A Cultural Tourism Map Search System and Method Based on Deep Learning and BeiDou Database

Abstract

This invention discloses a cultural tourism map search system and method based on deep learning and the BeiDou database, comprising the following steps: acquiring user location and trajectory data, loading geographic objects, and constructing a heterogeneous graph; receiving natural language requests, extracting semantic vectors and writing them into the graph; constructing a graph contrastive learning network, generating views and performing embedding learning; optimizing structural parameters using a pigeon flocking algorithm and outputting the optimal model; constructing the current location node, extracting local subgraphs, and generating a sorted list; performing path reasoning, generating recommendation results, and providing feedback updates. This invention, by integrating deep semantic understanding and geographic trajectory modeling, achieves high-precision, interpretable, and self-learning intelligent search and recommendation of cultural tourism maps driven by the user's current location.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a cultural tourism map search system and method based on deep learning and the BeiDou database. Background Technology

[0002] With the development of geographic information systems, big data analytics, and intelligent recommendation technologies, location-based map search and recommendation services have been widely applied in transportation, cultural tourism, and navigation. In the context of cultural and tourism integration, tourists have placed higher demands on personalized and intelligent destination search experiences. They not only expect systems to understand their complex query intentions expressed in natural language, but also expect search results to be location-relevant, interest-matching, and transparent in interpretation, thereby improving overall travel efficiency and user experience satisfaction.

[0003] Current mainstream map search recommendation methods primarily rely on keyword matching, rule-based retrieval mechanisms, or center point recommendations based on points of interest. While these methods offer fast response times for simple route navigation, they have significant limitations in handling queries with complex semantic instructions, considering users' historical behavioral paths, or incorporating real-time location and scene modeling. On one hand, existing systems are mostly based on vector space models, relying solely on keywords or tags for text query matching, lacking deep semantic understanding of natural language search requests and failing to identify implicit location preferences, scene preferences, or behavioral needs within user intent. On the other hand, many recommendation algorithms lack effective modeling of users' historical behavior, geographical movement trajectories, and access sequences, resulting in recommendations that lack specificity and contextual adaptability, and are prone to semantic matching errors, redundant location recommendations, and unreasonable route planning.

[0004] Furthermore, existing map recommendation systems generally employ flat graph modeling methods, which cannot express the complex relationships between various types of nodes and multi-dimensional edges in cultural and tourism data. In the cultural and tourism field, user behavior, geographic objects, and path connections often exhibit heterogeneity and semantic overlap. Without constructing a graph structure that can express the fusion of semantic, spatial, and behavioral information, it will be difficult to achieve accurate and efficient multi-dimensional intelligent recommendations. In addition, although graph neural networks have been gradually introduced into recommendation systems, most of them use fixed structures and single parameter configurations, lacking structure optimization mechanisms for specific task objectives, making them prone to overfitting, embedding failure, or poor model robustness. Summary of the Invention

[0005] One objective of this invention is to propose a cultural tourism map search system and method based on deep learning and the BeiDou database. This invention integrates heterogeneous graph modeling, graph contrastive learning, and pigeon flocking optimization algorithms. By constructing a heterogeneous graph structure containing users, geographic objects, and path nodes, it captures users' historical trajectories and geographic relationships. Using an optimized graph contrastive learning network, it performs semantic embedding calculations and relevance ranking to achieve intelligent map search and recommendation driven by the user's current location. Furthermore, it achieves adaptive updates of the graph structure and model through user feedback behavior data, possessing advantages such as deep semantic understanding, strong location association, high recommendation accuracy, and strong system interpretability.

[0006] The cultural tourism map search method based on deep learning and BeiDou database according to an embodiment of the present invention includes the following steps:

[0007] The system retrieves the user's current location data and historical movement trajectory data from the BeiDou database, loads a static information set containing multiple cultural and tourism geographic objects, and constructs a heterogeneous graph structure.

[0008] Receive natural language search requests input by users, perform context semantic vector extraction processing on the natural language search requests, obtain semantic representation results, and write them into the heterogeneous graph structure as the initial features of user nodes;

[0009] A graph contrast learning network is constructed based on a heterogeneous graph structure. The graph structure perturbation is performed to generate two sets of views. The two sets of views are then input into the graph neural network for node embedding learning.

[0010] The structural parameters of the graph contrast learning network are optimized by using the pigeon flock optimization algorithm, and the optimal structure and parameter configuration of the graph contrast learning network are finally output, forming the optimized graph contrast learning network.

[0011] When a user initiates a map search request, the system extracts the user's current location coordinates from the BeiDou database, constructs graph nodes based on the user's current location coordinates and adds them to the heterogeneous graph structure, and extracts local subgraphs from the heterogeneous graph structure as the central node. The system then calls an optimized graph contrast learning network to perform node embedding representation calculations on the local subgraphs and generates a semantic relevance ranking list.

[0012] Based on the semantic relevance ranking list and the connection relationships of path nodes in the heterogeneous graph structure, multi-hop path reasoning is performed to generate recommendation results that include the optimal path sequence from the user's current location to the target geographic object and recommendation explanation information.

[0013] Optionally, the BeiDou database is a navigation data set that records the user's real-time location and movement path, including latitude and longitude coordinates, timestamps, speed and direction information, used to provide the user's spatial location and temporal trajectory information on the map, and supports graph node construction, subgraph extraction and path recommendation calculation;

[0014] The cultural and tourism geographic objects include cultural and tourism-related entities with geographical locations, tag classifications, and rating records, which are used to construct geographic entity nodes in the graph and support user interest matching, semantic association calculation, and path recommendation generation.

[0015] The heterogeneous graph structure includes user nodes, geographic object nodes, and path nodes. Access edges, spatial adjacency edges, and trajectory sequence edges are established between nodes to represent user behavior, geographic location relationships, and travel trajectories, supporting graph comparison learning and personalized recommendation calculation.

