Text feature-based travel activity information collection and analysis method and system

Abstract

The application relates to the technical field of intelligent processing of tourism information, and discloses a tourism activity information intelligent processing method and system based on text features. The tourism activity information intelligent processing method based on text features comprises the following steps: constructing a feature automatic labeling model through self-supervised learning, processing original text to obtain structured tourism activity data; generating a feature vector by using multi-level feature extraction and fusion technology; obtaining an activity information set after deduplication and integration by using self-adaptive threshold clustering and information fusion; combining migration learning and collaborative filtering to perform dynamic quality scoring and confidence evaluation; and finally generating a personalized tourism activity recommendation sequence by multi-objective optimization based on user demand, an activity correlation graph and a feature vector. The application realizes automatic collection, cross-platform deduplication, dynamic quality evaluation and accurate recommendation of tourism activity information, and improves recommendation accuracy and user satisfaction.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing technology for tourism information, and more specifically, to a method and system for collecting and analyzing tourism activity information based on text features. Background Technology

[0002] With the rapid development of the tourism market and the upgrading of consumption, tourists' demand for personalized and customized tourism services is increasing. Tourism activity information is scattered across various channels such as tourism platforms, social media, and official websites. This information exists in the form of unstructured text, including activity descriptions, user reviews, and travel guides. How to efficiently collect and accurately analyze tourism activity information from massive amounts of heterogeneous text and provide users with precise personalized recommendations has become an important issue in the field of tourism information services.

[0003] Existing methods for processing tourism activity information primarily rely on manual annotation and simple keyword matching techniques. Manual annotation requires a large number of domain experts to annotate the activity text, resulting in high costs, low efficiency, and difficulty in handling massive amounts of data. Keyword matching methods can only extract surface-level information from the text, failing to understand its deeper semantics and contextual relationships, leading to inaccurate feature extraction. Furthermore, descriptions of the same activity vary across different platforms; existing methods, using fixed thresholds for deduplication, struggle to adapt to the diverse feature distributions of different activity types, easily leading to misjudgments. In addition, existing quality assessment methods are mainly based on statistical aggregation, which cannot provide reliable quality scores for new or niche activities with sparse evaluation data, impacting recommendation effectiveness.

[0004] Therefore, there is a need for a technical solution that can automatically extract textual features of tourism activities, intelligently integrate cross-platform information, dynamically evaluate activity quality, and provide personalized recommendations, in order to solve the technical problems of existing technologies, such as feature annotation relying on manual methods, low accuracy of cross-platform deduplication, and incomplete quality assessment coverage. Summary of the Invention

[0005] This invention provides a method and system for collecting and analyzing tourism activity information based on text features, which solves the technical problems of feature annotation relying on manual methods, low accuracy of cross-platform deduplication, and incomplete quality assessment coverage in related technologies.

[0006] This invention provides a method for collecting and analyzing tourism activity information based on text features, comprising the following steps: S1. Collect the original tourism activity text dataset and use a self-supervised learning method to build an automatic feature annotation model to obtain tourism activity data; S2, acquire tourism activity data, and use a multi-level feature extraction framework and feature fusion method to obtain the feature vector of tourism activities; S3 receives the feature vectors of tourism activities and uses an adaptive threshold clustering algorithm and information fusion strategy to obtain a set of tourism activity information. S4. Based on the tourism activity information set, a method combining transfer learning and collaborative filtering is used to obtain activity information with dynamic quality scores and confidence levels. S5 receives activity information with dynamic quality scores and confidence levels, as well as feature vectors of tourism activities. It then uses a hierarchical sampling strategy to construct an activity association graph and performs graph neural network embedding learning to obtain the activity association graph and activity embedding representation. S6. Based on the user demand text, activity association graph, and feature vector of tourism activities, a multi-objective optimization algorithm is used to obtain a tourism activity sequence recommendation scheme.

