Travel route planning method and system based on artificial intelligence

Abstract

The invention discloses a tourist route planning method and system based on artificial intelligence. The method comprises the following steps: step 1, collecting multi-modal input data and constructing a multi-modal data set; step 2, inputting the multi-modal data set into a Perciver IO model for encoding; 3, constructing an original urban environment map; 4, generating a personalized urban environment map based on the intention-environment bidirectional remapping network; 5, constructing a tourism path set; step 6, optimizing the travel path set based on an improved Rhododendron search algorithm, and introducing a heterogeneous reconstruction egg exchange mechanism to realize jumping type evolution updating; and step 7, ending evolution updating when a preset evolution termination condition is met, generating an optimized travel path set, and selecting the travel path with the highest fitness as a current recommended path. According to the method, multi-modal modeling and the improved Rhododendron search algorithm are fused, and efficient planning of the personalized travel route is realized.

Description

Technical Field

[0001] This invention relates to the field of tourism recommendation technology, and in particular to a tourism route planning method and system based on artificial intelligence. Background Technology

[0002] With the rapid growth in demand for smart travel services and personalized travel recommendations, how to automatically plan high-quality travel routes based on user interests and urban spatial characteristics has become an important research direction in the fields of artificial intelligence and smart cities. Existing travel route planning methods mostly rely on preset attraction levels, historical check-in popularity, or user ratings for route construction, but in practical applications, they generally have the following problems:

[0003] The lack of deep semantic understanding of multimodal user preference data (such as text, images, and behavioral records) makes it difficult to accurately depict individualized tourism intentions; the construction of urban environmental maps is based solely on static traffic or attraction distribution, failing to dynamically adjust according to user interests, resulting in low matching degree between generated paths and actual user preferences; most optimization algorithms adopt traditional genetic algorithms or ant colony algorithms, with small granularity and slow convergence in the path update process, making it easy to get trapped in local optima; path replacement is only based on single points or edges, lacking a structural-level reconstruction mechanism, which limits the leapfrog evolution capability of tourism paths and makes it difficult to effectively explore diverse and high-quality path spaces.

[0004] Therefore, how to provide a tourism route planning method and system based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an artificial intelligence-based tourism route planning method and system. This invention accurately analyzes user intent by constructing a Perceiver IO model, generates personalized urban environment maps by combining intent-environment bidirectional remapping networks, and improves the diversity of path evolution and global search capabilities by introducing an improved cuckoo search algorithm with a heterogeneous reconstruction egg exchange mechanism. Finally, it selects the tourism route with the highest fitness as the current recommended route. It has comprehensive advantages such as high personal matching degree, strong evolution update efficiency, and high recommendation result quality, effectively overcoming the problems of single path and optimization getting stuck in local optima in traditional methods. It is applicable to various application scenarios such as intelligent tourism services.

[0006] According to an embodiment of the present invention, a tourism route planning method based on artificial intelligence includes the following steps:

[0007] Step 1: Collect multimodal input data from users and construct a multimodal dataset;

[0008] Step 2: Input the multimodal dataset into the Perceiver IO model for encoding to obtain the user intent representation vector;

[0009] Step 3: Construct the original urban environment map, which consists of interest nodes, reachable path edges, and initial edge weights;

[0010] Step 4: Input the user intent representation vector into the intent-environment bidirectional remapping network, calculate the importance weights of interest nodes and personalized edge weights, adjust the original urban environment map, and generate a personalized urban environment map.

[0011] Step 5: Based on the personalized urban environment map, generate multiple tourist routes according to the set time budget constraints and maximum number of nodes constraints. The tourist routes are composed of interest nodes arranged in the order of visit, and construct a set of tourist routes.

[0012] Step Six: Optimize the set of tourist routes based on the improved cuckoo search algorithm. The improvement of the improved cuckoo search algorithm is that it introduces a heterogeneous reconstruction egg exchange mechanism. In each round of tourist route replacement, the tourist route is divided into several functional segments to achieve leapfrog evolution and update.

[0013] Step 7: When the preset evolution termination condition is met, end the evolution update, generate an optimized set of travel routes, and select the travel route with the highest fitness as the current recommended route.

[0014] Optionally, the multimodal input data specifically includes text information, image information, user history behavior records, and interest preference information.

[0015] Optionally, step two specifically involves:

[0016] The text information is segmented into words, and words are converted into text token embedding vectors through word vector encoding.

[0017] The image information is divided into multiple image blocks, and each image block is flattened and convolutionally encoded to obtain the image token embedding vector.

[0018] The user's historical behavior records and interest preference information are structured and each behavior record is represented as a multi-field data structure including behavior timestamp, interest tag, access object type and interaction operation type. The structured fields are mapped to a vector representation of a set dimension through a multilayer perceptron coding network to obtain the behavior token embedding vector.

[0019] The text token embedding vector, image token embedding vector, and behavior token embedding vector are concatenated to form a unified token sequence, and modality type encoding and location information encoding are added to each token to form the model input tensor;

[0020] The model input tensor is input into the cross-attention encoding module of the Perceiver IO model. Cross-modal associations are established among all input tokens through the attention mechanism. Multimodal token information is integrated using global attention weighting to generate latent representations.

[0021] The latent representation is input into a stacked self-attention module for multi-layer representation updates. The updated latent representation is then mapped to an output token sequence through a reverse cross-attention module. The output token sequence is then pooled to obtain the user intent representation vector.

[0022] Optionally, step three specifically includes:

[0023] Based on the city's geospatial information data, we screen geographical entities with tourism, leisure, culture and entertainment functions, construct a set of interest nodes, and extract structured attributes such as geographical coordinates, category, opening time and popularity level for each interest node.