[0016] Optionally, the initial features of user nodes are written into the heterogeneous graph structure, including:

[0017] Receive natural language search requests input by users through terminals, the natural language search requests including geographical preferences, behavioral intentions, travel purposes, and topic keywords;

[0018] Perform text preprocessing on natural language search requests, including word segmentation, stop word removal, lexical unification, and invalid symbol removal, to generate standardized sentence input;

[0019] The preprocessed natural language input text is subjected to context semantic vector extraction processing to generate a context-dependent semantic vector representation;

[0020] The extracted semantic vector representations are normalized in dimension and aligned in feature to meet the requirements of dimensional consistency and structural constraints of node feature representations in heterogeneous graph structures.

[0021] By binding semantic vector representations with user identity identifiers, a complete user node feature representation data structure is generated.

[0022] The generated user node features are written into the corresponding user node positions in the heterogeneous graph structure, completing the semantic embedding mapping process from natural language search requests to graph structures.

[0023] Optionally, the graph contrast learning network constructed based on heterogeneous graph structures includes:

[0024] Constructing the original graph based on heterogeneous graph structure ,in, This represents a set of nodes, including user nodes, geographic object nodes, and path nodes. This represents a set of edges, including visit edges, spatially adjacent edges, and trajectory order edges.

[0025] Introducing a weighted coefficient based on geographic location similarity Construct a location-aware similarity function This is used to fuse the spatial relationships between nodes in the contrastive loss;

[0026] Introducing a subgraph perception scale control factor Used to determine the adjacency depth of the local subgraph extracted from the central node, in the first view Second view In the construction, the number of jumps is respectively used as The neighborhood is used as the subgraph extraction range, and node embedding extraction is performed at the subgraph level;

[0027] Introducing a node occlusion ratio adjustment factor This is used to control the occlusion ratio of node features during view perturbation, in the construction of the first view. Second view At that time, according to the ratio Masking the node feature dimensions in the original feature matrix;

[0028] First view Second view Input a graph neural network encoder to obtain the embedding representation of each node in two sets of views. , ;

[0029] Constructing a position-aware contrastive loss function This is used to optimize the consistency of the embedded representation of the same node in two views;

[0030] The graph neural network is trained based on the location-aware contrastive loss function, and finally a graph contrastive learning network representation structure that integrates geographic awareness, adjustable granularity and dynamic perturbation strategy is obtained.

[0031] Optionally, forming the optimized graph contrast learning network includes:

[0032] The initial population size for the pigeon flock optimization algorithm is set to... Initialize the set of individual pigeons in the flock. Each individual This represents the combination of structural parameters of a graph-based contrast learning network, where... , This represents the weighted coefficient of geographical location similarity corresponding to the i-th individual. Let be the subgraph perception scale control factor corresponding to the i-th individual. is the node occlusion ratio adjustment factor corresponding to the i-th individual;

[0033] Randomly initialize velocity vector for each individual And set individual memory vectors This is used to record an individual's historical optimal parameter configuration;

[0034] For each individual parameter combination Based on the constructed graph contrastive learning network structure, node embedding computation and contrastive learning training processes are performed to obtain the corresponding contrastive loss function values. Recommendation accuracy with node ranking Weighted construction of fitness function ;

[0035] All individuals were sorted according to their fitness values ​​from highest to lowest, and then divided into groups using an equal-division strategy. Within each subgroup, the individual with the highest fitness value is selected as the local leader. The remaining individuals will perform localized guidance and updates around the local navigator;

[0036] Calculate the current iteration round Dynamic inertia factor ;

[0037] Based on dynamic inertia factor Calculate candidate geomagnetic navigation positions for each individual. Landmark-guided location candidate values :

[0038] Geomagnetic navigation update:

[0039] ;

[0040] Landmark guidance update:

[0041] ;

[0042] in, Let be the position vector of the i-th individual in the t-th iteration. Let i be the velocity vector of individual i in round t. To guide the weighting coefficients, This refers to the individual memory weighting coefficient. This represents the position of the individual with the best fitness in the current population. This represents the historical best memory position of the i-th individual;

[0043] Introducing stage control factors The two candidate positions are dynamically balanced through a fusion mechanism, the position is updated, and the individual velocity is updated in the next round based on the position update result of this round.

[0044] Calculate the new round of individual locations The contrastive loss of the corresponding graph contrastive learning network And recommendation accuracy Update the fitness function ;

[0045] If the current individual fitness Then the memory vector will be updated to the current solution. ;

[0046] Repeat the process until the maximum number of iterations is met. Alternatively, a fitness convergence threshold condition may be used to ultimately output the individual with the best fitness. , which serves as the optimal combination of structural parameters for graph contrastive learning networks.

[0047] Optionally, generating the semantic relevance ranking list includes:

[0048] Receive map search requests initiated by users on the terminal and trigger the location-aware recommendation process initialization operation;

[0049] Extract the latitude and longitude coordinates of the user's current location from the BeiDou database, and use the latitude and longitude coordinates of the user's current location as location nodes to construct graph node entities;

[0050] The location node is added as the user's current location node to the currently maintained heterogeneous graph structure, thereby expanding the spatial representation capability of the graph structure;

[0051] Using the user's current location node as the central node, the associated local subgraph structure is extracted from the heterogeneous graph structure based on spatial adjacency and access relationships;

[0052] The optimized graph contrast learning network is invoked to perform embedding representation calculation on the extracted local subgraphs and generate a semantic representation for each candidate node.

[0053] Based on the semantic relevance between the embedded representations of the user's current location node and candidate nodes, a ranked list is generated and output as a set of candidate results for search recommendations.

[0054] Optionally, the user feedback behavior data specifically includes click records, browsing dwell time, historical search paths, and actual navigation behavior, which are used to optimize the retraining process of heterogeneous graph structures and graph contrastive learning networks.