[0007] In a preferred embodiment, S1 includes: Based on the original tourism activity text dataset, data cleaning and preprocessing operations were performed to obtain standardized text data; Based on standardized text data, a masked language model is constructed to mask some words in the text. By predicting the masked words, a self-supervised pre-training is performed to obtain a deep semantic representation of the text. Based on deep semantic representation of text, a sentence order prediction task is designed to learn the logical structure of activity descriptions. A contrastive learning task is designed to maximize the feature similarity of different descriptions of the same activity and minimize the feature similarity of different activity descriptions. The mask language model task, sentence order prediction task and contrastive learning task are jointly trained to obtain enhanced discriminative feature representations. Based on enhanced discriminative feature representation and a small amount of manually labeled seed data, a multi-dimensional feature labeling model is constructed, which includes a shared feature extraction layer and multiple task-specific classification layers. The model is then fine-tuned using transfer learning. The text of tourism activities is labeled with features based on a multi-dimensional feature automatic labeling model. The highest probability value of the categorized features is used as the confidence level. Labeling results with confidence levels lower than a preset threshold are marked as pending manual review, resulting in tourism activity data with multi-dimensional feature labels and confidence levels.

[0008] In a preferred embodiment, S2 includes: Based on large-scale tourism-related texts, domain pre-training was performed using masked language modeling tasks and next sentence prediction tasks to obtain a tourism domain pre-trained language model. Based on a pre-trained language model in the tourism field, a multi-level feature extraction framework is constructed, which includes a lexical encoder, a sentence encoder, and a document encoder, to obtain feature representations at the lexical, sentence, and document levels. Explicit features are extracted from multi-dimensional feature annotation results, and implicit features are extracted from multi-level feature representations. A gating fusion mechanism is designed to dynamically adjust the fusion weights of explicit and implicit features based on the confidence level of feature annotations, thereby obtaining a comprehensive feature vector.

[0009] In a preferred embodiment, S3 includes: Design a platform feature encoder, learn a feature transformation matrix for each platform, and train it through a contrastive learning method to map the feature vectors of different platforms to a unified semantic space; Design an adaptive threshold determination method based on local density awareness to calculate the local density of each activity in the feature space; A hierarchical clustering algorithm is used to select activity pairs with similarity exceeding an adaptive threshold from the similarity matrix and merge them until no activity pair has a similarity exceeding its adaptive threshold. Each activity information within a cluster is evaluated for quality and its overall weight is calculated. A weighted fusion method is then used to merge the activity information within the clusters to obtain a unified activity information after deduplication and integration.

[0010] In a preferred embodiment, S4 includes: A sentiment analysis model was used to classify the sentiment tendencies of user review texts, and key information extraction technology was used to extract quality-related features from the review texts. Tourism activities are grouped according to type and regional attributes. Activities with more than a preset threshold of evaluations are selected as source domain data. A quality assessment model is constructed and trained by learning the mapping relationship between evaluation features and quality scores to obtain the source domain quality assessment model. For activities with fewer than a preset threshold of evaluations, the parameters of the source domain model are used as initialization for fine-tuning to obtain a transfer learning score; neighbor activities are retrieved based on the cosine similarity of the comprehensive feature vector, and a weighted average method is used to infer the collaborative filtering score of the target activity; a time series analysis method is used to track the trend of quality changes to obtain a trend adjustment factor. The transfer learning score, collaborative filtering score, and trend adjustment factor are weighted and fused to obtain the dynamic quality score and confidence level.

[0011] In a preferred embodiment, S5 includes: The importance score of an activity is calculated by combining quality score, popularity, and timeliness. Based on the importance score, the activity is divided into multiple levels and incorporated into a relationship diagram. Calculate the thematic, geographical, temporal, and user behavior relationships between activities, and construct an activity relationship graph; A graph neural network model is constructed, and the feature vectors of the neighboring nodes of each active node are collected and aggregated to obtain the graph embedding representation of the activity.

[0012] In a preferred embodiment, S6 includes: A multi-level feature extraction framework is used to extract features from user requirement text, named entity recognition technology is used to identify key entities, and user requirement feature vectors are constructed. A combination of rule matching and semantic understanding is used to parse constraints and construct a set of constraints. A multi-dimensional weighted matching algorithm is used to calculate the matching degree of features in different dimensions and perform weighted aggregation to obtain the activity's demand matching degree score. Activities are filtered based on the set of constraints to obtain a set of candidate activities that meet the constraints.

