Tourism recommendation method and system based on multi-modal data

Abstract

The invention belongs to the field of tourism recommendation, and provides a tourism recommendation method and system based on multi-modal data, and the method comprises the steps: constructing a multi-modal tourism causality cognitive map; natural interaction information input by a user is collected, and tourism intention variables and implicit preference features of the user are extracted through an intention recognition module; performing forward causal reasoning in the causal cognitive map based on the tourism intention variable, determining a target scenic spot sequence meeting the intention of the user, and performing path feasibility evaluation; according to the target scenic spot sequence, executing an anti-fact reasoning operation in a cognitive map, identifying adverse factors and reasons thereof in a potential path, and performing interpretation enhancement and risk avoidance correction on a recommended path; and a recommendation result is output, the result comprises a recommendation sequence of multiple paths, and each path is attached with user intention interpretation, a causal path diagram, a risk analysis diagram and an accessibility weight score.

Description

Technical Field

[0001] The present invention belongs to the field of tourism recommendation, and in particular relates to a tourism recommendation method and system based on multimodal data. Background Art

[0002] With the development of mobile internet, smart devices, and social media, travel recommendation systems have become an essential component of online travel platforms and digital travel services. Existing travel recommendation methods primarily rely on collaborative filtering, tag matching, and content retrieval techniques. These methods rely on user historical behavior or explicit ratings, recommending tourist attractions or routes through similarity calculations. While these methods have proven effective in some scenarios, they still suffer from significant shortcomings in the following areas:

[0003] Traditional methods mostly use structured tags or historical clicks as input, making it difficult to process unstructured information such as user text descriptions, image preferences, and voice interactions, resulting in an inaccurate understanding of the user's true intentions. Existing models mostly use black-box algorithms, and the process of generating recommendation results cannot be traced, making it difficult to explain to users "why the recommendation is made", which affects user trust and acceptance. Faced with complex environmental conditions such as plateau topography, weather changes, and road regulations, existing models have difficulty identifying and dynamically adjusting risks at the path level, which can easily lead to unfeasible recommendations, poor user experience, and even safety hazards. Many systems can only recognize users' explicit clicks or tags, ignoring implicit preference information that users have not expressed but can be obtained through data mining such as behavioral patterns, voice tone, and image style. Recommendations lack deep personalization.

[0004] Therefore, there is an urgent need for a new tourism recommendation method that integrates multimodal perception and can realize a closed-loop reasoning process from user intention to route execution. It is especially suitable for special scenarios such as plateau areas with complex natural environments and changeable traffic conditions (such as the Sichuan-Tibet Highway). Summary of the Invention

[0005] In order to solve the problems in the prior art, the present invention provides a travel recommendation method based on multimodal data, comprising the following steps:

[0006] Constructing a multimodal tourism causal cognitive map, wherein the multimodal tourism causal cognitive map comprises a plurality of nodes and multiple types of directed edges, wherein the nodes include user status nodes, tourism destination nodes, environmental status nodes, traffic event nodes, and scenic spot nodes, and the directed edges are used to represent causal dependencies between the nodes;

[0007] Collect natural interaction information input by users and extract users' travel intention variables and implicit preference features through the intention recognition module;

[0008] Based on the travel intention variable, forward causal reasoning is performed in the causal cognitive map to determine a target attraction sequence that meets the user's intention, and a path feasibility assessment is performed in combination with the current geographical location, traffic conditions, and environmental condition variables;

[0009] Based on the target attraction sequence, counterfactual reasoning operations are performed in the cognitive map to identify unfavorable factors and their causes in potential routes, and to enhance explanations and modify risk avoidance for the recommended routes;

[0010] Output recommendation results, including a recommended sequence of multiple paths, each path accompanied by a user intent explanation, a causal path diagram, a risk analysis diagram, and a reachability weight score.

[0011] Furthermore, collecting the natural interaction information input by the user specifically includes the following steps:

[0012] Collect multi-modal input information including text, voice, image, trajectory behavior, and physiological status;

[0013] The above-mentioned multiple modal input information is respectively input into the corresponding feature extraction module for vectorized encoding. The natural language processing model is used to extract semantic features for text, and the image convolutional neural network is used to extract visual features for images. The speech is transcribed into text and then processed uniformly.

[0014] The encoding results of the multiple modal input information are fused to form a unified feature vector and input into the intention recognition module to extract the user's current travel intention variables and implicit preference features to drive the subsequent causal reasoning process.

[0015] Furthermore, the process of performing forward causal reasoning in the causal cognitive map includes:

[0016] Starting from the user's travel intention variable, the intention node is located in the causal cognitive map and directed propagation is performed to its downstream connected attraction nodes;

[0017] During the propagation process, the inference score is calculated for each attraction node based on the path structure and edge weight information between the intent node and the attraction node in the graph structure. The edge weight is obtained based on the statistical results of user historical behavior, expert rules, or graph learning model training.

[0018] The candidate scenic spot nodes are sorted according to the scores and screened based on geographical distance, traffic accessibility and environmental status variables, and finally a preliminary target scenic spot sequence is generated as the candidate recommended route.

[0019] Furthermore, the process of performing counterfactual reasoning operations in the cognitive map includes the following steps:

[0020] For the target attraction sequence generated by forward reasoning, multiple counterfactual path versions are constructed in sequence. The counterfactual paths are implemented by replacing some attraction nodes, path edges, or external variables in the original path.

[0021] Input the replaced counterfactual path into the causal cognitive map, re-execute reasoning and evaluation, and compare the differences between the original path and the counterfactual path in terms of intention matching, accessibility, safety, and user satisfaction indicators;

[0022] If the original path is significantly inferior to the counterfactual path in some dimensions, the relevant nodes or paths in the path are marked as potential unfavorable factors, and their causal sources are determined through structural backtracking;

[0023] Based on the analysis results, the original path is modified to avoid risks, including removing high-risk nodes, adjusting the access path, replacing inaccessible attractions, and outputting explanatory information to explain the reasons for the adjustment and the optimization strategy.

[0024] Furthermore, the process of outputting recommendation results includes:

[0025] The resulting multiple paths are structured and output, each path includes multiple tourist attraction nodes, which are arranged in time or space order to form a complete travel recommendation sequence;

[0026] Add user intention explanation information to each path to explain how the path meets the user's travel motivation and preference characteristics. The explanation information is generated based on the causal path between the intention node and the attraction node;

[0027] Output a causal path diagram associated with each path, which is used to graphically display the generation logic of the path, including the causal relationship between user status, intention, environmental conditions and scenic spots;

[0028] A route risk analysis diagram is provided, marking high-risk sections, impassable areas, or areas where users have poor physiological adaptability, and providing avoidance suggestions;

[0029] Calculate and output a route accessibility score that takes into account factors such as geographical distance, traffic conditions, weather conditions, the availability of scenic spots, and the user's physical condition, so that users can make route selection decisions.