[0055] The cultural tourism map search system based on deep learning and the BeiDou database includes the following modules:

[0056] The request processing module is used to receive user map search requests and extract semantic representations as user node features;

[0057] The mapping module is used to extract user location and trajectory data from the BeiDou database and combine it with geographic object information to construct a heterogeneous map structure.

[0058] The graph learning optimization module is used to generate perturbation views based on heterogeneous graphs, train the graph contrast learning network, and adjust the structural parameters through the pigeon flocking optimization algorithm.

[0059] The subgraph calculation module is used to extract local subgraphs and generate node embedding representations;

[0060] The recommendation module generates ranking results based on the semantic relevance of the embedded vectors, and outputs recommended paths and explanatory information;

[0061] The feedback update module is used to collect user feedback behavior data and use it for graph structure updates and model retraining.

[0062] The beneficial effects of this invention are:

[0063] This invention integrates the BeiDou database, graph comparison learning network, and pigeon flocking optimization algorithm to address the problems of weak semantic understanding, insufficient graph structure representation, low recommendation accuracy, and lack of adaptive model updates in existing map search and recommendation technologies. It proposes an intelligent search method with an overall closed-loop structure and multi-source information fusion capability, which significantly improves the system's response quality and recommendation accuracy in complex user query contexts.

[0064] This invention constructs a heterogeneous graph structure comprising user nodes, geographic object nodes, and path nodes, integrating real-time user location and historical trajectory data provided by the BeiDou database. This effectively preserves spatial adjacency relationships, access time sequence relationships, and path evolution relationships, enhancing the system's ability to model geographic information and behavioral data. By introducing a graph contrastive learning mechanism, a perturbation view is constructed based on the heterogeneous graph structure, and node embedding alignment is performed. This effectively enhances the robustness of semantic representation and the expressive power of node similarity, overcoming the shortcomings of traditional graph learning models that easily lose structural semantics when fusing multiple types of data.

[0065] To further enhance the adaptability and optimization performance of graph neural network structures, this invention introduces a pigeon flock optimization algorithm. Geographic perception weights, subgraph scale control factors, and feature perturbation ratios are set as optimization variables. Individual parameter vectors are constructed, and the graph contrastive learning network is structurally tuned using a swarm search approach. The algorithm employs a dynamic inertia factor, an individual memory mechanism, and a stage fusion guidance strategy to achieve a smooth transition between global exploration and local convergence, significantly improving the adaptability of structural parameter configuration and the discriminative ability of graph embeddings.

[0066] This invention also constructs graph nodes by extracting the user's current location in real time, extracts its spatial adjacency subgraph as the central node and performs embedding calculations, and completes recommendation ranking by combining semantic relevance and path reachability, outputting search results with geographical significance and semantic consistency. After the recommendation results are generated, the system collects the user's click records, dwell behavior, and actual navigation path as feedback data, which is applied to the dynamic incremental update of the heterogeneous graph structure and the retraining of the graph comparison learning model, thus constructing a recommendation system with continuous learning and dynamic adaptability. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is a flowchart of the cultural tourism map search method based on deep learning and BeiDou database proposed in this invention;

[0069] Figure 2 This is a schematic diagram of the structure of the cultural tourism map search system based on deep learning and Beidou database proposed in this invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0071] refer to Figures 1-2 The cultural tourism map search method based on deep learning and BeiDou database includes the following steps:

[0072] The system retrieves the user's current location data and historical movement trajectory data from the BeiDou database, loads a static information set containing multiple cultural and tourism geographic objects, and constructs a heterogeneous graph structure.

[0073] Receive natural language search requests input by users, perform context semantic vector extraction processing on the natural language search requests, obtain semantic representation results, and write them into the heterogeneous graph structure as the initial features of user nodes;

[0074] A graph contrast learning network is constructed based on a heterogeneous graph structure. The graph structure perturbation is performed to generate two sets of views. The two sets of views are then input into the graph neural network for node embedding learning.

[0075] The structural parameters of the graph contrast learning network are optimized by using the pigeon flock optimization algorithm, and the optimal structure and parameter configuration of the graph contrast learning network are finally output, forming the optimized graph contrast learning network.

[0076] When a user initiates a map search request, the system extracts the user's current location coordinates from the BeiDou database, constructs graph nodes based on the user's current location coordinates and adds them to the heterogeneous graph structure, and extracts local subgraphs from the heterogeneous graph structure as the central node. The system then calls an optimized graph contrast learning network to perform node embedding representation calculations on the local subgraphs and generates a semantic relevance ranking list.

[0077] Based on the semantic relevance ranking list and the connection relationships of path nodes in the heterogeneous graph structure, multi-hop path reasoning is performed to generate recommendation results containing the optimal path sequence from the user's current location to the target geographic object and recommendation explanation information. The recommendation results are returned to the user terminal, and user feedback behavior data is collected for incremental updates of the heterogeneous graph structure and retraining of the graph contrastive learning network.

[0078] This invention constructs a multi-source heterogeneous graph structure by integrating BeiDou database positioning data, user behavior trajectories, and natural language semantic features, effectively improving the matching ability between user intent and geographical context during map search. Utilizing graph contrastive learning networks and graph structure perturbation mechanisms, the system can extract highly robust node embedding representations from the graph's semantic structure, enhancing the adaptability of recommendation results to query semantics. Simultaneously, a pigeon flock optimization algorithm is introduced to dynamically fine-tune graph structure parameters, achieving synergistic optimization between the graph neural network structure and task objectives, significantly improving the model's generalization ability and accuracy across different user query scenarios. Furthermore, this invention extracts the user's current location to construct a central node, extracts local subgraphs, and performs high-quality embedding computation to generate a candidate list with higher ranking accuracy and stronger spatial relevance. Combined with path connection information, multi-hop inference is performed to generate interpretable navigation recommendation results. By using feedback behavioral data to drive graph structure and model retraining, the system possesses continuous evolution and personalized recommendation capabilities. Overall, this invention achieves comprehensive improvements in semantic understanding, structural modeling, recommendation interpretability, and system adaptability, significantly outperforming traditional keyword retrieval methods.