[0013] 8. The method for collecting and analyzing tourism activity information based on text features according to claim 7, wherein step S6 further includes: Define a multi-objective optimization problem, with optimization objectives including maximizing demand matching degree, maximizing activity quality score, maximizing activity sequence coherence, and minimizing time and space costs; A genetic algorithm is used to solve a multi-objective optimization problem. The activity sequences are encoded as chromosomes. The chromosomes in the population are sorted using a non-dominated sorting method. Genetic operations such as selection, crossover, and mutation are performed to obtain a Pareto optimal set of activity sequences. Calculate a comprehensive score for each activity sequence in the Pareto optimal solution set based on user preference weights, and select the activity sequence with the highest comprehensive score as the final recommendation.

[0014] In a preferred embodiment, the text feature-based tourism activity information collection and analysis method further includes: S7, continuously optimizing the model using an active learning strategy based on newly collected tourism activity text and user feedback data; specifically including: The newly collected activity texts are labeled with features using an automatic feature labeling model. An uncertainty assessment method is used to quantify the model’s prediction uncertainty for each sample, and the sample with the highest uncertainty is selected as a high-value sample to be labeled. High-value unlabeled samples are submitted to domain experts for manual annotation and quality review to obtain high-quality labeled data. An incremental learning method is used to fine-tune the model by mixing new labeled data with some historical labeled data, resulting in an updated feature extraction model. User feedback data is collected, and the recommendation process is modeled as a Markov decision process. A deep Q-network method is used to optimize the recommendation strategy, resulting in an updated recommendation model.

[0015] In a preferred embodiment, the text-feature-based tourism activity information collection and analysis system is used to perform the steps in the above-described text-feature-based tourism activity information collection and analysis method, including: The feature annotation module is used to collect raw tourism activity text datasets and use a self-supervised learning method to build an automatic feature annotation model to obtain tourism activity data. The feature extraction module is used to acquire tourism activity data. It adopts a multi-level feature extraction framework and feature fusion method to obtain the feature vector of tourism activities. The information integration module receives feature vectors of tourism activities and uses an adaptive threshold clustering algorithm and information fusion strategy to obtain a set of tourism activity information. The quality assessment module is used to obtain activity information with dynamic quality scores and confidence levels by using a combination of transfer learning and collaborative filtering based on a set of tourism activity information. The graph embedding module is used to receive activity information with dynamic quality scores and confidence levels, as well as feature vectors of tourism activities. It uses a hierarchical sampling strategy to construct an activity association graph and performs graph neural network embedding learning to obtain the activity association graph and activity embedding representation. The recommendation generation module is used to generate a tourism activity sequence recommendation scheme based on user demand text, activity association graph and feature vector of tourism activities, using a multi-objective optimization algorithm.

[0016] The beneficial effects of this invention are as follows: A self-supervised learning approach is employed to construct an automatic feature annotation model. Through joint training of a masked language model, sentence order prediction, and contrastive learning tasks, a general semantic representation is learned from unlabeled data. Automatic annotation of multi-dimensional features can be achieved with only a small amount of seed data for fine-tuning, significantly reducing the cost of manual annotation and improving the efficiency and accuracy of feature annotation. Simultaneously, a gated fusion mechanism dynamically fuses explicit and implicit features, adaptively adjusting the fusion weights based on annotation confidence. This achieves complementary advantages between structured attribute information and deep semantic information, enhancing the quality of feature representation. A local density-aware adaptive threshold determination method was designed. This method dynamically adjusts the similarity judgment threshold based on the local density of activities in the feature space, using a higher threshold for activities with dense feature distributions and a lower threshold for those with sparse feature distributions. This effectively solves the problem that fixed thresholds are difficult to adapt to the differences in feature distributions across different activity types, improving the accuracy of deduplication across platforms. Simultaneously, a quality assessment method combining transfer learning and collaborative filtering is employed. For activities with sparse evaluation data, transfer learning leverages the evaluation capabilities of activities with sufficient evaluation data, while collaborative filtering draws on the quality information of similar activities, achieving comprehensive and dynamic quality assessment and providing reliable quality assurance for recommendations. Attached Figure Description