[0030] Another aspect of the present invention provides a tourism recommendation system based on multimodal data, comprising the following modules:

[0031] A map construction module is used to construct a multimodal tourism causal cognitive map, wherein the multimodal tourism causal cognitive map includes multiple nodes and multiple types of directed edges. The nodes include user status nodes, tourism destination nodes, environmental status nodes, traffic event nodes, and scenic spot nodes. The directed edges are used to represent the causal dependency relationship between nodes.

[0032] The intention recognition module is used to collect natural interaction information input by users and extract the user's travel intention variables and implicit preference features from them;

[0033] a forward reasoning module for performing forward causal reasoning in the causal cognitive map based on the travel intention variable, determining a target attraction sequence that satisfies the user's intention, and performing a path feasibility assessment based on current geographic location, traffic conditions, and environmental condition variables;

[0034] a counterfactual reasoning module, configured to perform counterfactual reasoning operations in the causal cognitive map based on the target scenic spot sequence, identify unfavorable factors and their causes in potential paths, and perform explanation enhancement and risk avoidance correction on the recommended path;

[0035] The path output module is used to output recommendation results, which include a recommendation sequence of multiple paths, each path is accompanied by a user intention explanation, a causal path diagram, a risk analysis diagram and a reachability weight score.

[0036] Furthermore, the intent recognition module specifically includes:

[0037] Multimodal data acquisition unit, used to collect multiple modal input information including text, voice, image, trajectory behavior, and physiological status;

[0038] The feature extraction unit is used to input multiple modal input information into the corresponding feature extraction module for vectorized encoding. The text information is processed through the natural language processing model to extract semantic features, the image information is processed through the image convolutional neural network to extract visual features, and the speech information is transcribed into text and then processed uniformly.

[0039] The feature fusion unit is used to fuse the above multimodal features to form a unified user feature vector, and to extract the user's travel intention variables and implicit preference features for the forward reasoning module to call.

[0040] Furthermore, the forward reasoning module specifically includes:

[0041] The intention node positioning unit is used to locate the intention node in the causal cognitive map with the travel intention variable as the starting point;

[0042] The reasoning propagation unit is used to perform directed propagation from the intention node to the downstream connected scenic spot nodes, and calculate the reasoning score of each scenic spot node based on the path structure and edge weight information between the intention node and the scenic spot node;

[0043] The scenic spot screening unit is used to sort the scenic spot nodes with high scores, and screen them based on geographical location, traffic accessibility and environmental status variables, and output the target scenic spot sequence.

[0044] Furthermore, the counterfactual reasoning module specifically includes:

[0045] An alternative path construction unit is used to construct multiple counterfactual path versions by replacing some scenic spots nodes, path edges or external variables in the recommended path;

[0046] The simulation evaluation unit is used to input the replaced counterfactual path into the causal cognitive map to re-execute reasoning and evaluation, and compare the differences between the original path and the counterfactual path in terms of intention matching, accessibility, safety and satisfaction;

[0047] The unfavorable factor identification unit is used to mark nodes or edges in the original path that perform worse than the counterfactual path as potential unfavorable factors, and identify their causal path sources through structural backtracking;

[0048] The path correction unit is used to adjust the original path to avoid risks based on the identification results and generate explanatory information to explain the reasons for the path modification and the optimization effect.

[0049] Furthermore, the path output module specifically includes:

[0050] The path sorting unit is used to output multiple recommended paths in a structured manner. Each path consists of multiple tourist attraction nodes, which are arranged in time or space order.

[0051] An intention interpretation generation unit, configured to generate a description of the correspondence between the path and the user's intention based on the causal path from the intention node to the attraction node;

[0052] A causal path diagram generation unit is used to graphically display the causal chain on which the route recommendation is based, including the causal relationship between user status, travel intention, environmental conditions and scenic spot nodes;

[0053] The risk analysis unit is used to output a risk analysis map corresponding to each route, mark high-risk sections, impassable areas or plateau risk areas on the route, and provide avoidance suggestions;

[0054] The accessibility scoring unit is used to calculate the accessibility score of each path based on multi-dimensional factors, including but not limited to geographical distance, road conditions, weather, user status, and attraction availability.

[0055] This paper proposes a tourism recommendation method and system based on multimodal data. By leveraging multimodal causal cognitive graphs, user intent recognition, forward causal reasoning, and counterfactual path optimization, this method achieves deep understanding, adaptive generation, and interpretable output of tourism recommendations. This method has the following beneficial effects:

[0056] By integrating multimodal inputs such as text, voice, images, and behavioral trajectories, a multi-layer neural network model is constructed to accurately extract users' travel intention variables and implicit preference characteristics, thereby achieving more personalized and proactive recommendations.

[0057] The travel process is innovatively modeled as a causal graph structure that includes factors such as user status, intention, environment, transportation and attractions. The forward reasoning mechanism is used to replace the traditional similarity matching, making the recommendation process have causal logic, semantic interpretability and dynamic adjustability.

[0058] By simulating alternatives and evaluating hypothetical scenarios for recommended routes, we can identify potential high-risk or inaccessible factors, and timely adjust or replace route nodes to ensure the feasibility and safety of recommended routes in complex environments. This is particularly suitable for traffic risk areas such as the Sichuan-Tibet Plateau.

[0059] The method supports multi-path recommendation output. Each path is accompanied by a user intention matching description, a causal path diagram, a risk analysis diagram, and a reachability score, which comprehensively presents the reasons and background for the recommendation, helping users understand, compare, and make independent decisions.

[0060] At the same time, it comprehensively considers external environmental variables such as real-time weather, road closures, altitude, and health status to achieve dynamic adjustment and real-time adaptation of recommended routes, significantly improving the practicality and intelligence level in complex geographical environments (such as the Sichuan-Tibet Plateau and self-driving routes).

[0061] In summary, the present invention overcomes the shortcomings of existing technologies in multimodal understanding, semantic expression, path stability and plateau risk adaptability, and has good prospects for promotion and application and commercial value. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] Figure 1 is a flow chart of the method of the present invention;

[0064] Figure 2 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0065] Below, the invention is preferably described with reference to the accompanying drawings and specific embodiments.