[0079] In this embodiment, the BeiDou database is a navigation data set that records the user's real-time location and movement path, including latitude and longitude coordinates, timestamps, speed and direction information. It is used to provide the user's spatial location and temporal trajectory information on the map, and supports graph node construction, subgraph extraction and path recommendation calculation. The cultural and tourism geographic objects include cultural and tourism-related entities with geographic location, tag classification, and rating records, which are used to construct geographic entity nodes in the graph, and support user interest matching, semantic association calculation and path recommendation generation. The heterogeneous graph structure includes user nodes, geographic object nodes and path nodes. Access edges, spatial adjacency edges and trajectory order edges are established between nodes to represent user behavior, geographic location relationships and travel trajectories, and support graph comparison learning and personalized recommendation calculation.

[0080] In this embodiment, writing the initial features of the user node into the heterogeneous graph structure includes:

[0081] Receive natural language search requests input by users through terminals, the natural language search requests including geographical preferences, behavioral intentions, travel purposes, and topic keywords;

[0082] Perform text preprocessing on natural language search requests, including word segmentation, stop word removal, lexical unification, and invalid symbol removal, to generate standardized sentence input;

[0083] The preprocessed natural language input text is subjected to context semantic vector extraction processing to generate a context-dependent semantic vector representation;

[0084] The extracted semantic vector representations are normalized in dimension and aligned in feature to meet the requirements of dimensional consistency and structural constraints of node feature representations in heterogeneous graph structures.

[0085] By binding semantic vector representations with user identity identifiers, a complete user node feature representation data structure is generated.

[0086] The generated user node features are written into the corresponding user node positions in the heterogeneous graph structure, completing the semantic embedding mapping process from natural language search requests to graph structures.

[0087] This invention, by constructing a standardized natural language processing workflow, achieves precise mapping of user search requests to semantic embeddings in heterogeneous graph structures, significantly improving the map search system's ability to understand the user's true intent. Natural language requests often contain complex expressions and diverse query content, such as geographical preferences, behavioral intentions, and travel purposes. This invention effectively extracts deep semantic features and regularizes them into structured embedding vectors through text preprocessing, contextual semantic vector extraction, feature alignment, and normalization steps. By binding semantic vectors to user identity identifiers and writing them into the heterogeneous graph structure, not only is the consistency between node feature dimensions and the graph network structure ensured, but personalized semantic queries and structural modeling are also integrated. This technical solution solves the problems of traditional map systems' difficulty in parsing fuzzy language and the inability of node features to directly carry natural language intent. It enhances the graph structure's ability to express and extend user semantics, laying a high-quality input foundation for subsequent graph neural network training and recommendation path generation, and has the beneficial effects of high accuracy, high generalization ability, and excellent structural compatibility.

[0088] In this embodiment, the graph comparison learning network constructed based on heterogeneous graph structure includes:

[0089] Constructing the original graph based on heterogeneous graph structure ,in, This represents a set of nodes, including user nodes, geographic object nodes, and path nodes. This represents a set of edges, including visit edges, spatially adjacent edges, and trajectory order edges.

[0090] Introducing a weighted coefficient based on geographic location similarity Construct a location-aware similarity function This is used to fuse spatial relationships between nodes in contrastive loss:

[0091] ;

[0092] in, Represents a node and Geographic similarity, The cosine similarity between embedded vectors is represented. For nodes The embedding vector representation, For nodes Embedded vector representation;

[0093] Introducing a subgraph perception scale control factor Used to determine the adjacency depth of the local subgraph extracted from the central node, in the first view Second view In the construction, the number of jumps is respectively used as The neighborhood is used as the subgraph extraction range, and node embedding extraction is performed at the subgraph level;

[0094] Introducing a node occlusion ratio adjustment factor This is used to control the occlusion ratio of node features during view perturbation, in the construction of the first view. Second view At that time, according to the ratio Masking the node feature dimensions in the original feature matrix enhances the structural robustness of the embedding.

[0095] First view Second view Input a graph neural network encoder to obtain the embedding representation of each node in two sets of views. , ;

[0096] Constructing a position-aware contrastive loss function This is used to optimize the consistency of the embedded representation of the same node in two views:

[0097] ;

[0098] in, To compare temperature coefficients, It is a logarithmic function. It is an exponential function;

[0099] The graph neural network is trained based on the location-aware contrastive loss function, and finally a graph contrastive learning network representation structure that integrates geographic awareness, adjustable granularity and dynamic perturbation strategy is obtained.

[0100] This invention improves graph contrastive learning networks by introducing three innovative structural parameters, significantly enhancing the model's robustness and representational capabilities in geographic perception, structural perturbation, and semantic representation. First, a geographic location similarity weighting coefficient is introduced to construct a location-aware similarity function, allowing spatial adjacency relationships between nodes to directly influence the contrastive loss term, effectively improving the model's ability to capture spatial associations between users and geographic objects. Second, a subgraph-aware scale control factor is introduced to dynamically adjust the neighborhood range extracted from the central node, supporting subgraph construction with different hop counts and achieving granular local structural modeling, enhancing adaptability to multi-level semantic relationships in the graph. Third, a node occlusion ratio adjustment factor is set to control the intensity of feature occlusion during graph-view perturbation, improving the model's stability in adversarial perturbation and unsupervised view enhancement. Through the combined effect of these structural parameters, the node embedding representations extracted by the system from two perturbed views exhibit stronger consistency and discriminativeness, and the location-aware contrastive loss function further optimizes the semantic alignment effect between different views. The final graph contrastive learning network possesses stronger geographic representation capabilities, semantic accuracy, and structural generalization performance, providing a high-quality embedding representation foundation for subsequent recommendation ranking and path reasoning.