[0024] Based on road network data, pedestrian access data, and public transportation route maps, access paths in the city are extracted and analyzed to identify traffic paths connecting different interest nodes and obtain access path edges. The access path edges include the identifiers of the starting and ending interest nodes, path length, estimated travel time, and traffic mode type attribute information.

[0025] The path length and estimated travel time between interest nodes on the reachable path are weighted and calculated to generate initial edge weights.

[0026] Based on the interest nodes, access path edges, and initial edge weights, an original urban environment graph is constructed to represent the connection relationships between interest nodes in the city.

[0027] Optionally, the intent-environment bidirectional remapping network specifically includes an interest node enhancement module, an edge weight adjustment module, and a graph fusion module;

[0028] The user intent representation vector is input into the interest node enhancement module, and the cosine similarity between the user intent representation vector and the attribute vector of each interest node in the original urban environment map is calculated to generate the importance weight of the interest node.

[0029] Based on the importance weight of interest nodes, the set of interest nodes in the original urban environment map is filtered out, and interest nodes whose importance weight is less than the preset importance weight threshold are removed. Also, based on the connection relationship between interest nodes, associated access path edges are removed simultaneously.

[0030] The edge weight adjustment module extracts the importance weights of the starting interest nodes at both ends of each access path edge, and performs a weighted fusion based on the importance weights of the starting interest nodes at both ends and the initial edge weight to calculate the personalized edge weight of each access path edge.

[0031] The initial edge weight of each reachable path edge is replaced with a personalized edge weight, and reachable path edges with personalized edge weights greater than a preset edge weight threshold are subjected to connection pruning operations.

[0032] A personalized urban environment map is generated based on the retained interest nodes, the updated access path edges, and the personalized edge weights.

[0033] Optionally, step five specifically includes:

[0034] Based on the set of interest nodes and access path edge information in the personalized urban environment map, the importance weights of the interest nodes are sorted in descending order, and the top N interest nodes are selected to construct the initial candidate node set.

[0035] Each interest node is selected sequentially from the set of initial candidate nodes as the starting point of the path, the travel path is initialized, and time budget constraints and maximum number of nodes constraints are set for the construction of the travel path.

[0036] Using the last interest node in the current travel path as a reference node (when the travel path just begins to expand, the starting candidate node is the starting point of the path and also the last interest node), traverse the adjacent interest nodes that are connected by accessible path edges, sort them based on the personalized edge weights of the accessible path edges, and select the interest node with the smallest personalized edge weight that does not appear in the current travel path as the next expansion node.

[0037] The cumulative travel time is updated in real time after each node expansion. When the cumulative travel time exceeds the preset time budget, or the number of visited nodes exceeds the preset maximum number of nodes, the path expansion process is terminated.

[0038] Complete the tourism path construction process for all initial candidate nodes in sequence, generate multiple tourism paths that meet the time budget constraint and the maximum number of nodes constraint, and construct a tourism path set.

[0039] Optionally, step six specifically includes:

[0040] The population is initialized as a set of travel routes, and the fitness of each travel route is calculated as follows:

[0041] The importance weights of all nodes of interest in the travel route are summed and multiplied by a preset first weighting factor to obtain the first calculated value; the personalized edge weights of all accessible path edges in the travel route are summed and multiplied by a preset second weighting factor to obtain the second calculated value; the fitness of the travel route is obtained by subtracting the second calculated value from the first calculated value.

[0042] In each round of tourism route evolution and update, the tourism route with the lowest fitness value is selected from the current tourism route set as the target route, and a tourism route is randomly selected from the top 20% of tourism routes with fitness values ​​from high to low as the parent route.

[0043] Based on the category attribute of the interest nodes, consecutive interest nodes of the same category in the parent path are divided into a functional segment to construct a sequence of functional segments of the tourism path;

[0044] The heterogeneous reconstruction egg exchange mechanism of the improved cuckoo search algorithm specifically includes: randomly selecting one or more functional segments from the functional segment sequence and replacing functional segments with the same category attributes in the target path;

[0045] If no functional segment of the same category exists in the target path, the selected functional segment is inserted into the target path in the original access order to generate a candidate path.

[0046] The candidate paths are checked for connectivity integrity. If there are two adjacent interest nodes that are not connected, the shortest path connecting the two adjacent interest nodes is found in the personalized urban environment map, and an intermediate interest node is inserted to connect the paths.

[0047] The fitness of the candidate paths after connection integrity verification is recalculated. If the fitness of the candidate path is higher than that of the original target path, the target path is replaced by the candidate path and it enters the next generation of the population. Otherwise, the original target path remains unchanged.

[0048] Optionally, dividing consecutive interest nodes of the same category in the parent path into a functional segment specifically involves:

[0049] The parent path is scanned according to the category attributes of the interest nodes. Two or more interest nodes that are visited consecutively and have the same category attributes are divided into a functional segment. The position index of the start and end interest nodes in the parent path is recorded for each functional segment. This is used by the heterogeneous reconstruction egg exchange mechanism to find matching segments and determine the insertion position in the target path.

[0050] Optionally, the preset evolution termination condition includes at least one of the following:

[0051] Scenario 1: The current evolutionary generation has reached the preset maximum generation threshold;

[0052] Scenario 2: The maximum fitness improvement in the set of tourism routes across several consecutive generations is less than the set convergence threshold;

[0053] Scenario 3: The fitness standard deviation of all travel routes in the current population is less than the set stability threshold;

[0054] During the tourism route optimization process, the preset evolution termination conditions are continuously monitored. If the preset evolution termination conditions are met (i.e., any one or more of the above situations occur), the evolution process is immediately terminated, the final generation is output as the optimized tourism route set, and the tourism route with the highest fitness is selected as the current recommended route.