[0066] This embodiment solves the above problem through the following steps:

[0067] In one embodiment, reference Figure 1 The present invention provides a tourism recommendation method based on multimodal data, which belongs to the technical field of intelligent information processing and tourism travel services. It specifically involves joint modeling and causal reasoning of multimodal data such as images, text, voice, location information, and user behavior. By constructing a tourism causal cognitive model, it can realize dynamic recognition of users' personalized tourism intentions, deep matching of scenic spot semantic structures, and comprehensive analysis of path feasibility and risk, thereby providing users with explainable, accessible, and context-adaptive multi-path tourism recommendation services.

[0068] Specifically, the method comprises the following steps:

[0069] Step S10: construct a multimodal tourism causal cognitive map, wherein the multimodal tourism causal cognitive map includes multiple nodes and multiple types of directed edges. The nodes include user status nodes, tourism destination nodes, environmental status nodes, traffic event nodes, and scenic spot nodes. The directed edges are used to represent the causal dependency relationship between nodes.

[0070] In this embodiment of the present invention, step S10 is used to construct a multimodal tourism causal cognitive map, which serves as the core data structure and knowledge representation model for subsequent tourism intention reasoning and route recommendation. The specific implementation steps are as follows:

[0071] S101: Define node collection type

[0072] Construct a node set V in the causal cognitive graph, where the node types include at least:

[0073] User status node: used to represent the static and dynamic status information of the user during the travel process. The static status includes gender, age, budget level, travel mode, etc., and the dynamic status includes physiological parameters (such as heart rate, oxygen saturation), emotional state (such as anxiety, fatigue), current location, etc. The user's status node can be input by the user, selected through given options, or obtained through the user's social media, wearable devices, etc.

[0074] Travel purpose node: used to indicate the user's expressed or implied travel intention, such as "appreciating natural scenery", "participating in ethnic festivals", "healing and relaxing", "photography check-in", "self-driving challenge", etc.

[0075] Environmental status node: used to represent external natural and social environmental factors related to the tourism process, including real-time weather, altitude, air oxygen content, light intensity, scenic spot opening status, holiday status, etc.

[0076] Traffic event node: used to indicate the status or abnormality of the tourism transportation process, such as road congestion, landslide warning, road closure, construction information, traffic accidents, etc.

[0077] Scenic spot nodes are used to represent candidate tourist resources in the target tourist area, including natural landscapes, intangible cultural heritage projects, historical sites, service facilities, etc. The nodes are associated with multimodal data features, such as scenic spot images, comment text, label attributes, geographic location, accessibility, etc.

[0078] S102: Establishing directed edge sets and causal dependencies

[0079] Construct a causal relationship graph edge set E between nodes, each edge is a directed edge v i →v j , used to represent the source node v i Changes will affect the target node v j The state of the causal influence, including:

[0080] User status node → travel purpose node, for example: the altitude sickness index increases → the travel purpose changes from "high-intensity exploration" to "short-distance leisure".

[0081] Environmental status node → scenic spot node, for example: increased probability of rainfall → outdoor open scenic spots are not recommended.

[0082] Traffic event node → scenic spot accessibility node, for example: road collapse → scenic spot "Medog Village" is inaccessible.

[0083] Tourism destination node → attraction node, for example: "National Cultural Experience" → recommend "Wenchuan Qiang Village" and "Batang National Square".

[0084] The causal path from user state node → intention node → attraction node.

[0085] The causal strength of edges can be assigned through historical sample learning, rule system definition or manual annotation, and supports probabilistic graph structures (such as Bayesian network form) or symbolic logic graph representation.

[0086] S103: Fusion of multimodal features into node representation

[0087] For each of the above nodes, we extract their corresponding multimodal features and construct a node embedding representation vector. The features include:

[0088] Image features; text features; location information (latitude and longitude, altitude, map fragment); user trajectory behavior (trajectory graph vector, dwell time); sensor data (such as oxygen saturation, device acceleration); scene labels (such as "red culture", "Tibetan folk customs", "extreme driving").

[0089] S104: Form an initial causal cognitive map and store it in the database

[0090] Organize the nodes and edges into a graph structure G causal=(V,E,A), where A represents the multimodal attribute set of nodes and edges, forming a multimodal tourism causal cognitive graph, which is stored in a graph database (such as Neo4j) or used for subsequent graph neural network modeling (using.

[0091] This step formalizes the tourism recommendation problem into a causal graph structure modeling problem, uses graph nodes and edges to represent the reasoning path between "user-environment-intention-resource-behavior", and transforms the traditional correlation matching method into a "structured causal reasoning" method, so that the system has causal interpretability, adaptability and semantic expression capabilities.

[0092] Step S20: Collect the natural interaction information input by the user, and extract the user's travel intention variables and implicit preference features through the intention recognition module.

[0093] Step S20 is used to achieve semantic bridging between the information actively or passively expressed by the user and the system's understanding. The core goal is to identify the user's travel intention category and implicit preference tendency from natural interaction input, and provide semantic constraints for subsequent causal reasoning and recommended path generation.

[0094] The specific implementation steps include:

[0095] S201: Collecting Natural Interaction Information

[0096] Collect natural interaction information input by users through user terminal devices (such as smartphones, car terminals, wearable devices, etc.), including but not limited to:

[0097] Text input: such as travel purpose, comments, search keywords, and tag selection filled in the App.

[0098] Voice input: such as the user's intentions or feelings expressed through a voice assistant or in-vehicle system.

[0099] Image input: such as uploaded travel photos, selfies, and screenshot collections.

[0100] Behavioral trajectory: such as historical visited attractions, stay time, click records, and trajectory paths.

[0101] Physiological / environmental signals (optional): such as heart rate, oxygen saturation, temperature and other status data collected by wearable devices.

[0102] The above information is organized into an input vector sequence for analysis by the intent recognition module.

[0103] S202: Preprocessing and feature extraction

[0104] Perform data cleaning, modality conversion, and vectorization processing on the multimodal natural interaction information:

[0105] Text processing, including the use of methods such as word segmentation, denoising, and part-of-speech tagging, combined with pre-trained language models such as BERT and RoBERTa to generate sentence vectors or keyword vectors.

[0106] Speech processing, including using an automatic speech recognition module to transcribe it into text and then perform NLP processing, or directly using speech emotion recognition to extract emotion vectors.

[0107] Image processing, including using CLIP or ResNet to extract image embedding features and identify tourism themes contained in the image (such as snow-capped mountains, folk customs, scenery, and roads).

[0108] Behavioral data processing, including trajectory vectorization, time frequency statistics, and spatial distribution heat map generation.

[0109] Normalization of physiological / environmental data, including for inferring user status (such as "fatigue", "anxiety", "excitement").

