in real time. Dynamic operational status data relies on the city-level transportation IoT, acquired through induction coils, ultrasonic detectors, and vehicle-mounted GPS terminals deployed on key road sections. Real-time pedestrian flow data is obtained through infrared counters or Wi-Fi sniffing devices within transportation hubs, with a sampling period set to once per minute to ensure sensitive capture of passenger flow fluctuations. Vehicle punctuality rate and average speed are pushed in real time by the operator's backend, combined with historical periodic data to form a dynamic congestion index.

[0054] For individual passenger attribute data, the system collects anonymized data on the user's end to ensure compliance with privacy protection protocols. This data includes the frequency of the passenger's historical travel routes, tolerance scores for specific modes of transportation such as walking or subway (using a five-point rating system), the user's current precise geographic coordinates, and pre-selected route preferences in the system settings, such as "minimize transfers" or "avoid long walks." Geographic environmental information, as an external constraint affecting travel, includes the operating attributes, opening hours, and ratings of millions of Points of Interest (POIs) in the city. Meteorological parameters are accessed down to the hour, covering precipitation, visibility, and wind speed levels. Open text data from the internet mainly comes from government microblogs, official traffic radio accounts, and real-time user comments from mainstream navigation software. The system aggregates this fragmented information in real time to the data layer through high-frequency web crawling or subscribing to specific webhooks.

[0055] The multi-source heterogeneous data acquisition and preprocessing layer integrates a structured processing unit. The core task of this unit is to transform unstructured text descriptions into machine-understandable knowledge tuples. In terms of specific engineering steps, the system uses a pre-trained deep semantic model to process the collected emergency traffic announcements. During entity extraction, the model annotates the text sequence word by word or character by character, accurately identifying the affected locations, time periods, event types, and severity levels. Subsequently, in the relation extraction stage, the system utilizes dependency parsing technology to clarify the association between the event and the affected traffic routes or physical coordinates, encapsulating it into a "event-impact-entity" triple structure conforming to the ontology definition.

[0056] like Figure 3 As shown, during the implementation of the travel knowledge graph construction and management layer, the system predefines a complete ontology model based on expert knowledge in the transportation field. This model abstracts the physical world into multiple interconnected entity classes. The "Passenger" class includes attributes such as age group and preference tags; the "Transportation" class is subdivided into subclasses such as subway, bus, taxi, and shared bicycle, and assigned physical constraints such as rated passenger capacity and average speed; the "Route Segment" class describes the connection characteristics between two physical nodes, including the distribution of travel time and real-time cost standards. The knowledge fusion unit plays a crucial role at this level, solving the identity recognition problem under multi-source data through complex entity alignment logic. Specifically, when two geographical entities from different data sources have similar names but slightly offset coordinates, the system calculates the Euclidean spatial distance between them. If the distance is lower than a set threshold, such as 30 meters, the system further compares the character edit distance and the cosine similarity of the semantic description vectors. Only when the comprehensive score exceeds the preset confidence threshold does the system determine that the two entities are the same physical object and perform attribute merging and redundancy elimination operations.

[0057] To efficiently store and retrieve this complex relational data, the travel knowledge graph management layer employs a distributed graph database based on an attribute graph structure. In this database, nodes represent entities, edges represent relationships between entities, and attributes are stored on nodes or edges. The graph database supports the storage of billions of nodes and can perform real-time graph traversals with depths of six hops or more. This structure enables the system to quickly calculate the path connectivity between any two points. As a crucial component of this layer, the global consistency verification unit is responsible for maintaining the logical integrity of the knowledge graph. Given the high frequency of data updates, this unit automatically detects temporal inconsistencies or spatial overlaps by executing a series of pre-defined ontology logic rules. For example, if the system detects that a passenger appears at two different stations ten kilometers apart within the same minute, the consistency verification unit will trigger an alert and automatically correct the data based on the credibility weight of the data source, ensuring the authority of the knowledge graph as the system's "single source of truth."