[0055] According to an embodiment of the present invention, a tourism route planning system based on artificial intelligence includes the following modules:

[0056] The multimodal data acquisition module is used to collect users' multimodal input data and construct multimodal datasets;

[0057] The user intent modeling module is used to input the multimodal dataset into the Perceiver IO model for encoding and to generate user intent representation vectors.

[0058] The urban environment map construction module is used to construct the original urban environment map, which consists of interest nodes, reachable path edges, and initial edge weights.

[0059] The map personalization adjustment module is used to input the user intent representation vector into the intent-environment bidirectional remapping network, calculate the importance weight of interest nodes and the personalized edge weight of access paths, adjust the original urban environment map, and generate a personalized urban environment map.

[0060] The initial path generation module is used to generate multiple tourist routes based on a personalized urban environment map, according to the set time budget constraints and maximum node number constraints, and to construct a set of tourist routes.

[0061] The path optimization module is used to optimize the set of tourist paths based on the improved rhododendron search algorithm. The improved rhododendron search algorithm includes a heterogeneous reconstruction egg exchange mechanism, which divides the parent path into functional segments and performs structural-level reconstruction in each round of evolution to achieve leapfrog evolution updates, while performing connection integrity verification and fitness calculation.

[0062] An evolution control module is used to set and detect preset evolution termination conditions. If the preset evolution termination conditions are met, the evolution process is stopped.

[0063] The route recommendation module is used to select the most adaptive tourist route from the final optimized tourist route set after the evolution process is completed, and output it as the current recommended route.

[0064] The beneficial effects of this invention are:

[0065] This invention addresses the problems of ambiguous user intent modeling, low path matching accuracy, and local optima in existing tourism route planning by introducing the Perceiver IO model and the intent-environment bidirectional remapping network. It adopts a unified multimodal data fusion and personalized urban map reconstruction mechanism, combined with an improved cuckoo search algorithm based on a heterogeneous reconstruction egg exchange mechanism, to achieve leapfrog evolution optimization of tourism routes. In the user intent modeling stage, a cross-attention mechanism is used to deeply fuse and encode text, image, and behavioral data, outputting a high-dimensional unified user intent representation vector that effectively captures multimodal preference features. In the environment modeling stage, the weights of interest nodes and access paths in the urban environment map are dynamically adjusted based on user intent, achieving map reconstruction that is highly coupled with individual preferences. In the path optimization stage, the parent path is divided into functional segments and a heterogeneous reconstruction egg exchange mechanism is introduced to enhance the path evolution span and diversity. The accessibility and superiority of the path are ensured through connection integrity verification and fitness evaluation. Ultimately, the system can efficiently generate personalized tourism recommendation paths with reasonable structure, high preference matching, and strong convergence in complex urban traffic environments based on user interests, significantly improving the intelligence and personalization level of path planning. Attached Figure Description

[0066] 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:

[0067] Figure 1 This is an overall flowchart of a tourism route planning method based on artificial intelligence proposed in this invention;

[0068] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based tourism route planning system proposed in this invention. Detailed Implementation

[0069] 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.

[0070] refer to Figure 1 An artificial intelligence-based tourism route planning method includes the following steps:

[0071] Step 1: Collect multimodal input data from users and construct a multimodal dataset;

[0072] Step 2: Input the multimodal dataset into the Perceiver IO model for encoding to obtain the user intent representation vector;

[0073] Step 3: Construct the original urban environment map, which consists of interest nodes, reachable path edges, and initial edge weights;

[0074] Step 4: Input the user intent representation vector into the intent-environment bidirectional remapping network, calculate the importance weights of interest nodes and personalized edge weights, adjust the original urban environment map, and generate a personalized urban environment map.

[0075] Step 5: Based on the personalized urban environment map, generate multiple tourist routes according to the set time budget constraints and maximum number of nodes constraints. The tourist routes are composed of interest nodes arranged in the order of visit, and construct a set of tourist routes.

[0076] Step Six: Optimize the set of tourist routes based on the improved cuckoo search algorithm. The improvement of the improved cuckoo search algorithm is that it introduces a heterogeneous reconstruction egg exchange mechanism. In each round of tourist route replacement, the tourist route is divided into several functional segments to achieve leapfrog evolution and update.

[0077] Step 7: When the preset evolution termination condition is met, end the evolution update, generate an optimized set of travel routes, and select the travel route with the highest fitness as the current recommended route.

[0078] In this embodiment, the multimodal input data specifically includes text information, image information, user history behavior records, and interest preference information.

[0079] In this embodiment, step two specifically includes:

[0080] The text information is segmented into words, and words are converted into text token embedding vectors through word vector encoding.

[0081] The image information is divided into multiple image blocks, and each image block is flattened and convolutionally encoded to obtain the image token embedding vector.

[0082] The user's historical behavior records and interest preference information are structured and each behavior record is represented as a multi-field data structure including behavior timestamp, interest tag, access object type and interaction operation type. The structured fields are mapped to a vector representation of a set dimension through a multilayer perceptron coding network to obtain the behavior token embedding vector.

[0083] The text token embedding vector, image token embedding vector, and behavior token embedding vector are concatenated to form a unified token sequence, and modality type encoding and location information encoding are added to each token to form the model input tensor;

[0084] The model input tensor is input into the cross-attention encoding module of the Perceiver IO model. Cross-modal associations are established among all input tokens through the attention mechanism. Multimodal token information is integrated using global attention weighting to generate latent representations.

[0085] The latent representation is input into a stacked self-attention module for multi-layer representation updates. The updated latent representation is mapped to an output token sequence through a reverse cross-attention module. The output token sequence is then pooled to obtain the user intent representation vector.