[0110] S203: Intent Recognition and Implicit Preference Extraction

[0111] The above multimodal vector is input into the intent recognition module to extract the following two core semantic objects:

[0112] Travel intention variables:

[0113] Indicates the travel goal that the user currently hopes to achieve. It is a high-level semantic variable and includes but is not limited to the following types:

[0114] Natural scenery appreciation

[0115] Cultural immersion experience

[0116] Photography check-in

[0117] Challenge the self-driving route

[0118] Plateau healing and relaxation

[0119] Food Exploration

[0120] Red Research

[0121] The intent variable can be predicted using a classification model (such as Softmax), or a multi-label output method (such as Sigmoid+BCE loss function) can be used to predict multiple concurrent intents.

[0122] In a specific implementation method, intent recognition is achieved through the following process:

[0123] First, a large number of data samples with manually labeled travel intentions are collected. The samples include users' multimodal input information, such as text descriptions, image content, voice input, or historical behavior. At the same time, each sample is associated with one or more real travel intention labels.

[0124] Preprocess and extract features from the data of each modality. For example, use a pre-trained model to extract sentence vectors from text and visual feature vectors from images, and concatenate or fuse multimodal features as model input.

[0125] Build a neural network model consisting of an input layer, several hidden layers, and an output layer. The dimensions of the output layer correspond to the number of intent labels. If each sample corresponds to only one intent label, use the Softmax function to output a single-class probability. If a sample may contain multiple intent labels, use the Sigmoid function to output the independent probability of each label.

[0126] The error between the predicted result and the true label is calculated using the cross-entropy loss function (for single label) or the binary cross-entropy loss function (for multi-label), and the model parameters are iteratively updated through optimization algorithms such as backpropagation and gradient descent.

[0127] Repeat the training for multiple cycles until the accuracy or F1 value of the model on the validation set meets the expected requirements.

[0128] In actual use, it receives natural interactive information currently input by the user, such as text input "I want to see Tibetan culture", uploaded travel photos or voice descriptions.

[0129] The input is processed through the same feature extraction as during training to generate a vector representation in a unified format, which is then input into the trained model.

[0130] The model outputs the predicted probability of each travel intention, and the system judges the user's current travel intention based on the maximum probability (single label) or the category exceeding the set threshold (multi-label).

[0131] The identified intent variables are used as semantic inputs to participate in the subsequent recommendation reasoning and path generation process.

[0132] Implicit preference characteristics:

[0133] Indicates preferences that are not directly expressed by the user but are inferred from their behavior / images / comments, etc., including:

[0134] Landscape preference type (snow scene, forest, Tibetan village, etc.)

[0135] Transportation preference (driving, public transportation)

[0136] Path length tolerance (fast pace or slow experience)

[0137] Security sensitivity

[0138] Willingness to pay and consumption budget

[0139] Cultural proximity (acceptance of national culture)

[0140] Implicit preference features can be used to learn latent space representation vectors through variational autoencoders (VAE) or contrastive learning methods.

[0141] In a specific implementation method, implicit preference feature extraction is achieved through the following process:

[0142] Collect a large number of user data samples, each sample contains user behavior information (such as click records, travel trajectories, comment content, stay time, etc.) and corresponding user static attributes (such as age, gender, travel mode, etc.).

[0143] The above data is converted into vector form and fed into the encoder part of the VAE model as input. This part compresses and maps the high-dimensional data into a low-dimensional latent space (hidden space) through a neural network, and outputs a mean vector and a variance vector.

[0144] According to the principle of variational inference, a latent vector is sampled from this latent distribution as the user's implicit preference representation.

[0145] This vector is input into the decoder part, which attempts to reconstruct the original input data. The training goal is to minimize the reconstruction error and KL divergence loss (used to make the potential distribution close to the standard normal distribution), thereby approximating the user's true preference distribution.

[0146] Through multiple rounds of iterative training, a model is obtained that can compress user behavior into a stable latent vector representation.

[0147] In actual use, a portion of the current user's behavioral information or static portrait is obtained as input.

[0148] The input is fed into the trained VAE encoder to obtain the corresponding preference latent space vector, which represents the abstract expression of the user's personalized preferences.

[0149] This preference vector is used in subsequent recommendation models, such as calculating similarity with scenic spot feature vectors and inputting it into graph neural networks.

[0150] S204: Semantic Output Standardization

[0151] Finally, the system outputs the above extraction results uniformly as follows:

[0152] Explicit intent tags, such as "Photography + Tibetan cultural experience"

[0153] Implicit preference vector, such as P u =[0.8,0.2,0.6,0.1,0.9], indicating high acceptance of plateau scenic spots, strong safety requirements, and medium budget, etc.

[0154] This step addresses the challenge of traditional recommendation algorithms accurately understanding unstructured expressions of user intent by introducing artificial intelligence (AI) tools such as natural language processing, image recognition, speech understanding, and behavior modeling. Through multimodal deep encoding, expression normalization, and hierarchical intent-preference modeling, we extract clear recommendation semantics from ambiguous and incomplete user behavior. The core concept is to transform "user input" into "reasonable semantic variables."

[0155] Step S30: Based on the travel intention variable, forward causal reasoning is performed in the causal cognitive map to determine a target attraction sequence that meets the user's intention, and a path feasibility assessment is performed in combination with the current geographical location, traffic conditions, and environmental condition variables.

[0156] Step S30 is used to use the user's travel intention variable as the starting point of reasoning on the basis of the constructed multimodal tourism causal cognitive map, screen the candidate attraction sequence that matches the user's intention through the forward causal reasoning mechanism, and comprehensively evaluate the accessibility and path rationality of each attraction in combination with factors such as the user's real-time location information, road traffic conditions, and natural environment conditions, thereby outputting a set of structured and executable travel recommendation candidate paths.

[0157] The specific implementation steps include:

[0158] S301: Locating user intention nodes in the causal cognitive graph

[0159] The travel intention variable identified in step S20 is used as the initial node. The travel destination node associated with this intention is located in the causal cognitive graph as the starting point for causal reasoning. Each intention node is connected to multiple attraction nodes via a "travel destination → attraction" causal edge, indicating that the attraction can satisfy a certain type of travel motivation.

[0160] For example, the downstream attractions intended to be connected by the "cultural experience" include nodes such as the "Tibetan Village Museum" and the "Qiang Nationality Intangible Cultural Heritage Training Center".

[0161] S302: Perform forward causal reasoning

[0162] Based on graph reasoning algorithms (such as graph traversal, graph attention network GAT or Bayesian network reasoning), starting from the intention node, the signal is propagated downstream along the directed edge, the causal graph structure is traversed layer by layer, the causal scores are accumulated, and a set of candidate attractions is generated.