[0017] Figure 1 This is a flowchart of the tourism activity information collection and analysis method based on text features of the present invention; Figure 2 This is a flowchart of the text feature-based tourism activity information collection and analysis method of the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a method for collecting and analyzing tourism activity information based on text features, such as... Figures 1 to 2 As shown, it includes the following steps: S1. Collect the original tourism activity text dataset and use a self-supervised learning method to build an automatic feature annotation model to obtain tourism activity data; This step aims to address the problem of tourism activity feature annotation relying on large amounts of manually labeled data. It utilizes self-supervised learning methods to learn features using the text's own structural and semantic information, achieving automated feature annotation. Specifically, it includes the following steps: S11, based on the original tourism activity text dataset, uses data cleaning and preprocessing operations to obtain standardized text data; Raw text data of tourism activities is collected from multiple tourism platforms, and the data includes multiple text types such as activity titles, activity descriptions, activity details, user reviews, etc. For the collected raw text, data cleaning operations are performed to remove noise content such as HTML tags, special symbols, and advertising information. Word segmentation processing is performed on the text, and a word segmentation tool for Chinese text is used to segment continuous text into word sequences. Part-of-speech tagging is performed on the word segmentation results to identify different parts of speech such as nouns, verbs, adjectives, etc., so as to provide part-of-speech information for subsequent feature extraction. Stop-word removal processing is performed on the text to remove high-frequency words that contribute little to semantics such as "de", "le", "zai" and the like. Text length standardization processing is performed: context completion is performed for excessively short texts, and segmentation processing is performed for excessively long texts, so as to ensure that the length of each text fragment is within a reasonable range. Normalization processing is performed on time expressions, place expressions and numerical expressions in the text: time in different formats is unified into a standard format, place names are unified into standard geographical names, and numerical expressions are unified into Arabic numeral form. After the above preprocessing operations, a standardized text data set of tourism activities is obtained, and each piece of data in this data set includes information such as cleaned text content, word segmentation results, and part-of-speech tagging results.

[0020] S12, based on the standardized text data, self-supervised pre-training is performed using a masked language model task to obtain deep semantic representations of the text; A masked language model is constructed, which is based on the Transformer architecture and includes multi-layer self-attention mechanisms and feed-forward neural networks. For the standardized text data, part of the words in the text are randomly selected for masking processing, the masking ratio is set to 15%, and the masked words are replaced with special tokens. The masked text is input into the model, and the model captures the dependencies between words through the self-attention mechanism and learns context information. The training objective of the model is to predict the masked words, and the model parameters are optimized by minimizing the cross-entropy loss between the predicted words and the real words. During the training process, the model learns the deep semantic representation of the text, which includes not only the surface information of words, but also the semantic information of words in context. Large-scale text data of tourism activities is used for pre-training, and the scale of training data reaches the million level, covering tourism activity descriptions of different types, different regions and different styles. After pre-training is completed, the model can generate high-quality semantic representation vectors for any input text, and the vector captures the semantic features and context information of the text.

[0021] S13, based on the deep semantic representation of the text, a sentence order prediction task and a contrastive learning task are used to obtain an enhanced discriminative feature representation; Building upon pre-trained masked language models, a sentence order prediction task is designed as an auxiliary task. For activity description text containing multiple sentences, the sentence order is randomly shuffled to construct a scrambled text. This scrambled text is then input into the model, which must predict the correct sentence order. In this way, the model learns the logical structure of the activity description and the relationships between sentences, understanding how activity information is organized. The loss function for the sentence order prediction task uses ranking loss to measure the difference between the predicted and actual order.

[0022] A contrastive learning task is designed to further enhance the discriminative power of feature representations. For the same tourism activity, descriptive texts collected from different platforms or at different times, although expressed differently, describe the same activity and constitute positive sample pairs. For different tourism activities, their descriptive texts constitute negative sample pairs. The goal of contrastive learning is to maximize the feature similarity of positive sample pairs and minimize the feature similarity of negative sample pairs. A contrastive loss function is used to bring the feature vectors of positive sample pairs closer in semantic space, and to increase the distance between the feature vectors of negative sample pairs. Through contrastive learning, the model learns discriminative features that can distinguish different activities, while also being robust to different descriptions of the same activity.