[0086] The Perceiver IO model in this invention is a general neural network structure suitable for multimodal information processing. It encodes multimodal input data from text, images, user behavior records, and interest preference information into a unified user intent representation vector. The Perceiver IO model mainly consists of three parts: a cross-attention module, a self-attention module, and a reverse cross-attention module. First, the encoded multimodal token sequence is input to the cross-attention module. This module, based on the attention mechanism in Transformer, establishes cross-modal associations among all input tokens, achieving semantic fusion between different modalities. This module does not directly perform self-attention calculation on all tokens; instead, it aggregates the embedded information of all tokens into a compressed representation through the attention mechanism, improving processing efficiency and semantic abstraction capabilities.

[0087] After generating the initial fused latent representation, it is input into a stacked self-attention module for multi-layer representation updates. This module adopts a standard multi-head self-attention structure, containing multiple layers of Transformer encoder units. Each layer models the semantic dependencies within the representation through a self-attention mechanism, and combines a feedforward network to complete feature transformation and normalization processing, thereby progressively improving the abstract expressive power of the representation.

[0088] The updated latent representation is further fed into the inverse cross-attention module. This module is structurally similar to the cross-attention module but operates in the opposite direction. Its core principle is to construct an attention mechanism using the token embeddings from the input phase as queries and the latent representations as keys and values. By calculating the attention distribution of each input token to each unit in the latent representation, the most relevant semantic information is extracted, achieving the reconstruction from abstract representations to specific modal semantics. Finally, pooling operations are used to aggregate the output token sequence into vectors, resulting in a fixed-dimensional user intent representation vector.

[0089] In this embodiment, step three specifically includes:

[0090] Based on the city's geospatial information data, we screen geographical entities with tourism, leisure, culture and entertainment functions, construct a set of interest nodes, and extract structured attributes such as geographical coordinates, category, opening time and popularity level for each interest node.

[0091] Based on road network data, pedestrian access data, and public transportation route maps, access paths in the city are extracted and analyzed to identify traffic paths connecting different interest nodes and obtain access path edges. The access path edges include the identifiers of the starting and ending interest nodes, path length, estimated travel time, and traffic mode type attribute information.

[0092] The path length and estimated travel time between interest nodes on the reachable path are weighted and calculated to generate initial edge weights.

[0093] Based on the interest nodes, access path edges, and initial edge weights, an original urban environment graph is constructed to represent the connection relationships between interest nodes in the city.

[0094] In this embodiment, the intention-environment bidirectional remapping network specifically includes an interest node enhancement module, an edge weight adjustment module, and a graph fusion module;

[0095] The user intent representation vector is input into the interest node enhancement module, and the cosine similarity between the user intent representation vector and the attribute vector of each interest node in the original urban environment map is calculated to generate the importance weight of the interest node.

[0096] Based on the importance weight of interest nodes, the set of interest nodes in the original urban environment map is filtered out, and interest nodes whose importance weight is less than the preset importance weight threshold are removed. Also, based on the connection relationship between interest nodes, associated access path edges are removed simultaneously.

[0097] The edge weight adjustment module extracts the importance weights of the starting interest nodes at both ends of each access path edge, and performs a weighted fusion based on the importance weights of the starting interest nodes at both ends and the initial edge weight to calculate the personalized edge weight of each access path edge.

[0098] ;

[0099] in, Represents personalized edge weights. Represents the original edge weights. Representing interest nodes Importance weight, Representing interest nodes Importance weight, and The value represents the weighting coefficient, where the personalized edge weight represents the travel cost. The smaller the value, the more the access path matches the user's interests and preferences, and the stronger the user's willingness to travel.

[0100] The initial edge weight of each reachable path edge is replaced with a personalized edge weight, and reachable path edges with personalized edge weights greater than a preset edge weight threshold are subjected to connection pruning operations.

[0101] A personalized urban environment map is generated based on the retained interest nodes, the updated access path edges, and the personalized edge weights.

[0102] In this embodiment, step five specifically includes:

[0103] Based on the set of interest nodes and access path edge information in the personalized urban environment map, the importance weights of the interest nodes are sorted in descending order, and the top N interest nodes are selected to construct the initial candidate node set.

[0104] Each interest node is selected sequentially from the set of initial candidate nodes as the starting point of the path, the travel path is initialized, and time budget constraints and maximum number of nodes constraints are set for the construction of the travel path.

[0105] Using the last interest node in the current travel path as a reference node (when the travel path just begins to expand, the starting candidate node is the starting point of the path and also the last interest node), traverse the adjacent interest nodes that are connected by accessible path edges, sort them based on the personalized edge weights of the accessible path edges, and select the interest node with the smallest personalized edge weight that does not appear in the current travel path as the next expansion node.

[0106] The cumulative travel time is updated in real time after each node expansion. When the cumulative travel time exceeds the preset time budget, or the number of visited nodes exceeds the preset maximum number of nodes, the path expansion process is terminated.

[0107] Complete the tourism path construction process for all initial candidate nodes in sequence, generate multiple tourism paths that meet the time budget constraint and the maximum number of nodes constraint, and construct a tourism path set.

[0108] In this embodiment, step six specifically includes:

[0109] The population is initialized as a set of travel routes, and the fitness of each travel route is calculated as follows:

[0110] The importance weights of all nodes of interest in the travel route are summed and multiplied by a preset first weighting factor to obtain the first calculated value; the personalized edge weights of all accessible path edges in the travel route are summed and multiplied by a preset second weighting factor to obtain the second calculated value; the fitness of the travel route is obtained by subtracting the second calculated value from the first calculated value.

[0111] In each round of tourism route evolution and update, the tourism route with the lowest fitness value is selected from the current tourism route set as the target route, and a tourism route is randomly selected from the top 20% of tourism routes with fitness values ​​from high to low as the parent route.