[0058] like Figure 4As shown, the task of the dynamic spatiotemporal semantic perception layer is to transform high-dimensional and sparse knowledge graph data into a vector form that can be efficiently processed by deep learning models. The feature encoding module utilizes graph convolution techniques to fuse the features of a specific node with the features of its neighboring nodes in each iteration. In terms of specific parameter settings, different types of edges are assigned different propagation weights; for example, the weight of a strong connection like "belonging to a certain route" is usually higher than that of a weak connection like "close to a point of interest." This convolution process can effectively capture the spatial local characteristics of the traffic network. Simultaneously, the perception layer combines a long short-term memory network to process the time-dimensional data, encoding the traffic flow fluctuation trend over the past three hours into the dynamic feature vector. The final generated node embedding vector typically has high dimensionality, such as 128 or 256 dimensions, where each dimension represents some abstract spatiotemporal or semantic feature.

[0059] To achieve a keen capture of the real-time environment, the dynamic spatiotemporal semantic perception layer is also equipped with a context constraint module. This module dynamically adjusts the activation threshold of the connection edges in the knowledge graph based on current external environmental variables, such as heavy rainfall, peak congestion, or late-night periods. For example, during heavy rain, this module automatically increases the weight of "travel cost" for ground public transportation routes and decreases the weight of "environmental resistance" for indoor transfer routes, thereby giving the embedding vector output by the perception layer a strong environmental bias.

[0060] like Figure 5 As shown, in the knowledge-guided deep recommendation decision layer, the system employs a multi-task neural network framework to implement complex decision-making logic. The path candidate set generation module, as the first step, uses an improved graph search algorithm to filter potential solutions from a massive number of possible combinations. This search process not only relies on the shortest physical path but also introduces a "semantic cost factor" calculated from the knowledge graph. This factor is a comprehensive variable that reflects in real time the probability of a path segment being affected by unforeseen events and the current weather risk. The personalized ranking module then performs a refined evaluation of these candidate paths. This module integrates a deep interest evolution network, which can analyze the similarity between a passenger's current request and historical behavioral trajectories through an attention mechanism. When calculating the score, the system comprehensively considers the total time spent on the path, economic expenditure, transfer frequency, and environmental comfort.

[0061] The explanatory reasoning module is one of the system's technological highlights. While generating recommended routes, it traces the critical paths leading to the decision back from the knowledge graph. When the system recommends a detour bypassing the main road, this module extracts logical chains from the knowledge graph, such as "main road - rear-end collision," "the accident - resulted in a 70% reduction in traffic capacity," and "alternative route - has a high on-time rate." These logical chains are then transformed into templated natural language explanations and displayed on the terminal interface in easily understandable text, greatly improving the transparency of the decision-making process.

[0062] like Figure 6 As shown, the interactive path presentation and feedback layer directly faces the end user. It not only displays multimodal paths graphically, such as color-coding different modes of transportation on an electronic map, but more importantly, it constructs a closed-loop learning feedback system. The real-time feedback loop continuously monitors passenger behavior. If a passenger chooses the second option instead of the top-ranked one, or deviates from the recommended route during their journey, the feedback loop immediately captures this deviation and re-labels the geographical coordinates, environmental parameters, and corresponding time of the deviation point, sending this information back to the recommendation engine. The system utilizes incremental learning technology to fine-tune model parameters without interrupting service, ensuring the system continuously evolves with user interaction.

[0063] As a deeper level of optimization, the system's built-in hub micropath modeling subsystem performs ultra-high-precision triplet modeling for large integrated transportation hubs. This subsystem details the operating status of each elevator within the hub, the average queuing time at each ticket gate, and the physical distance between different floors for transfers. When the deep recommendation decision layer calculates the total time, the refined data provided by this subsystem can refine the originally vague "transfer time" into a precise time with statistical distribution patterns, thereby solving the "last mile" uncertainty problem commonly found in multimodal connections.