[0023] By jointly training a masked language modeling task, a sentence order prediction task, and a contrastive learning task using a multi-task learning framework, the three tasks share the underlying feature extraction network. By jointly optimizing the loss functions of the three tasks, an enhanced discriminative feature representation is obtained. This feature representation not only contains the semantic information of the text but also the structural information of the activity descriptions and the discriminative information between activities.

[0024] S14, based on enhanced discriminative feature representation and a small amount of manually labeled seed data, is fine-tuned using transfer learning to obtain a multi-dimensional feature automatic labeling model; In one embodiment of the present invention, although self-supervised learning can learn general semantic representations from unlabeled data, accurate multi-dimensional feature annotation still requires a small amount of seed data annotated by domain experts to guide the model in learning specific annotation rules. A small amount of manually annotated seed data is prepared, with a scale of approximately 1000 to 5000 entries, significantly reducing annotation costs compared to the tens of thousands of annotation data required for traditional supervised learning. The seed data includes text on tourism activities and their corresponding multi-dimensional feature annotations. Feature dimensions include activity type, spatiotemporal attributes, participant attributes, experience attributes, cultural attributes, interactivity, educational significance, safety, difficulty level, and cost range. Activity types adopt a hierarchical classification system, including primary categories such as cultural experiences, outdoor adventures, leisure and entertainment, and educational study tours, as well as subcategories under each primary category. Spatiotemporal attributes include the activity's geographical location, suitable season, and activity duration. Participant attributes include suitable age range and suitable group types such as families, couples, and groups. Experience attributes include experience intensity and experience type such as spectator, participatory, and immersive experiences.

[0025] A multi-dimensional feature annotation model is constructed based on the enhanced feature representation obtained through self-supervised learning. This model employs a multi-task learning architecture, comprising a shared feature extraction layer and multiple task-specific classification layers. The shared feature extraction layer is initialized using pre-trained model parameters obtained through self-supervised learning. Each task-specific classification layer corresponds to a feature dimension and is responsible for the annotation prediction of that dimension. For categorical features such as activity type, a multi-classifier is used to output the probability distribution of each category; for numerical features such as difficulty level, a regressor is used to output the numerical prediction result.

[0026] A transfer learning approach is employed, using a small amount of seed data to fine-tune the model. During fine-tuning, the parameters of the shared feature extraction layer are updated slightly to retain the general semantic representation capabilities obtained through self-supervised learning, while the parameters of the task-specific classification layer are updated more significantly to learn the annotation patterns of specific feature dimensions. Cross-entropy loss is used to optimize the classification task, and mean squared error loss is used to optimize the regression task. The model parameters are updated using the backpropagation algorithm. After fine-tuning, a multi-dimensional feature automatic annotation model is obtained, capable of automatically annotating multi-dimensional features of any input tourism activity text.

[0027] S15, based on the multi-dimensional feature automatic annotation model, the tourism activity text is labeled with features and the confidence score is calculated to obtain tourism activity data with multi-dimensional feature annotation and confidence score; The text of the tourism activity to be labeled is input into a multi-dimensional feature automatic labeling model. The model predicts for each feature dimension. For categorical features, the model outputs the probability distribution of each category, selects the category with the highest probability as the predicted label, and uses the probability value of that category as the label's confidence level. For numerical features, the model outputs numerical prediction results. The confidence level of the prediction results is calculated by analyzing the model's prediction variance or using ensemble learning methods. The confidence level reflects the model's certainty about the labeling result; high confidence level indicates that the model is quite confident in the labeling, while low confidence level indicates that the model has uncertainty about the labeling.

[0028] For annotations with a confidence level below a preset threshold, they are marked for manual review and submitted to domain experts for manual annotation or correction. The preset threshold is set according to the importance of different feature dimensions; the threshold for key feature dimensions such as activity type is set to 0.8, and the threshold for secondary feature dimensions such as difficulty level is set to 0.6. The manually reviewed annotations are used as new high-quality training data and are periodically used for incremental model optimization, forming a closed-loop mechanism of model prediction - manual review - model optimization. For annotations with a confidence level above the threshold, they are directly adopted as feature annotations for the activity and enter the subsequent processing flow.