[0112] Based on the category attribute of the interest nodes, consecutive interest nodes of the same category in the parent path are divided into a functional segment to construct a sequence of functional segments of the tourism path;

[0113] The heterogeneous reconstruction egg exchange mechanism of the improved cuckoo search algorithm specifically includes: randomly selecting one or more functional segments from the functional segment sequence and replacing functional segments with the same category attributes in the target path;

[0114] If no functional segment of the same category exists in the target path, the selected functional segment is inserted into the target path in the original access order to generate a candidate path.

[0115] The candidate paths are checked for connectivity integrity. If there are two adjacent interest nodes that are not connected, the shortest path connecting the two adjacent interest nodes is found in the personalized urban environment map, and an intermediate interest node is inserted to connect the paths.

[0116] The fitness of the candidate paths after connection integrity verification is recalculated. If the fitness of the candidate path is higher than that of the original target path, the target path is replaced by the candidate path and enters the next generation of the population; otherwise, the original target path is kept unchanged.

[0117] In the AI-based tourism route planning method of this invention, an improved cuckoo search algorithm is proposed to address the limitations of existing cuckoo search algorithms in optimizing complex path combinations. This algorithm introduces a heterogeneous reconstruction egg exchange mechanism to achieve leapfrog evolution updates, thereby significantly improving the diversity of path generation and personalized matching effects.

[0118] The core of the improved Cuckoo Search algorithm lies in dividing the travel route into multiple functional segments. Each functional segment consists of interest nodes with consistent category attributes and continuous visits from the parent path, representing a certain type of travel preference scenario (such as cultural tours, food tours, and nature tours). After the division, the start and end position indices of each functional segment in the parent path are recorded as a structural reference. In each iteration, the path with the lowest fitness is selected from the current travel route set as the target path, and a parent path is randomly selected from the top 20% of paths in terms of fitness. At the same time, a heterogeneous reconstruction egg exchange mechanism is introduced to randomly select one or more functional segments from the functional segment sequence of the parent path to replace the functional segments with consistent category attributes in the target path. If there are no functional segments of the same category in the target path, the selected functional segments are inserted into the target path in the original access order.

[0119] The key to achieving leapfrog evolutionary updates lies in the fact that selected functional segments do not need to maintain their original structural positions or orders. Instead, by matching categories, they are directly replaced with segments of the same category in the target path, or new functional segments are inserted sequentially at appropriate positions in the target path. This breaks through the limitations of traditional path point fine-tuning and achieves structural-level leap mutations. This segment-level reconstruction operation gives path evolution a greater capacity for structural mutations, avoids getting trapped in local optima, and improves global exploration efficiency.

[0120] To ensure the feasibility of the reconstructed path, this invention checks the connectivity integrity of adjacent interest nodes in the new path to see if there is a connection. If there is no connection, the shortest reachable path is automatically found in the personalized urban environment map, and necessary intermediate nodes are inserted to complete the path. The fitness of the candidate paths is recalculated, and those with better fitness are included in the next generation of the population, gradually approaching the optimal recommended path.

[0121] In this embodiment, dividing consecutive interest nodes of the same category in the parent path into a functional segment specifically means:

[0122] The parent path is scanned according to the category attributes of the interest nodes. Two or more interest nodes that are visited consecutively and have the same category attributes are divided into a functional segment. The position index of the start and end interest nodes in the parent path is recorded for each functional segment. This is used by the heterogeneous reconstruction egg exchange mechanism to find matching segments and determine the insertion position in the target path.

[0123] In this embodiment, the preset evolution termination condition includes at least one of the following:

[0124] Scenario 1: The current evolutionary generation has reached the preset maximum generation threshold;

[0125] Scenario 2: The maximum fitness improvement in the set of tourism routes across several consecutive generations is less than the set convergence threshold;

[0126] Scenario 3: The fitness standard deviation of all travel routes in the current population is less than the set stability threshold;

[0127] During the tourism route optimization process, the preset evolution termination conditions are continuously monitored. If the preset evolution termination conditions are met (i.e., any one or more of the above situations occur), the evolution process is immediately terminated, the final generation is output as the optimized tourism route set, and the tourism route with the highest fitness is selected as the current recommended route.

[0128] refer to Figure 2 An artificial intelligence-based tourism route planning system includes the following modules:

[0129] The multimodal data acquisition module is used to collect users' multimodal input data and construct multimodal datasets;

[0130] The user intent modeling module is used to input the multimodal dataset into the Perceiver IO model for encoding and to generate user intent representation vectors.

[0131] The urban environment map construction module is used to construct the original urban environment map, which consists of interest nodes, reachable path edges, and initial edge weights.

[0132] The map personalization adjustment module is used to input the user intent representation vector into the intent-environment bidirectional remapping network, calculate the importance weight of interest nodes and the personalized edge weight of access paths, adjust the original urban environment map, and generate a personalized urban environment map.

[0133] The initial path generation module is used to generate multiple tourist routes based on a personalized urban environment map, according to the set time budget constraints and maximum node number constraints, and to construct a set of tourist routes.

[0134] The path optimization module is used to optimize the set of tourist paths based on the improved rhododendron search algorithm. The improved rhododendron search algorithm includes a heterogeneous reconstruction egg exchange mechanism, which divides the parent path into functional segments and performs structural-level reconstruction in each round of evolution to achieve leapfrog evolution updates, while performing connection integrity verification and fitness calculation.

[0135] An evolution control module is used to set and detect preset evolution termination conditions. If the preset evolution termination conditions are met, the evolution process is stopped.

[0136] The route recommendation module is used to select the most adaptive tourist route from the final optimized tourist route set after the evolution process is completed, and output it as the current recommended route.