[0163] The reasoning process can define the following rules:

[0164] The score of each scenic spot node is the weighted sum of the multiple paths it receives;

[0165] The weights of the corresponding edges are trained based on historical behavior data or labeled by expert knowledge;

[0166] If a probabilistic graphical model is introduced, maximum a posteriori path search is adopted.

[0167] The inference output is a sequence of candidate scenic spots with the highest scores, denoted as S = {s1, s2, ..., s n}

[0168] S303: Introduce real variable constraints to conduct path feasibility assessment

[0169] To improve the practicality and safety of recommended routes, we use the following three real-time variables to screen the feasibility of candidate attractions and reorder the routes:

[0170] Geographic location constraints: Calculate the path length and estimated time based on the user's current location and the geographic coordinates of the attractions, and consider the relative positions of consecutive attractions to form a recommended path with reasonable geographical order;

[0171] Traffic condition constraints: Introducing road status data (e.g., open / closed / under construction), marking high-altitude risk sections, and eliminating routes to scenic spots that are inaccessible or too risky;

[0172] Environmental condition constraints: including weather conditions, altitude, user health status (such as altitude sickness risk), and penalizing or excluding high-risk attractions.

[0173] Based on the above conditions, an accessibility score C is calculated for each scenic spot. i , combined with the inference score, calculate the final ranking score:

[0174] Score i =α·P i +β·C i

[0175] in:

[0176] P i Score forward reasoning;

[0177] C i Score accessibility assessment;

[0178] α, β are weight parameters.

[0179] S304: Output target scenic spot sequence

[0180] The sorted scenic spot set is output as the target path sequence matching the current user's intention for subsequent counterfactual reasoning (step S40) and recommendation presentation.

[0181] This step, based on a causal graph modeling approach, considers travel intention as the upstream "cause" and attraction behavior as the downstream "effect." Structured graph reasoning is used to find the semantic path between intention and resources. Real-world data is also incorporated to establish constraints, ensuring that recommendations meet both user needs and the physical feasibility of the path.

[0182] Forward causal reasoning embodies the active recommendation logic of starting from "what the user wants" and deducing "where to go", which is different from the traditional passive collaborative filtering method based on historical behavior.

[0183] For example, if a user is located near Xinduqiao, Ganzi Prefecture, Sichuan Province, and enters the intent of "Tibetan cultural experience + photographing cultural landscapes," the system executes the following process:

[0184] Locate intention nodes in the causal graph: “cultural experience” and “humanistic photography”;

[0185] Forward reasoning to identify relevant scenic spots: such as "Tagong Temple", "Daofu Folk Houses", and "Bamei Grassland Tibetan Architecture Complex";

[0186] The system accessed the road condition information and found that there was a landslide in the Daofu to Luhuo section, so the "Luhuo Ancient Town" scenic spot was excluded;

[0187] The user's current oxygen saturation is low, excluding high-altitude areas (such as the Gongga Mountain Plateau Trail);

[0188] Combining geographical order and traffic efficiency, the final recommended scenic spot sequence is:

[0189] “Tagong Temple → Bamei Tibetan Village → Xinduqiao Old Street → Folk Culture and Creative Store”

[0190] It also includes reasons for recommendation, estimated time, and road risk warnings.

[0191] Step S40 , performing counterfactual reasoning operations in the cognitive map based on the target scenic spot sequence, identifying unfavorable factors and their causes in the potential path, and performing explanation enhancement and risk avoidance correction on the recommended path.

[0192] Step S40 is used to introduce a counterfactual reasoning mechanism based on the generated target attraction sequence, simulate substitution and hypothesis evaluation of the recommended path, identify high-risk, unsuitable or low-satisfaction factors and their causal paths in the potential path, and optimize the recommendation results accordingly, thereby achieving improved robustness and personalized safety assurance of the recommended path.

[0193] The specific implementation steps include:

[0194] S401: Constructing counterfactual settings

[0195] For the target scenic spot sequence S={s1,s2,…,s n}, construct the counterfactual path settings one by one:

[0196] The counterfactual path is a replacement version of a node or edge in the recommended path, in the form of:

[0197] Replace a certain scenic spot in the current path i For another similar attraction s′ i ;

[0198] Replace a certain path e ij is the backup path e′ ij ;

[0199] Replace external variables (such as worsening weather and road closures) to perform scenario simulations.

[0200] Set the goal as: "If the current path is not recommended, is there a better / safer / more reasonable path?"

[0201] S402: Simulating Causal Propagation and Comparison

[0202] Input the constructed counterfactual path into the causal cognitive graph, re-execute the reasoning process, and calculate the scoring indicators under this path, such as:

[0203] Attraction matching degree (how well it matches the user’s intention);

[0204] accessibility score;

[0205] Environmental adaptability (such as avoiding high altitudes);

[0206] Safety risk factors (such as disaster history along the route);

[0207] User satisfaction simulation score (estimated based on the regression model of historical user feedback data).

[0208] Compare and analyze the original recommended path with all counterfactual paths to extract "which node or path is most likely to cause problems" and its "causal path source".

[0209] S403: Identify adverse factors and causal pathways

[0210] The part of the original path that is significantly worse than the counterfactual path is marked as a "potential disadvantage factor", such as:

[0211] Attraction s3 leads to lower satisfaction in the counterfactual simulation;

[0212] Edge 3,4Traffic efficiency is significantly improved after replacement;

[0213] Environment variable E w (such as weather deterioration) affecting the reliability of the recommended route.

[0214] The source path of unfavorable factors (such as the path chain consisting of weather node → road condition node → scenic spot node) is identified through graph structure backtracking method for explanation enhancement.

[0215] S404: Path correction and interpretation enhanced output

[0216] Based on the above analysis results, the original recommended path is modified as follows:

[0217] Eliminate unfavorable nodes;

[0218] Replace with an equivalent node but with lower risk and higher matching degree;

[0219] Output the recommended explanation text: "To avoid possible...", a certain attraction has been replaced with..."

[0220] The final recommendation results include the following information:

[0221] Reasons for recommending each pathway and explanation of the risks;

[0222] List of alternative paths and their advantages;

[0223] Description of potential adverse factors and correction strategies;

[0224] Path robustness score.

[0225] The core of this step is based on the principle of counterfactual reasoning, that is, without changing the user's intention, constructing a parallel hypothetical world of "If you don't take the current recommended path, what other routes can you take?" By simulating these possibilities, potential risks or suboptimal links in the existing path can be identified.

[0226] In terms of specific implementation, path replacement and result comparison are performed with the help of causal graph structure, which can be achieved by using causal influence estimation functions (such as influence functions, Shapley value approximation), simulation path scoring models (regression prediction, graph neural network scoring), and other methods.