[0064] In addition, the system includes a sudden event simulation and evolution prediction unit. This unit uses a recurrent neural network to learn from tens of thousands of historical traffic emergencies, and can simulate the spatiotemporal diffusion pattern of the current event within the next one to four hours based on its initial intensity and location. The prediction results are input into the recommendation decision engine in the form of dynamic constraints, giving it proactive obstacle avoidance capabilities. For example, when a subway station is temporarily closed, the system can predict the peak passenger flow at nearby bus stations within seconds of the event and guide passengers to avoid potentially extremely congested areas in subsequent recommendations.

[0065] Regarding data security and privacy protection, the system performs strict hash de-identification operations at the data collection and preprocessing layers. All information involving ID card numbers, mobile phone numbers, or specific home addresses is converted into irreversible encrypted digests. In terms of physical storage, access to knowledge graph attributes is restricted by role-based access control policies, ensuring that sensitive preference data can only be used for model training within an encrypted sandbox and will not be leaked to third parties.

[0066] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A passenger multimodal travel route recommendation system based on knowledge graphs and machine learning, characterized in that, The system is deployed on a distributed cloud computing cluster in terms of physical architecture, including a multi-source heterogeneous data acquisition and preprocessing layer, a travel knowledge graph construction and management layer, a dynamic spatiotemporal semantic perception layer, a knowledge-guided deep recommendation decision layer, and an interactive path presentation and feedback layer. Each level achieves closed-loop flow of data and instructions through a high-speed internal bus and communication protocol; The multi-source heterogeneous data acquisition and preprocessing layer is configured to access and integrate static traffic facility data, dynamic operational status data, passenger individual attribute data, geographic environment information, and open text data from the Internet in real time through a standard interface. The travel knowledge graph construction and management layer is configured to build and maintain a high-dimensional travel knowledge network based on a preset traffic domain ontology model. The ontology model defines entity classes including passengers, vehicles, transportation stations, route segments, points of interest, emergencies, and environmental factors, as well as attribute constraint relationships between entity classes. The dynamic spatiotemporal semantic perception layer is configured to use embedding representation technology to map high-dimensional discrete nodes in the travel knowledge graph to a low-dimensional continuous dense vector space, generating node embedding vectors that include four dimensions: static attributes, spatial topology, temporal dynamics, and semantic environment. The knowledge-guided deep recommendation decision layer is configured to implement end-to-end path recommendation logic using a multi-task neural network framework, and integrates a path candidate set generation module, a personalized ranking module, and an interpretive reasoning module. The interactive path presentation and feedback layer is configured to display recommended solutions on the user terminal and has a real-time feedback loop to feed back passenger preference corrections and abnormal offset data to the knowledge-guided deep recommendation decision layer.

2. The passenger multimodal travel route recommendation system based on knowledge graphs and machine learning according to claim 1, characterized in that, The multi-source heterogeneous data acquisition and preprocessing layer is equipped with a structured processing unit, which uses natural language processing technology to extract entities and relationships from the open text data on the Internet. The entity extraction process uses a pre-trained language model to perform sequence labeling on the text, identifying four core elements: location, time, event type, and degree of impact. The relation extraction process is based on dependency parsing, establishing semantic associations between events and geographic coordinates and transportation routes, transforming unstructured text into structured triples that conform to the ontology definition, and providing standardized input for the dynamic updating of the travel knowledge graph.

3. The passenger multimodal travel route recommendation system based on knowledge graphs and machine learning according to claim 1, characterized in that, The travel knowledge graph construction and management layer includes a knowledge fusion unit, which is used to align attributes from different sources that point to the same physical entity. The knowledge fusion unit determines the uniqueness of entities by calculating the spatial distance similarity, name character edit distance, and semantic vector cosine similarity, eliminating redundancy and filling in missing attributes; The travel knowledge graph construction and management layer uses a distributed graph database for storage, uses an attribute graph structure to represent the complex topological relationships between nodes, and calculates the path connectivity between any two points and their reachability strength in a specific semantic context based on graph traversal reasoning.