[0137] Example 1:

[0138] To verify the feasibility of this invention in practice, it was applied to a tourism route planning platform. A simulated user environment was constructed and system modules were deployed. The planning effect of personalized tourism routes for users was tested and compared in detail.

[0139] In this embodiment, the platform first collects a set of multimodal input data from test users, including interest keywords, uploaded images, historical collection records, and click behavior logs. After unified modeling using the Perceiver IO model to obtain the user intent vector, it is input into the intent-environment bidirectional remapping network to adjust the importance weights of each interest node and the personalized edge weights of the access paths in the original urban environment map, thereby generating a personalized urban environment map that is highly correlated with the user's intent.

[0140] The system generates an initial set of travel routes based on the set time budget and the maximum number of visits to interest nodes. Each route consists of interest nodes in the order of visits and satisfies connectivity and travel cost constraints. Then the system enters the route optimization stage, the core of which is the improved cuckoo search algorithm based on the heterogeneous reconstruction egg exchange mechanism proposed in this invention.

[0141] Traditional path optimization methods mostly employ fixed-granularity node replacement or whole-path fine-tuning strategies, which suffer from problems such as local convergence, narrow search space, and lack of structural leaps. To address these issues, this invention proposes for the first time to divide the path into semantically continuous functional segments, each composed of consecutive interest nodes of the same category, and then perform structural-level leap-based reconstruction based on these segments.

[0142] In each round of evolution, the system selects the tourism path with the lowest fitness from the current path set as the target path, and randomly selects a parent path from the top 20% of paths in terms of fitness. After performing a category scan on the parent path, a sequence of functional segments is constructed, and 1 to 3 functional segments are randomly selected to execute a heterogeneous reconstruction egg exchange mechanism: if a functional segment of the same category exists in the target path, it is directly replaced; otherwise, the selected functional segment is inserted into the target path to form a new candidate path.

[0143] This mechanism enables leapfrog reorganization of paths at the structural level, overcoming the limitations of traditional algorithms that rely on point-by-point replacement and local fine-tuning, and significantly improving population diversity and convergence efficiency. After candidate paths are constructed, the system also checks path connectivity. If disconnected node pairs exist, the system uses a personalized urban environment map to find the shortest reachable path and inserts intermediate nodes to ensure path integrity.

[0144] Finally, the candidate paths are re-evaluated and compared with the original paths: if the fitness is better, the candidate path is replaced and introduced into the next generation of the population; otherwise, the original path is kept unchanged. The system executes this process in each round of evolution until the final preset evolution termination condition is met.

[0145] In the experimental comparison, we selected the traditional genetic algorithm (GA), the standard cuckoo search algorithm (CSA), and the improved cuckoo search algorithm (ICSA) of this invention for path optimization and compared their performance in different aspects.

[0146] Table 1. Comparison of the performance of three optimization algorithms in tourism route planning tasks.

[0147] As can be clearly seen from the data in Table 1 above, the improved Cuckoo Search Algorithm (ICSA algorithm) proposed in this invention performs excellently in several key performance indicators, outperforming the compared GA algorithm and the standard CSA algorithm. In terms of average path fitness, the ICSA algorithm reaches 91.4, significantly higher than the GA algorithm's 83.2 and the standard CSA algorithm's 86.7, indicating that this method can generate higher-quality personalized travel routes that are more closely matched to user interests and path constraints.

[0148] The number of convergence algebras is an important indicator of the efficiency of optimization algorithms. ICSA only requires 18 generations to reach the convergence criterion, while the GA algorithm requires 35 generations and the standard CSA algorithm requires 28 generations. This clearly demonstrates that the ICSA algorithm significantly improves the evolutionary efficiency of the search process through functional segment partitioning and heterogeneous reconstruction exchange mechanisms. In terms of optimization time, the ICSA algorithm maintains a time of 3.9 seconds, which is slightly higher than CSA's 3.6 seconds, but much lower than GA's 4.1 seconds. Combined with other indicators, it can be seen that a good balance is achieved between time consumption and performance.

[0149] In terms of point-of-interest (POI) coverage, the ICSA algorithm achieved 88.2% coverage, a significant improvement compared to GA's 76.5% and CSA's 81.4%. This means the route better covers the areas of interest to users, enhancing the richness of the route's content and the travel value. Regarding user satisfaction ratings, the ICSA algorithm also achieved a high score of 9.0, significantly outperforming GA's 7.2 and the standard CSA's 8.1, further validating that the improved evolutionary mechanism brings a significant increase in user satisfaction in actual user experience.

[0150] Overall, the method of this invention, by introducing leapfrog evolution and structure-level reconstruction operations, not only significantly outperforms traditional algorithms in path optimization quality, but also demonstrates comprehensive performance improvements in optimization efficiency, interest point matching degree, and user satisfaction, verifying its feasibility and advancement in intelligent tourism path planning tasks.

[0151] Furthermore, to further demonstrate the leapfrog evolutionary capability of the heterogeneous reconstructed egg exchange mechanism, we recorded the path diversity and fitness fluctuations of each generation of the population. Compared to other algorithms, this invention can quickly differentiate fitness levels in early iterations, indicating that it possesses strong global exploration and population differentiation capabilities.

[0152] Table 2. Statistics on path fitness and population diversity during optimization.

[0153] As can be seen from the data listed in Table 2 above, the improved Cuckoo Search Algorithm (ICSA) exhibits stronger fitness convergence ability and population path diversity maintenance ability during evolutionary path optimization. Firstly, the fitness fluctuation value of the ICSA algorithm rapidly decreases from 4.5 in generation 4 to 0.9 in generation 16, a significantly larger decrease than that of the GA algorithm (from 1.4 to 0.2) and the standard CSA algorithm (from 2.1 to 0.3). This indicates that during evolution, ICSA can more effectively compress the fluctuation space of invalid path solutions and accelerate convergence to a better solution. Especially after generation 8, ICSA still maintains a high fitness change amplitude, indicating that its optimization potential remains strong in the mid-to-late stages.