[0227] A specific example:

[0228] The user is currently located in Litang County, Sichuan Province. The system recommends the following path:

[0229] “Litang Temple → Haizi Mountain Observation Deck → Batang Tibetan Cultural Center → Mangkang Salt Well Ancient Village”

[0230] Systematic construction of counterfactual paths:

[0231] Replace "Haizi Mountain Observation Deck" with "Changqingchun Kerr Temple";

[0232] Replace the "Mangkang Salt Well" direction with the "Daocheng Riverside Plank Road" direction;

[0233] The simulated weather variables are "sudden snowfall + strong wind".

[0234] System reasoning found that:

[0235] There is a recent landslide warning in the Mangkang direction, and the probability of altitude sickness is high;

[0236] If using alternative routes can maintain the cultural experience and reduce travel risks;

[0237] Explanation output: "Due to the high altitude and poor traffic conditions in the Mangkang section, the last section of the route has been rerouted to Daocheng. This is expected to reduce risk by 20% and increase satisfaction by 15%."

[0238] Finally, the system outputs the correction path and explanation, along with a graphical causal path diagram.

[0239] Step S50: output the recommendation results, which include a recommendation sequence of multiple paths, each path is accompanied by a user intention explanation, a causal path diagram, a risk analysis diagram, and a reachability weight score.

[0240] Step S50 is used to structure the output of the multiple recommended paths that are finally screened out after completing the forward causal reasoning (step S30) and counterfactual analysis and correction (step S40) of the target scenic spot, and generate user intention matching instructions associated with each path, causal links generated by the path, possible risk warning information, and a quantitative score of the overall accessibility of the path, so as to achieve personalized presentation of the recommended content, explainable support, and risk control.

[0241] Specific implementations include:

[0242] S501: Generate multiple recommended path sequences

[0243] Based on the results of steps S30 and S40, the system finally selects several path sequences that meet the user's intention, have good environmental adaptability, and are highly accessible. Each path consists of several scenic spots in a temporal and spatial order, expressed as:

[0244] P i ={s i1 ,s i2 ,...,s im},i=1,2,...,N

[0245] Where N represents the number of output paths, and each path is an alternative recommended route that meets different trade-off requirements (such as safest, shortest time, most cultural depth, etc.).

[0246] S502: Adding user intention explanation information

[0247] For each path, the system traces back to its source, extracts the corresponding relationship between it and the user's travel intention variable, and outputs the recommendation reason in natural language or graphical form, explaining how the path meets the user's specific needs. For example:

[0248] "Path P1 includes three Tibetan cultural sites and two photography viewing platforms, meeting the user's dual purpose of 'culture + photography'."

[0249] The explanatory information can be derived from the path analysis of “intention node → attraction node” in the causal graph.

[0250] S503: Generate causal path diagram

[0251] Based on the causal cognitive map, the causal chain relationship between each node in the path is extracted and a causal path diagram is drawn to show "why this path is generated" and "which factors jointly contribute to this recommendation", including but not limited to:

[0252] User intent node;

[0253] Environmental status nodes (such as weather, altitude);

[0254] Safety / traffic event nodes;

[0255] Scenic spot node.

[0256] The causal path diagram can be represented by a directed graph, where the path direction is visualized to reveal the recommendation logic.

[0257] S504: Output risk analysis diagram and risk warning

[0258] Combined with the counterfactual reasoning results, mark the potential high-risk factors in the path, including but not limited to:

[0259] High altitude attractions;

[0260] Road sections prone to geological disasters;

[0261] areas prone to congestion;

[0262] Areas with unknown road conditions;

[0263] Warning about the user's own health condition (such as high altitude sickness index).

[0264] Each path provides a risk analysis diagram with textual prompts based on actual variables, such as:

[0265] "Path P2 passes through Daocheng to Litang, with an average altitude of over 4,100 meters. Please be aware of altitude sickness."

[0266] S505: Calculate and add accessibility weight score

[0267] The system evaluates the accessibility of each path based on the following factors and calculates a weighted score (e.g., out of 100):

[0268] Geographical distance and duration;

[0269] Current traffic conditions;

[0270] Route continuity and closure;

[0271] weather and altitude conditions;

[0272] User health parameter adaptability;

[0273] Attractions are open.

[0274] The scoring results serve as auxiliary information for sorting and guiding selection, showing users the overall feasibility of the path.

[0275] S506: Output recommendation results in a structured manner

[0276] The data structure or interface information of the final output recommendation results includes:

[0277] Recommended path sequence;

[0278] User intent matching instructions;

[0279] Paths form a cause-effect graph;

[0280] Path risk analysis diagram and prompts;

[0281] accessibility score;

[0282] Alternative path suggestions (optional).

[0283] The output result can be directly used for mobile app interface display, vehicle navigation system access or linkage with the user dialogue system.

[0284] This step, based on the recommendation generation principles of "explainable results, traceable processes, and risk awareness," transforms the recommendation system from a traditional "content matching black box" into a comprehensive decision-making tool characterized by "causal transparency, risk avoidance, and dynamic optimization." Through a multi-path candidate mechanism, a causal graph generation mechanism, and a multi-dimensional scoring system, users can choose the path that suits them best, regardless of their preferences and circumstances.

[0285] A specific example:

[0286] The user is currently in Kangding, Ganzi Prefecture, and wants to experience ethnic culture and take highland scenery photos. The system recommends the following three routes:

[0287] Path P1 (robust):

[0288] Scenic spot sequence: Kangding Love Song Scenic Route → Tagong Temple → Bamei Tibetan Houses

[0289] Intent explanation: Culture and photography combined, moderate average altitude;

[0290] Accessibility score: 92 / 100;

[0291] Risk warning: There is a slight risk of altitude sickness, so it is recommended to bring an oxygen tank;

[0292] Causal path diagram: shows the path formed from intention node to attraction node.

[0293] Path P2 (Challenging):

[0294] Scenic spot sequence: Xinduqiao → Litang Grassland → Daocheng Yading

[0295] Intentional explanation: Emphasis on the impact of the landscape;

[0296] Accessibility score: 78 / 100;

[0297] Risk Warning: The altitude is over 4,100 meters, and there is a risk of altitude sickness;

[0298] Alternative suggestion: "Yading" can be replaced with "Gongga Mountain Observation Deck".

[0299] Path P3 (Cultural Depth):

[0300] Scenic spot sequence: Luding Bridge → Red Army Long March Memorial Hall → Jiaju Tibetan Village

[0301] Intentional explanation: Red + Tibetan culture in tandem;

[0302] Accessibility score: 88 / 100;

[0303] Risk warning: No major traffic risks, some road sections will experience speed reduction due to construction.

[0304] The system outputs the above results in a structured manner and provides users with selection through a graphical interface.