[0101] In this embodiment, forming the optimized graph contrast learning network includes:

[0102] The initial population size for the pigeon flock optimization algorithm is set to... Initialize the set of individual pigeons in the flock. Each individual This represents the combination of structural parameters of a graph-based contrast learning network, where... , This represents the weighted coefficient of geographical location similarity corresponding to the i-th individual. Let be the subgraph perception scale control factor corresponding to the i-th individual. is the node occlusion ratio adjustment factor corresponding to the i-th individual;

[0103] Randomly initialize velocity vector for each individual And set individual memory vectors This is used to record an individual's historical optimal parameter configuration;

[0104] For each individual parameter combination Based on the constructed graph contrastive learning network structure, node embedding computation and contrastive learning training processes are performed to obtain the corresponding contrastive loss function values. Recommendation accuracy with node ranking Construct the fitness function :

[0105] ;

[0106] in, To compare the weighting coefficients of the loss in the fitness function, which are used to control the degree of influence of structural similarity, This is the weighting coefficient of recommendation accuracy in the fitness function, used to control the degree of contribution to recommendation quality;

[0107] All individuals were sorted according to their fitness values ​​from highest to lowest, and then divided into groups using an equal-division strategy. Within each subgroup, the individual with the highest fitness value is selected as the local leader. The remaining individuals will perform localized guidance and updates around the local navigator;

[0108] Calculate the current iteration round Dynamic inertia factor :

[0109] ;

[0110] in, The initial maximum inertia weight value, The maximum number of iterations, The minimum inertia weight value set;

[0111] Calculate candidate geomagnetic navigation positions for each individual. Landmark-guided location candidate values :

[0112] Geomagnetic navigation update:

[0113] ;

[0114] Landmark guidance update:

[0115] ;

[0116] in, Let be the position vector of the i-th individual in the t-th iteration. Let i be the velocity vector of individual i in round t. To guide the weighting coefficients, This refers to the individual memory weighting coefficient. This represents the position of the individual with the best fitness in the current population. This represents the historical best memory position of the i-th individual;

[0117] Introducing stage control factors The system dynamically balances two candidate positions through a fusion mechanism, performs position updates, and updates the individual velocity in the next round based on the position update results of this round.

[0118] ;

[0119] in, For the i-th individual in the i-th... The position vector during round iteration;

[0120] Calculate the new round of individual locations The contrastive loss of the corresponding graph contrastive learning network And recommendation accuracy Update the fitness function ;

[0121] If the current individual fitness Then the memory vector will be updated to the current solution. ;

[0122] Repeat the process until the maximum number of iterations is met. Alternatively, a fitness convergence threshold condition may be used to ultimately output the individual with the best fitness. , which serves as the optimal combination of structural parameters for graph contrastive learning networks.

[0123] This invention addresses the problem of traditional graph neural network structural parameters not being automatically optimized. It introduces a pigeon flock optimization algorithm with a phased navigation mechanism, achieving adaptive optimization of structural parameters within an unsupervised graph contrastive learning framework, significantly improving the model's structural expressiveness and recommendation performance. First, the algorithm sets three key structural parameters—geographical location similarity weighting coefficient, subgraph perception scale control factor, and node occlusion ratio adjustment factor—corresponding to spatial sensitivity, subgraph extraction granularity, and feature perturbation intensity in the graph model, respectively, constituting the parameter combination for each individual. A fitness function is used to jointly evaluate each individual's performance in both node semantic alignment (contrastive loss) and recommendation accuracy, ensuring that the optimization objective balances embedding quality and recommendation effect. During the search process, the algorithm sorts the population by fitness and divides it into subgroups, using the individual with the highest fitness as a local leader to guide the remaining individuals in local updates. Simultaneously, a dynamic inertia factor, geomagnetic navigation location candidates, and landmark-guided location candidate mechanisms are introduced to simulate the pigeon flock's search behavior at different stages: emphasizing exploration ability in the early stages and convergence accuracy in the later stages. The phase control factor integrates the two guiding directions to ensure a smooth transition of the search path. Furthermore, each individual updates its memory vector based on its actual optimization performance after each iteration, enabling the utilization and learning of historical experience. Ultimately, the system outputs a graph contrastive learning network model with optimal structure through multiple rounds of optimization iterations. The introduction of this algorithm significantly solves the problem that traditional fixed structures cannot adapt to different task characteristics, achieving adaptive optimization of the graph contrastive learning model under different data graph structures and semantic scenarios. This improves the accuracy of recommendations, the model's generalization ability, and robustness, providing a sustainable evolutionary optimization mechanism for large-scale map search and recommendation systems.

[0124] In this embodiment, generating the semantic relevance ranking list includes:

[0125] Receive map search requests initiated by users on the terminal and trigger the location-aware recommendation process initialization operation;

[0126] Extract the latitude and longitude coordinates of the user's current location from the BeiDou database, and use the latitude and longitude coordinates of the user's current location as location nodes to construct graph node entities;

[0127] The location node is added as the user's current location node to the currently maintained heterogeneous graph structure, thereby expanding the spatial representation capability of the graph structure;

[0128] Using the user's current location node as the central node, the associated local subgraph structure is extracted from the heterogeneous graph structure based on spatial adjacency and access relationships;

[0129] The optimized graph contrast learning network is invoked to perform embedding representation calculation on the extracted local subgraphs and generate a semantic representation for each candidate node.

[0130] Based on the semantic relevance between the embedded representations of the user's current location node and candidate nodes, a ranked list is generated and output as a set of candidate results for search recommendations.