[0154] Furthermore, regarding the path diversity index, although the fitness fluctuations of the ICSA algorithm gradually converged during the evolution process, its path diversity did not decay rapidly, remaining at 0.65 from 0.83 in the 4th generation to 0.65 in the 16th generation. This result reflects that by introducing a heterogeneous reconstruction egg exchange mechanism, the ICSA algorithm can maintain path diversity while steadily promoting the overall population to evolve towards a better region, effectively avoiding premature convergence.

[0155] This embodiment improves the traditional rhododendron search algorithm by introducing a heterogeneous reconstruction egg exchange mechanism, significantly enhancing its optimization capabilities in personalized tourism route planning tasks. This mechanism divides the parent path into functional segments based on the category of interest nodes, achieving a leapfrog evolutionary update based on category matching during path replacement. This overcomes the slow convergence and local optima traps caused by local replacements in traditional evolutionary strategies. The improved algorithm balances the overall diversity of path structure and the trend of fitness improvement in each iteration, effectively enhancing the global search capability and local optimization accuracy of the population. Simultaneously, it integrates personalized urban environmental maps to achieve path connection repair, ensuring the structural integrity and smoothness of the generated paths. This not only improves the adaptability and coverage of tourism routes but also considers user preferences and urban accessibility, providing strong technical support for achieving more intelligent and personalized tourism services.

[0156] 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 tourism route planning method based on artificial intelligence, characterized in that, The steps include: Step 1: Collect multimodal input data from users and construct a multimodal dataset; Step 2: Input the multimodal dataset into the Perceiver IO model for encoding to obtain the user intent representation vector; Step 3: Construct the original urban environment map, which consists of interest nodes, reachable path edges, and initial edge weights; Step 4: Input the user intent representation vector into the intent-environment bidirectional remapping network, calculate the importance weights of interest nodes and personalized edge weights, adjust the original urban environment map, and generate a personalized urban environment map. Step 5: Based on the personalized urban environment map, generate multiple tourist routes according to the set time budget constraints and maximum number of nodes constraints. The tourist routes are composed of interest nodes arranged in the order of visit, and construct a set of tourist routes. Step Six: Optimize the set of tourist routes based on the improved cuckoo search algorithm. The improvement of the improved cuckoo search algorithm is that it introduces a heterogeneous reconstruction egg exchange mechanism. In each round of tourist route replacement, the tourist route is divided into several functional segments to achieve leapfrog evolution and update. Step 7: When the preset evolution termination condition is met, end the evolution update, generate an optimized set of travel routes, and select the travel route with the highest fitness as the current recommended route.

2. The method for planning tourist routes based on artificial intelligence according to claim 1, characterized in that, The multimodal input data specifically includes text information, image information, user history behavior records, and interest preference information.

3. The artificial intelligence-based tourism route planning method according to claim 1, characterized in that, Step two specifically involves: The text information is segmented into words, and words are converted into text token embedding vectors through word vector encoding. The image information is divided into multiple image blocks, and each image block is flattened and convolutionally encoded to obtain the image token embedding vector. The user's historical behavior records and interest preference information are structured and each behavior record is represented as a multi-field data structure including behavior timestamp, interest tag, access object type and interaction operation type. The structured fields are mapped to a vector representation of a set dimension through a multilayer perceptron coding network to obtain the behavior token embedding vector. The text token embedding vector, image token embedding vector, and behavior token embedding vector are concatenated to form a unified token sequence, and modality type encoding and location information encoding are added to each token to form the model input tensor; The model input tensor is input into the cross-attention encoding module of the Perceiver IO model. Cross-modal associations are established among all input tokens through the attention mechanism. Multimodal token information is integrated using global attention weighting to generate latent representations. The latent representation is input into a stacked self-attention module for multi-layer representation updates. The updated latent representation is then mapped to an output token sequence through a reverse cross-attention module. The output token sequence is then pooled to obtain the user intent representation vector.

4. The artificial intelligence-based tourism route planning method according to claim 1, characterized in that, Step three specifically involves: Based on the city's geospatial information data, we screen geographical entities with tourism, leisure, culture and entertainment functions, construct a set of interest nodes, and extract structured attributes such as geographical coordinates, category, opening time and popularity level for each interest node. Based on road network data, pedestrian access data, and public transportation route maps, access paths in the city are extracted and analyzed to identify traffic paths connecting different interest nodes and obtain access path edges. The access path edges include the identifiers of the starting and ending interest nodes, path length, estimated travel time, and traffic mode type attribute information. The path length and estimated travel time between interest nodes on the reachable path are weighted and calculated to generate initial edge weights. Based on the interest nodes, access path edges, and initial edge weights, an original urban environment graph is constructed to represent the connection relationships between interest nodes in the city.

5. The artificial intelligence-based tourism route planning method according to claim 1, characterized in that, The intent-environment bidirectional remapping network specifically includes an interest node enhancement module, an edge weight adjustment module, and a graph fusion module; The user intent representation vector is input into the interest node enhancement module, and the cosine similarity between the user intent representation vector and the attribute vector of each interest node in the original urban environment map is calculated to generate the importance weight of the interest node. Based on the importance weight of interest nodes, the set of interest nodes in the original urban environment map is filtered out, and interest nodes whose importance weight is less than the preset importance weight threshold are removed. Also, based on the connection relationship between interest nodes, associated access path edges are removed simultaneously. The edge weight adjustment module extracts the importance weights of the starting interest nodes at both ends of each access path edge, and performs a weighted fusion based on the importance weights of the starting interest nodes at both ends and the initial edge weight to calculate the personalized edge weight of each access path edge. The initial edge weight of each reachable path edge is replaced with a personalized edge weight, and reachable path edges with personalized edge weights greater than a preset edge weight threshold are subjected to connection pruning operations. A personalized urban environment map is generated based on the retained interest nodes, the updated access path edges, and the personalized edge weights.