[0305] See also Figure 2 In another embodiment, the present invention further provides a tourism recommendation system based on multimodal data, comprising:

[0306] A map construction module is used to construct a multimodal tourism causal cognitive map, wherein the multimodal tourism causal cognitive map includes multiple nodes and multiple types of directed edges. The nodes include user status nodes, tourism destination nodes, environmental status nodes, traffic event nodes, and scenic spot nodes. The directed edges are used to represent the causal dependency relationship between nodes.

[0307] The intention recognition module is used to collect natural interaction information input by users and extract the user's travel intention variables and implicit preference features from them;

[0308] a forward reasoning module for performing forward causal reasoning in the causal cognitive map based on the travel intention variable, determining a target attraction sequence that satisfies the user's intention, and performing a path feasibility assessment based on current geographic location, traffic conditions, and environmental condition variables;

[0309] a counterfactual reasoning module, configured to perform counterfactual reasoning operations in the causal cognitive map based on the target scenic spot sequence, identify unfavorable factors and their causes in potential paths, and perform explanation enhancement and risk avoidance correction on the recommended path;

[0310] The path output module is used to output recommendation results, which include a recommendation sequence of multiple paths, each path is accompanied by a user intention explanation, a causal path diagram, a risk analysis diagram and a reachability weight score.

[0311] In one embodiment, the intent recognition module specifically includes:

[0312] Multimodal data acquisition unit, used to collect multiple modal input information including text, voice, image, trajectory behavior, and physiological status;

[0313] The feature extraction unit is used to input multiple modal input information into the corresponding feature extraction module for vectorized encoding. The text information is processed through the natural language processing model to extract semantic features, the image information is processed through the image convolutional neural network to extract visual features, and the speech information is transcribed into text and then processed uniformly.

[0314] The feature fusion unit is used to fuse the above multimodal features to form a unified user feature vector, and to extract the user's travel intention variables and implicit preference features for the forward reasoning module to call.

[0315] In one embodiment, the forward reasoning module specifically includes:

[0316] The intention node positioning unit is used to locate the intention node in the causal cognitive map with the travel intention variable as the starting point;

[0317] The reasoning propagation unit is used to perform directed propagation from the intention node to the downstream connected scenic spot nodes, and calculate the reasoning score of each scenic spot node based on the path structure and edge weight information between the intention node and the scenic spot node;

[0318] The scenic spot screening unit is used to sort the scenic spot nodes with high scores, and screen them based on geographical location, traffic accessibility and environmental status variables, and output the target scenic spot sequence.

[0319] In one embodiment, the counterfactual reasoning module specifically includes:

[0320] An alternative path construction unit is used to construct multiple counterfactual path versions by replacing some scenic spots nodes, path edges or external variables in the recommended path;

[0321] The simulation evaluation unit is used to input the replaced counterfactual path into the causal cognitive map to re-execute reasoning and evaluation, and compare the differences between the original path and the counterfactual path in terms of intention matching, accessibility, safety and satisfaction;

[0322] The unfavorable factor identification unit is used to mark nodes or edges in the original path that perform worse than the counterfactual path as potential unfavorable factors, and identify their causal path sources through structural backtracking;

[0323] The path correction unit is used to adjust the original path to avoid risks based on the identification results and generate explanatory information to explain the reasons for the path modification and the optimization effect.

[0324] In one embodiment, the path output module specifically includes:

[0325] The path sorting unit is used to output multiple recommended paths in a structured manner. Each path consists of multiple tourist attraction nodes, which are arranged in time or space order.

[0326] An intention interpretation generation unit, configured to generate a description of the correspondence between the path and the user's intention based on the causal path from the intention node to the attraction node;

[0327] A causal path diagram generation unit is used to graphically display the causal chain on which the route recommendation is based, including the causal relationship between user status, travel intention, environmental conditions and scenic spot nodes;

[0328] The risk analysis unit is used to output a risk analysis map corresponding to each route, mark high-risk sections, impassable areas or plateau risk areas on the route, and provide avoidance suggestions;

[0329] The accessibility scoring unit is used to calculate the accessibility score of each path based on multi-dimensional factors, including but not limited to geographical distance, road conditions, weather, user status, and attraction availability.

[0330] It should be noted that the explanation of the above-mentioned embodiment of the travel recommendation method based on multimodal data is also applicable to the device of the embodiment of the present application and will not be repeated here.

[0331] Those skilled in the art will appreciate that the various units and algorithm steps described in the embodiments disclosed herein can be implemented using a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0332] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0333] In the several embodiments provided in this application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of this application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), magnetic disk or optical disk, and other media that can store program code.

[0334] The above is only a specific embodiment of the present application. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in this application, which should be included in the scope of protection of this application. For some module structures that are not particularly clear in the present invention, the content recorded in the prior art shall prevail. The prior art mentioned in the above background technology section and the specific embodiment section of the present invention can be regarded as part of the present invention and is used to understand the meaning of some technical features or parameters.

Claims

1. A tourism recommendation method based on multimodal data, characterized in that: The method comprises the following steps: Constructing a multimodal tourism causal cognitive map, wherein the multimodal tourism causal cognitive map comprises a plurality of nodes and multiple types of directed edges, wherein the nodes include user status nodes, tourism destination nodes, environmental status nodes, traffic event nodes, and scenic spot nodes, and the directed edges are used to represent causal dependencies between the nodes; Collect natural interaction information input by users and extract users' travel intention variables and implicit preference features through the intention recognition module; Based on the travel intention variable, forward causal reasoning is performed in the causal cognitive map to determine a target attraction sequence that meets the user's intention, and a path feasibility assessment is performed in combination with the current geographical location, traffic conditions, and environmental condition variables; Based on the target attraction sequence, counterfactual reasoning operations are performed in the cognitive map to identify unfavorable factors and their causes in potential routes, and to enhance explanations and modify risk avoidance for the recommended routes; Output recommendation results, including a recommended sequence of multiple paths, each path accompanied by a user intent explanation, a causal path diagram, a risk analysis diagram, and a reachability weight score.

2. The tourism recommendation method based on multimodal data according to claim 1, characterized in that: The collecting of natural interaction information input by the user specifically includes the following steps: Collect multi-modal input information including text, voice, image, trajectory behavior, and physiological status; The above-mentioned multiple modal input information is respectively input into the corresponding feature extraction module for vectorized encoding. The natural language processing model is used to extract semantic features for text, and the image convolutional neural network is used to extract visual features for images. The speech is transcribed into text and then processed uniformly. The encoding results of the multiple modal input information are fused to form a unified feature vector and input into the intention recognition module to extract the user's current travel intention variables and implicit preference features to drive the subsequent causal reasoning process.