[0131] This invention introduces a subgraph embedding computation mechanism driven by the user's current location, achieving deep integration of map search requests and the current location context, significantly improving the spatial matching and semantic relevance of recommendation results. Upon receiving a map search request, the system automatically extracts the user's current location coordinates from the BeiDou database, constructs them as graph nodes, and dynamically injects them into a heterogeneous graph structure. This allows the user's location to participate in the graph modeling process in real time, effectively expanding the spatial awareness capabilities of the graph structure. Extracting local subgraphs centered on the current location node, based on its spatial adjacency and access relationships, accurately captures geographic objects and path relationships strongly associated with the current location, avoiding interference and computational redundancy caused by full-graph modeling. By calling an optimized graph contrastive learning network, embedding computation is performed on the local subgraphs to obtain the semantic representations of candidate nodes. Combined with the embedding vectors of the user's current location node, semantic relevance is calculated, and a ranking list is generated, ensuring that the recommendation results not only match the user's query semantics but also satisfy the practicality of the geographic location. This mechanism enables spatial semantic fusion recommendations based on the user's current location, and has significant advantages such as fast response speed, strong location adaptability, accurate semantic matching, and good recommendation interpretability. It is especially suitable for mobile intelligent map search and recommendation tasks in complex scenarios.

[0132] In this embodiment, the user feedback behavior data specifically includes click records, browsing dwell time, historical search paths, and actual navigation behavior, which are used to optimize the retraining process of heterogeneous graph structures and graph contrastive learning networks.

[0133] refer to Figure 2 The cultural tourism map search system based on deep learning and the BeiDou database includes the following modules:

[0134] The request processing module is used to receive user map search requests and extract semantic representations as user node features;

[0135] The mapping module is used to extract user location and trajectory data from the BeiDou database and combine it with geographic object information to construct a heterogeneous map structure.

[0136] The graph learning optimization module is used to generate perturbation views based on heterogeneous graphs, train the graph contrast learning network, and adjust the structural parameters through the pigeon flocking optimization algorithm.

[0137] The subgraph calculation module is used to extract local subgraphs and generate node embedding representations;

[0138] The recommendation module generates ranking results based on the semantic relevance of the embedded vectors, and outputs recommended paths and explanatory information;

[0139] The feedback update module is used to collect user feedback behavior data and use it for graph structure updates and model retraining.

[0140] Example 1:

[0141] To verify the feasibility of this invention in practice, it was applied to a cultural tourism map search application test. A city's cultural tourism data center collaborated with an artificial intelligence company to conduct a comparative experiment on map recommendation systems based on users' natural language map queries. The test system deployed both a traditional keyword retrieval scheme and this invention, and provided services to users via mobile terminals, collecting their actual search and navigation behaviors. Participating users included local and out-of-town tourists who initiated map search requests in different usage scenarios such as daily travel, weekend leisure, and museum visits. After the system returned results, it recorded the number of clicks, average dwell time, recommendation accuracy, and actual access path.

[0142] In traditional systems, user search requests such as "quiet parks suitable for walks," "cultural buildings suitable for photography," and "museums with exhibitions" are often returned results solely through keyword matching, which can easily lead to recommendations that do not match the user's actual intent. For example, when searching for "cultural locations suitable for quiet reading," the system might recommend densely populated, commercially-oriented areas due to the high frequency of the keyword "culture," ignoring the user's preference for a quiet atmosphere, resulting in an unbalanced recommendation ranking. In contrast, when using the system of this invention, the search request is first converted into a semantic vector, which is then semantically matched with user nodes, geographic object nodes, and path nodes in a constructed heterogeneous graph structure. This, combined with the user's current location, generates a relevance ranking through subgraph embedding, achieving a deep understanding of the user's intent.

[0143] In the experiment, the system tested ten location types, including cultural and historical sites, urban parks, exhibition halls and museums, commercial districts, areas surrounding universities and bookstores, local characteristic streets, quiet cultural sites, nighttime attractions, popular photography spots, and family-friendly recreational areas. The control group system often failed to grasp the core intent when faced with fuzzy, highly descriptive natural language queries, resulting in low recommendation click-through rates, inaccurate ranking, and short user dwell time. In contrast, the system of this invention, after loading real-time user location and historical trajectory data recorded in the BeiDou database, constructs a heterogeneous graph structure including spatial adjacency, access relationships, and behavioral paths. Then, through graph comparison learning network embedding computation, it effectively improves the recommendation matching accuracy.

[0144] Taking "exhibition halls and museums" as an example, the traditional system's recommended clicks were 132, with an average dwell time of 61.87 seconds, ranking 5th. In contrast, the system of this invention, under the same conditions, increased clicks to 171, average dwell time to 76.20 seconds, and ranked 1st. Similarly, in the "universities and bookstores surrounding" scenario, the traditional method resulted in 102 clicks and a 5th-place recommendation, while the system of this invention increased clicks to 138, ranking 2nd, and dwell time increased from 59.03 seconds to 79.98 seconds, demonstrating higher user acceptance of the recommendations. Through an optimized graph contrastive learning model, the system can better understand the abstract semantics of "quiet atmosphere," "suitable for nighttime visits," and "rich cultural atmosphere," and combined with geospatial trajectory reasoning, the recommended paths are more realistic and reliable.

[0145] More importantly, in this invention's system, user clicks, browsing, and navigation behaviors are continuously collected as feedback data, constantly optimizing the graph structure's edge weights and node states. This data is also used for retraining the graph contrastive learning network, thus enabling the model to adaptively adjust and evolve. Experimental results show that, among the overall user group, the average click volume of this invention's system is approximately 31.6% higher than that of the traditional system, the average dwell time is increased by approximately 23%, the recommendation ranking is significantly improved, and the actual access rate increases from 67% to nearly 90%. User satisfaction scores in questionnaires rise from 3.4 (traditional system) to 4.6 (this invention's system).