6. The method for planning tourist routes based on artificial intelligence according to claim 1, characterized in that, Step five specifically involves: Based on the set of interest nodes and access path edge information in the personalized urban environment map, the importance weights of the interest nodes are sorted in descending order, and the top N interest nodes are selected to construct the initial candidate node set. Each interest node is selected sequentially from the set of initial candidate nodes as the starting point of the path, the travel path is initialized, and time budget constraints and maximum number of nodes constraints are set for the construction of the travel path. Using the last interest node in the current travel path as the reference node, traverse the adjacent interest nodes that are connected by accessible path edges, sort them based on the personalized edge weights of the accessible path edges, and select the interest node with the smallest personalized edge weight that does not appear in the current travel path as the next expansion node. The cumulative travel time is updated in real time after each node expansion. When the cumulative travel time exceeds the preset time budget, or the number of visited nodes exceeds the preset maximum number of nodes, the path expansion process is terminated. Complete the tourism path construction process for all initial candidate nodes in sequence, generate multiple tourism paths that meet the time budget constraint and the maximum number of nodes constraint, and construct a tourism path set.

7. The artificial intelligence-based tourism route planning method according to claim 1, characterized in that, Step six specifically involves: The population is initialized as a set of travel routes, and the fitness of each travel route is calculated as follows: The importance weights of all nodes of interest in the travel route are summed and multiplied by a preset first weighting factor to obtain the first calculated value; the personalized edge weights of all accessible path edges in the travel route are summed and multiplied by a preset second weighting factor to obtain the second calculated value; the fitness of the travel route is obtained by subtracting the second calculated value from the first calculated value. In each round of tourism route evolution and update, the tourism route with the lowest fitness value is selected from the current tourism route set as the target route, and a tourism route is randomly selected from the top 20% of tourism routes with fitness values ​​from high to low as the parent route. Based on the category attribute of the interest nodes, consecutive interest nodes of the same category in the parent path are divided into a functional segment to construct a sequence of functional segments of the tourism path; The heterogeneous reconstruction egg exchange mechanism of the improved cuckoo search algorithm specifically includes: randomly selecting one or more functional segments from the functional segment sequence and replacing functional segments with the same category attributes in the target path; If no functional segment of the same category exists in the target path, the selected functional segment is inserted into the target path in the original access order to generate a candidate path. The candidate paths are checked for connectivity integrity. If there are two adjacent interest nodes that are not connected, the shortest path connecting the two adjacent interest nodes is found in the personalized urban environment map, and an intermediate interest node is inserted to connect the paths. The fitness of the candidate paths after connection integrity verification is recalculated. If the fitness of the candidate path is higher than that of the original target path, the target path is replaced by the candidate path and it enters the next generation of the population. Otherwise, the original target path remains unchanged.

8. The artificial intelligence-based tourism route planning method according to claim 7, characterized in that, The step of dividing consecutive interest nodes of the same category in the parent path into a functional segment is as follows: The parent path is scanned according to the category attributes of the interest nodes. Two or more interest nodes that are visited consecutively and have the same category attributes are divided into a functional segment. The position index of the start and end interest nodes in the parent path is recorded for each functional segment. This is used by the heterogeneous reconstruction egg exchange mechanism to find matching segments and determine the insertion position in the target path.

9. The method for planning tourist routes based on artificial intelligence according to claim 1, characterized in that, The preset evolution termination condition includes at least one of the following: Scenario 1: The current evolutionary generation has reached the preset maximum generation threshold; Scenario 2: The maximum fitness improvement in the set of tourism routes across several consecutive generations is less than the set convergence threshold; Scenario 3: The fitness standard deviation of all travel routes in the current population is less than the set stability threshold; During the tourism route optimization process, the preset evolution termination condition is continuously monitored. If the preset evolution termination condition is met, the evolution process is terminated immediately, the final generation is output as the optimized tourism route set, and the tourism route with the highest fitness is selected as the current recommended route.

10. An artificial intelligence-based tourism route planning system, executing the artificial intelligence-based tourism route planning method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The multimodal data acquisition module is used to collect users' multimodal input data and construct multimodal datasets; The user intent modeling module is used to input the multimodal dataset into the Perceiver IO model for encoding and to generate user intent representation vectors. The urban environment map construction module is used to construct the original urban environment map, which consists of interest nodes, reachable path edges, and initial edge weights. The map personalization adjustment module is used to input the user intent representation vector into the intent-environment bidirectional remapping network, calculate the importance weight of interest nodes and the personalized edge weight of access paths, adjust the original urban environment map, and generate a personalized urban environment map. The initial path generation module is used to generate multiple tourist paths based on a personalized urban environment map, according to the set time budget constraints and maximum node number constraints, and to construct a set of tourist paths. The path optimization module is used to optimize the set of tourist paths based on the improved rhododendron search algorithm. The improved rhododendron search algorithm includes a heterogeneous reconstruction egg exchange mechanism, which divides the parent path into functional segments and performs structural-level reconstruction in each round of evolution to achieve leapfrog evolution updates, while performing connection integrity verification and fitness calculation. An evolution control module is used to set and detect preset evolution termination conditions. If the preset evolution termination conditions are met, the evolution process is stopped. The route recommendation module is used to select the most adaptive tourist route from the final optimized tourist route set after the evolution process is completed, and output it as the current recommended route.

Images