[0029] The annotation results are associated with the original text to construct structured tourism activity data. Each activity data entry includes the original text content, multi-dimensional feature annotations, the confidence level of each annotation, the data source platform, the collection time, and the review status. The review status indicates whether the data has passed automatic annotation, is awaiting manual review, or has completed manual review. This data is stored in the activity database to provide a data foundation for subsequent feature extraction and analysis.

[0030] S2, acquire tourism activity data, and use a multi-level feature extraction framework and feature fusion method to obtain the feature vector of tourism activities; This step aims to extract deep semantic features from tourism activity texts and fuse explicit structured feature annotations with implicit semantic features to generate a comprehensive feature representation; specifically, it includes the following steps: S21. Based on large-scale tourism-related texts, a domain-specific pre-training method is used to obtain a pre-trained language model for the tourism domain. We collected large-scale tourism-related text data from sources including travel guide websites, travel blogs, travel social media, official tourism websites, and user review platforms. We preprocessed the collected text data to construct a tourism-related vocabulary list, which includes general terms as well as specialized terms, place names, attraction names, event names, and other field-specific vocabulary.

[0031] A pre-trained language model is built based on the Transformer architecture, and pre-training is performed using masked language modeling and next-sentence prediction tasks. Large-scale tourism text data is used for pre-training, and a distributed training strategy is employed. After pre-training, the model learns language representations in the tourism domain, enabling it to understand tourism terminology, expression habits, and semantic features, providing high-quality semantic encoding capabilities for tourism activity texts.

[0032] S22, based on a pre-trained language model in the tourism field, adopts a multi-level coding structure to obtain feature representations at the lexical, sentence, and document levels; A multi-level feature extraction framework is constructed, which includes three levels: a lexical encoder, a sentence encoder, and a document encoder.

[0033] At the lexical level, dual encoding at the character level and word level is employed. For each word, a character-level convolutional neural network is used to extract character-level features, while word vectors are obtained from a pre-trained language model in the tourism domain. The two are then concatenated to obtain a comprehensive representation of the word.

[0034] At the sentence level, a bidirectional long short-term memory network is used to encode the lexical feature sequence, capturing the contextual semantics of the words through bidirectional processing. An attention mechanism is then used to weight and aggregate the lexical features to obtain the overall semantic vector of the sentence.

[0035] At the document level, the activity description text is treated as a document composed of multiple sentences. A Transformer structure and hierarchical attention mechanism are used to encode sentence features, and a self-attention mechanism is employed to capture the relationships between sentences, resulting in the overall semantic vector of the document.

[0036] Through a multi-level encoding structure, feature representations at the lexical, sentence, and document levels are obtained. These features characterize the semantic information of tourism activity texts at different granularities.

[0037] S23, based on multi-level feature representation and multi-dimensional feature annotation results, a gated fusion mechanism is adopted to obtain a comprehensive feature vector that integrates explicit and implicit features; From the multi-dimensional feature annotation results obtained in step S1, explicit features are extracted. Explicit features are structured annotation information, such as activity type, spatiotemporal attributes, and participant attributes. These categorical annotations are converted into one-hot encoded vectors, and numerical annotations are directly used as feature values ​​to construct explicit feature vectors. Explicit feature vectors are sparse, discrete feature representations that clearly reflect the attribute information of the activity.

[0038] Latent features are extracted from the multi-level feature representations obtained in step S22. Latent features are continuous vector representations learned from text semantics, containing deep semantic information and contextual relationships of the text. The feature representations at the lexical, sentence, and document levels are concatenated or weighted to obtain the latent feature vector. The latent feature vector is a dense, continuous feature representation that captures the semantic details and implicit information of the text.

[0039] A gating fusion mechanism is designed to dynamically adjust the fusion weights of explicit and implicit features. Since the explicit and implicit feature vectors may have different dimensions, they are first aligned in dimension. The explicit feature vector is transformed in dimension through a fully connected layer, mapping it to the same dimension space as the implicit feature vector, thus obtaining the aligned explicit feature representation.