[0146] Table 1. Comparison of Recommendation Performance of Map Search Systems

[0147]

[0148] Overall, the system of this invention exhibits higher click-through rates, longer user dwell time, and higher recommendation rankings across all location types compared to traditional keyword retrieval systems, demonstrating a significant performance advantage in map search and recommendation tasks. In the "Cultural and Historical Sites" category, the traditional system recorded 135 clicks, while the system of this invention increased this to 180, a growth of 33.3%. The average user dwell time increased from 56.43 seconds to 67.14 seconds, and the recommendation ranking jumped from 5th to 1st. This indicates that the system of this invention is better able to match the semantic needs of "cultural and historical" with geographical intent.

[0149] In the "Exhibition Halls and Museums" category, the traditional system received 132 clicks with an average dwell time of 61.87 seconds and ranked 5th. In contrast, the system of this invention, under the same query conditions, received 171 clicks, with a dwell time increasing to 76.20 seconds, and ranked 1st, indicating a higher user acceptance of the recommendation results. This trend is particularly evident in scenarios such as "Urban Parks," "Quiet Cultural Sites," and "Popular Photography Spots," which are sensitive to both semantics and location. The system of this invention effectively reflects user preferences as identifiable embedded features and outputs more relevant candidate locations.

[0150] It is worth noting that in the category of "family-friendly play area," where the target user is relatively clear but the semantic description is diverse, the traditional system recommended it in 7th place with 119 clicks. However, the system of this invention can recognize the semantic expressions "family-friendly," "suitable for children," and "safe and open," significantly improving its recommendation ranking to 1st place, increasing the number of clicks to 164, and raising the dwell time to 74.21 seconds. This result verifies the high adaptability of the system of this invention to fuzzy query intent.

[0151] From a user behavior perspective, increased clicks directly reflect improved relevance of recommended content, while increased average dwell time demonstrates a close connection between recommendations and user interests. Furthermore, improved recommendation ranking enhances user satisfaction and efficiency. Traditional systems often rank recommendations low in various scenarios, indicating that their retrieval logic fails to understand the true intent behind the query; the system merely performs word-based matching, leading to multiple clicks or low search efficiency. This invention, by fusing heterogeneous graph structures with an optimized graph contrastive learning network model, establishes an effective connection between spatial adjacency, behavioral path information, and semantic representation, making the recommendation results more reasonable in terms of ranking and interpretability.

[0152] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for searching cultural tourism maps based on deep learning and the BeiDou database, characterized in that, Includes the following steps: S1. Obtain the user's current location data and historical movement trajectory data from the Beidou database, load a static information set containing multiple cultural and tourism geographic objects, and construct a heterogeneous graph structure. The heterogeneous graph structure includes user nodes, geographic object nodes, and path nodes. Access edges, spatial adjacency edges, and trajectory order edges are established between nodes. S2. Receive the natural language search request input by the user, perform context semantic vector extraction processing on the natural language search request, obtain the semantic representation result, and write it into the heterogeneous graph structure as the initial feature of the user node. S3. Construct a graph contrast learning network based on heterogeneous graph structure, perform graph structure perturbation to generate two sets of views, input the two sets of views into the graph neural network respectively, and perform node embedding learning; S4. The structural parameters of the graph contrast learning network are optimized using the pigeon flock optimization algorithm, and the optimal structure and parameter configuration of the graph contrast learning network are finally output, forming the optimized graph contrast learning network. S5. When a user initiates a map search request, the system extracts the user's current location coordinates from the BeiDou database, constructs graph nodes based on the user's current location coordinates and adds them to the heterogeneous graph structure, and extracts local subgraphs from the heterogeneous graph structure as the central node. The system then calls an optimized graph contrast learning network to perform node embedding representation calculations on the local subgraphs and generates a semantic relevance ranking list. S6. Based on the semantic relevance ranking list and the connection relationship of path nodes in the heterogeneous graph structure, perform multi-hop path reasoning to generate recommendation results containing the optimal path sequence from the user's current location to the target geographic object and recommendation explanation information. Return the recommendation results to the user terminal and collect user feedback behavior data for incremental updates of the heterogeneous graph structure and retraining of the graph contrastive learning network.

2. The cultural tourism map search method based on deep learning and BeiDou database according to claim 1, characterized in that, The BeiDou database is a collection of navigation data that records the user's real-time location and movement path, including latitude and longitude coordinates, timestamps, speed and direction information; the cultural and tourism geographic objects include cultural and tourism-related entities with geographic locations, tag classifications, and rating records.

3. The cultural tourism map search method based on deep learning and BeiDou database according to claim 1, characterized in that, S2 specifically includes: S21. Receive a natural language search request input by the user through the terminal, wherein the natural language search request includes geographical preferences, behavioral intentions, travel purpose and topic keywords; S22. Perform text preprocessing on natural language search requests, including word segmentation, stop word removal, word unification, and invalid symbol removal, to generate standardized sentence input; S23. Perform context semantic vector extraction processing on the preprocessed natural language input text to generate a context-dependent semantic vector representation; S24. Perform dimension normalization and feature alignment on the extracted semantic vector representation to meet the requirements of dimension consistency and structural constraints of node feature representation in heterogeneous graph structures. S25. Bind the semantic vector representation with the user identity identifier to generate a complete user node feature representation data structure; S26. Write the generated user node features into the corresponding user node positions in the heterogeneous graph structure to complete the semantic embedding mapping process from natural language search request to graph structure.