3. The tourism recommendation method based on multimodal data according to claim 1, characterized in that: The process of performing forward causal reasoning in the causal cognitive map includes: Starting from the user's travel intention variable, the intention node is located in the causal cognitive map and directed propagation is performed to its downstream connected attraction nodes; During the propagation process, the inference score is calculated for each attraction node based on the path structure and edge weight information between the intent node and the attraction node in the graph structure. The edge weight is obtained based on the statistical results of user historical behavior, expert rules, or graph learning model training. The candidate scenic spot nodes are sorted according to the scores and screened based on geographical distance, traffic accessibility and environmental status variables, and finally a preliminary target scenic spot sequence is generated as the candidate recommended route.

4. The tourism recommendation method based on multimodal data according to claim 1, characterized in that: The process of performing counterfactual reasoning operations in the cognitive map includes the following steps: For the target attraction sequence generated by forward reasoning, multiple counterfactual path versions are constructed in sequence. The counterfactual paths are implemented by replacing some attraction nodes, path edges, or external variables in the original path. Input the replaced counterfactual path into the causal cognitive map, re-execute reasoning and evaluation, and compare the differences between the original path and the counterfactual path in terms of intention matching, accessibility, safety, and user satisfaction indicators; If the original path is significantly inferior to the counterfactual path in some dimensions, the relevant nodes or paths in the path are marked as potential unfavorable factors, and their causal sources are determined through structural backtracking; Based on the analysis results, the original path is modified to avoid risks, including removing high-risk nodes, adjusting the access path, replacing inaccessible attractions, and outputting explanatory information to explain the reasons for the adjustment and the optimization strategy.

5. The tourism recommendation method based on multimodal data according to claim 1, characterized in that: The process of outputting the recommendation results includes: The resulting multiple paths are structured and output, each path includes multiple tourist attraction nodes, which are arranged in time or space order to form a complete travel recommendation sequence; Add user intention explanation information to each path to explain how the path meets the user's travel motivation and preference characteristics. The explanation information is generated based on the causal path between the intention node and the attraction node; Output a causal path diagram associated with each path, which is used to graphically display the generation logic of the path, including the causal relationship between user status, intention, environmental conditions and scenic spots; A route risk analysis diagram is provided, marking high-risk sections, impassable areas, or areas where users have poor physiological adaptability, and providing avoidance suggestions; Calculate and output a route accessibility score that takes into account factors such as geographical distance, traffic conditions, weather conditions, the availability of scenic spots, and the user's physical condition, so that users can make route selection decisions.

6. A tourism recommendation system based on multimodal data, characterized in that: The system includes the following modules: A map construction module is used to construct a multimodal tourism causal cognitive map, wherein the multimodal tourism causal cognitive map includes multiple nodes and multiple types of directed edges. The nodes include user status nodes, tourism destination nodes, environmental status nodes, traffic event nodes, and scenic spot nodes. The directed edges are used to represent the causal dependency relationship between nodes. The intention recognition module is used to collect natural interaction information input by users and extract the user's travel intention variables and implicit preference features from them; a forward reasoning module for performing forward causal reasoning in the causal cognitive map based on the travel intention variable, determining a target attraction sequence that satisfies the user's intention, and performing a path feasibility assessment based on current geographic location, traffic conditions, and environmental condition variables; a counterfactual reasoning module, configured to perform counterfactual reasoning operations in the causal cognitive map based on the target scenic spot sequence, identify unfavorable factors and their causes in potential paths, and perform explanation enhancement and risk avoidance correction on the recommended path; The path output module is used to output recommendation results, which include a recommendation sequence of multiple paths, each path is accompanied by a user intention explanation, a causal path diagram, a risk analysis diagram and a reachability weight score.

7. The tourism recommendation system based on multimodal data according to claim 6, characterized in that: The intention recognition module specifically includes: Multimodal data acquisition unit, used to collect multiple modal input information including text, voice, image, trajectory behavior, and physiological status; The feature extraction unit is used to input multiple modal input information into the corresponding feature extraction module for vectorized encoding. The text information is processed through the natural language processing model to extract semantic features, the image information is processed through the image convolutional neural network to extract visual features, and the speech information is transcribed into text and then processed uniformly. The feature fusion unit is used to fuse the above multimodal features to form a unified user feature vector, and to extract the user's travel intention variables and implicit preference features for the forward reasoning module to call.

8. The tourism recommendation system based on multimodal data according to claim 6, characterized in that: The forward reasoning module specifically includes: The intention node positioning unit is used to locate the intention node in the causal cognitive map with the travel intention variable as the starting point; The reasoning propagation unit is used to perform directed propagation from the intention node to the downstream connected scenic spot nodes, and calculate the reasoning score of each scenic spot node based on the path structure and edge weight information between the intention node and the scenic spot node; The scenic spot screening unit is used to sort the scenic spot nodes with high scores, and screen them based on geographical location, traffic accessibility and environmental status variables, and output the target scenic spot sequence.

9. The tourism recommendation system based on multimodal data according to claim 6, characterized in that: The counterfactual reasoning module specifically includes: An alternative path construction unit is used to construct multiple counterfactual path versions by replacing some scenic spots nodes, path edges or external variables in the recommended path; The simulation evaluation unit is used to input the replaced counterfactual path into the causal cognitive map to re-execute reasoning and evaluation, and compare the differences between the original path and the counterfactual path in terms of intention matching, accessibility, safety and satisfaction; The unfavorable factor identification unit is used to mark nodes or edges in the original path that perform worse than the counterfactual path as potential unfavorable factors, and identify their causal path sources through structural backtracking; The path correction unit is used to adjust the original path to avoid risks based on the identification results and generate explanatory information to explain the reasons for the path modification and the optimization effect.

10. The tourism recommendation system based on multimodal data according to claim 6, characterized in that: The path output module specifically includes: The path sorting unit is used to output multiple recommended paths in a structured manner. Each path consists of multiple tourist attraction nodes, which are arranged in time or space order. An intention interpretation generation unit, configured to generate a description of the correspondence between the path and the user's intention based on the causal path from the intention node to the attraction node; A causal path diagram generation unit is used to graphically display the causal chain on which the route recommendation is based, including the causal relationship between user status, travel intention, environmental conditions and scenic spot nodes; The risk analysis unit is used to output a risk analysis map corresponding to each route, mark high-risk sections, impassable areas or plateau risk areas on the route, and provide avoidance suggestions; The accessibility scoring unit is used to calculate the accessibility score of each path based on multi-dimensional factors, including but not limited to geographical distance, road conditions, weather, user status, and attraction availability.

Images