4. The passenger multimodal travel route recommendation system based on knowledge graphs and machine learning according to claim 1, characterized in that, The dynamic spatiotemporal semantic perception layer includes a feature encoding module. The feature encoding module extracts structured features from the knowledge graph through a graph convolutional network. In each convolutional iteration, a specific node aggregates the attribute information of its first-order and multi-order neighbor nodes and assigns different weight coefficients according to the type of the connecting edge. The feature encoding module combines long short-term memory network to process time series data, encoding the periodic fluctuation pattern of traffic flow and the attenuation process of the impact of sudden events into the dynamic feature vector of the node; the dynamic spatiotemporal semantic perception layer also has a context constraint module, which dynamically adjusts the activation threshold of different edge weights in the knowledge graph according to the current time, weather conditions and regional congestion level.

5. The passenger multimodal travel route recommendation system based on knowledge graphs and machine learning according to claim 1, characterized in that, The path candidate set generation module retrieves a set of candidate paths that satisfy basic spatiotemporal constraints in a multimodal transportation network based on an improved graph search strategy. The improved graph search strategy automatically avoids areas affected by sudden events or harsh environments by introducing semantic cost factors from the knowledge graph. The semantic cost factor is calculated in real time based on the association strength between the "emergency event" entity and the "path segment" entity in the knowledge graph.

6. The passenger multimodal travel route recommendation system based on knowledge graphs and machine learning according to claim 1, characterized in that, The personalized ranking module uses a deep interest evolution network to capture the correlation between passengers' current travel intentions and historical behavior patterns through an attention mechanism. During the calculation process, the system performs fully connected layer operations on the feature vectors of candidate paths with the passengers' historical trajectory features and real-time preference vectors. The loss function of the personalized ranking module combines four dimensions: path time, economic cost, transfer convenience, and environmental comfort. Through training and optimization, the recommendation results are made closer to the actual choices of passengers.

7. The passenger multimodal travel route recommendation system based on knowledge graphs and machine learning according to claim 1, characterized in that, The explanatory reasoning module is configured to extract key reasoning chains from the travel knowledge graph while generating recommended routes. The reasoning chains demonstrate the logical relationship between recommended routes and passenger preferences, real-time events, and point-of-interest attributes. The interpretive reasoning module traces and extracts the correlation logic of the constraints of events on traffic nodes in the knowledge graph, and transforms it into a semantic description that the user can understand, so as to improve the interpretability of recommendation decisions.

8. The passenger multimodal travel route recommendation system based on knowledge graphs and machine learning according to claim 1, characterized in that, The system also includes a hub-within-a-hub micropath modeling subsystem, which establishes station internal triples in the travel knowledge graph to describe the walking distance, floor span, vertical elevator location and average transfer time required for different line transfers. When calculating the total path cost, the knowledge-guided deep recommendation decision layer calls upon the refined data from the micro-path modeling subsystem within the hub to address the path description accuracy issue at multimodal connection points.

9. The passenger multimodal travel route recommendation system based on knowledge graph and machine learning according to claim 1, characterized in that, The system also includes a sudden event simulation and evolution prediction unit. The sudden event simulation and evolution prediction unit uses a recurrent neural network to learn the evolution patterns of historical sudden events recorded in the knowledge graph. When a new sudden event is detected in its early stages, it predicts the diffusion impact on surrounding traffic flow and node connectivity within a preset time period in the future. The prediction results are then input as dynamic constraints to the knowledge-guided deep recommendation decision layer in real time.

10. The passenger multimodal travel route recommendation system based on knowledge graph and machine learning according to claim 1, characterized in that, The travel knowledge graph construction and management layer includes a global consistency verification unit, which is used to detect and correct logical contradictions in the knowledge graph by executing preset ontology logic rules when multi-source data is updated frequently. The multi-source heterogeneous data acquisition and preprocessing layer performs a feature normalization process. For continuous physical values, it uses range standardization or standard deviation standardization methods, and for discrete semantic features, it uses one-hot encoding or word embedding methods to convert them into fixed-length digital vectors. The interactive path presentation and feedback layer supports dynamic adjustment of multiple criteria. By adjusting the weight slider on the terminal interface, the output fusion ratio of the multi-task head in the deep recommendation decision layer can be finely adjusted in real time to achieve instant updates of the recommended path.

Images