4. The cultural tourism map search method based on deep learning and BeiDou database according to claim 1, characterized in that, S3 specifically includes: S31. Constructing the original graph based on heterogeneous graph structure ,in, This represents a set of nodes, including user nodes, geographic object nodes, and path nodes. This represents a set of edges, including visit edges, spatially adjacent edges, and trajectory order edges. S32. Introduce a weighted coefficient for geographic location similarity. Construct a location-aware similarity function This is used to fuse spatial relationships between nodes in contrastive loss: ； in, Represents a node and Geographic similarity, The cosine similarity between embedded vectors is represented. For nodes The embedding vector representation, For nodes Embedded vector representation; S33. Introducing a subgraph perception scale control factor Used to determine the adjacency depth of the local subgraph extracted from the central node, in the first view Second view In the construction, the number of jumps is respectively used as The neighborhood is used as the subgraph extraction range, and node embedding extraction is performed at the subgraph level; S34. Introduce a node occlusion ratio adjustment factor. This is used to control the occlusion ratio of node features during view perturbation, in the construction of the first view. Second view At that time, according to the ratio Masking the node feature dimensions in the original feature matrix enhances the structural robustness of the embedding. S35, respectively view the first view Second view Input a graph neural network encoder to obtain the embedding representation of each node in two sets of views. , ; S36. Construct a position-aware contrast loss function This is used to optimize the consistency of the embedded representation of the same node in two views: ； in, To compare temperature coefficients, It is a logarithmic function. It is an exponential function; S37. The graph neural network is trained based on the location-aware contrastive loss function, and finally a graph contrastive learning network representation structure that integrates geographic awareness, adjustable granularity and dynamic perturbation strategy is obtained.

5. The cultural tourism map search method based on deep learning and BeiDou database according to claim 4, characterized in that, S4 specifically includes: S41. Set the initial population size for the pigeon flock optimization algorithm as follows: Initialize the set of individual pigeons in the flock. Each individual This represents the combination of structural parameters of a graph-based contrast learning network, where... , This represents the weighted coefficient of geographical location similarity corresponding to the i-th individual. Let be the subgraph perception scale control factor corresponding to the i-th individual. is the node occlusion ratio adjustment factor corresponding to the i-th individual; S42. Randomly initialize the velocity vector for each individual. And set individual memory vectors This is used to record an individual's historical optimal parameter configuration; S43. For each individual parameter combination Based on the constructed graph contrastive learning network structure, node embedding computation and contrastive learning training processes are performed to obtain the corresponding contrastive loss function values. Recommendation accuracy with node ranking Construct the fitness function : ； in, To compare the weighting coefficients of the loss in the fitness function, which are used to control the degree of influence of structural similarity, This is the weighting coefficient of recommendation accuracy in the fitness function, used to control the degree of contribution to recommendation quality; S44. Sort all individuals according to their fitness values ​​from highest to lowest, and divide them into equal groups. Within each subgroup, the individual with the highest fitness value is selected as the local leader. The remaining individuals will perform localized guidance and updates around the local navigator; S45. Calculate the current iteration round. Dynamic inertia factor : ； in, The initial maximum inertia weight value, The maximum number of iterations, The minimum inertia weight value set; S46. Calculate the candidate geomagnetic navigation position values ​​for each individual. Landmark-guided location candidate values : Geomagnetic navigation update: ； Landmark guidance update: ； in, Let be the position vector of the i-th individual in the t-th iteration. Let i be the velocity vector of individual i in round t. To guide the weighting coefficients, This refers to the individual memory weighting coefficient. This represents the position of the individual with the best fitness in the current population. The best historical memory position for the i-th individual; S47. Introduction of Stage Control Factors The system dynamically balances two candidate positions through a fusion mechanism, performs position updates, and updates the individual velocity in the next round based on the position update results of this round. ； in, For the i-th individual in the i-th... The position vector during round iteration; S48. Calculate the new round of individual positions. The contrastive loss of the corresponding graph contrastive learning network And recommendation accuracy Update the fitness function ; S49. If the current individual fitness Then update the memory vector to the current solution. ; S410. Repeat steps S44 to S49 until the maximum number of iterations is met. Alternatively, a fitness convergence threshold condition may be used to ultimately output the individual with the best fitness. , which serves as the optimal combination of structural parameters for graph contrastive learning networks.

6. The cultural tourism map search method based on deep learning and BeiDou database according to claim 1, characterized in that, S5 specifically includes: S51. Receive the map search request initiated by the user on the terminal and trigger the location-aware recommendation process initialization operation; S52. Extract the latitude and longitude coordinates of the user's current location from the BeiDou database, and use the latitude and longitude coordinates of the user's current location as location nodes to construct graph node entities; S53. Add the location node as the user's current location node to the currently maintained heterogeneous graph structure to expand the spatial representation capability of the graph structure; S54. Using the user's current location node as the center node, extract the associated local subgraph structure from the heterogeneous graph structure based on spatial adjacency and access relationships; S55. Call the optimized graph contrast learning network to perform embedding representation calculation on the extracted local subgraphs and generate semantic representations for each candidate node; S56. Based on the semantic relevance between the embedded representations of the user's current location node and the candidate nodes, generate a ranking list and output it as a set of candidate results for search recommendation.

7. The cultural tourism map search method based on deep learning and BeiDou database according to claim 1, characterized in that, The user feedback behavior data specifically includes click records, browsing dwell time, historical search paths, and actual navigation behavior, which are used to optimize the retraining process of heterogeneous graph structures and graph contrastive learning networks.

8. A cultural tourism map search system based on deep learning and BeiDou database, comprising the cultural tourism map search method based on deep learning and BeiDou database as described in any one of claims 1 to 7, characterized in that, Includes the following modules: The request processing module is used to receive user map search requests and extract semantic representations as user node features; The mapping module is used to extract user location and trajectory data from the BeiDou database and combine it with geographic object information to construct a heterogeneous map structure. The graph learning optimization module is used to generate perturbation views based on heterogeneous graphs, train the graph contrast learning network, and adjust the structural parameters through the pigeon flocking optimization algorithm. The subgraph calculation module is used to extract local subgraphs and generate node embedding representations; The recommendation module generates ranking results based on the semantic relevance of the embedded vectors and outputs recommended paths and explanatory information. The feedback update module is used to collect user feedback behavior data and use it for graph structure updates and model retraining.

Images