[0040] The gating mechanism's input includes aligned explicit feature vectors, implicit feature vectors, and confidence vectors for feature annotations. The confidence vector contains the confidence values ​​for each feature dimension. The gating mechanism is implemented using a neural network. The network structure consists of two fully connected layers and an activation function. The input is a concatenated vector of explicit features, implicit features, and confidence values; the output is a scalar of fusion weights. These fusion weights are normalized to between zero and one using a sigmoid activation function, representing the relative importance of explicit and implicit features. For explicit features with high confidence, the gating network learns to assign larger fusion weights, making the fused features more dependent on explicit annotation information; for explicit features with low confidence, the gating network assigns smaller fusion weights, making the fused features more dependent on implicit semantic information.

[0041] Based on the fusion weights output by the gating mechanism, the aligned explicit and implicit features are weighted and fused. The fusion formula is: the fused feature vector equals the fusion weight multiplied by the aligned explicit feature vector plus one minus the fusion weight multiplied by the implicit feature vector. Since the two feature vectors are already dimensionally aligned, they can be directly weighted and summed. In this way, the fused features contain both explicit structured attribute information and implicit deep semantic information, achieving complementary advantages of the two types of features.

[0042] The fused features are input into a fully connected layer for feature dimension transformation and nonlinear mapping to obtain the final comprehensive feature vector. The dimension of the comprehensive feature vector is set to a fixed value, such as 512 or 1024, to facilitate subsequent similarity calculation and model processing. This comprehensive feature vector is a compact representation of tourism activities, containing multi-dimensional attribute information and deep semantic information of the activities, providing a high-quality feature foundation for subsequent information integration, quality assessment, and recommendation.

[0043] S3 receives the feature vectors of tourism activities and uses an adaptive threshold clustering algorithm and information fusion strategy to obtain a set of tourism activity information. This step aims to solve the problem of integrating and deduplicating heterogeneous information across platforms. Through an adaptive threshold clustering algorithm, it identifies descriptions of the same activity on different platforms, performs information fusion, and generates a unified activity representation. Specifically, it includes the following steps: S31, based on the text characteristics of different platforms, a platform feature encoder is used to perform feature space mapping to obtain feature vectors in a unified semantic space; The text description styles, levels of detail, and structuring levels differ across different travel platforms, leading to a shift in the semantic space distribution of text feature vectors for the same activity across different platforms. To eliminate these platform differences, a platform feature encoder is designed to learn a feature transformation matrix for each platform, mapping the platform's feature vectors to a unified semantic space.

[0044] The platform feature encoder is trained using a contrastive learning method. The training data includes descriptions of the same activity on different platforms as positive sample pairs, and descriptions of different activities as negative sample pairs. A triplet loss function is used for optimization, making positive sample pairs as close as possible in the mapped semantic space and negative sample pairs as far apart as possible.

[0045] The comprehensive feature vectors of tourism activities from different platforms are input into the corresponding platform feature encoders to obtain mapped feature vectors, which provide a basis for subsequent similarity calculation and clustering.

[0046] S32, based on feature vectors in a unified semantic space, adopts a local density-aware adaptive threshold determination method to obtain the similarity judgment threshold for each activity pair; Traditional deduplication methods use a fixed similarity threshold, but due to the differences in feature distribution among different activity types, a fixed threshold is difficult to adapt to all situations. For activity types with dense feature distribution, a fixed threshold may lead to different activities being misclassified as the same activity; for activity types with sparse feature distribution, a fixed threshold may lead to different descriptions of the same activity being misclassified as different activities.

[0047] An adaptive threshold determination method based on local density awareness is designed. For each pair of activities to be judged, the cosine similarity between the feature vectors of the two activities is calculated. The similarity value is between -1 and 1, with higher similarity indicating greater similarity between the two activities. Then, the local density of each activity in the feature space is calculated. The local density reflects the number of similar activities around that activity. The local density is calculated as follows: a neighborhood is defined in the feature space centered on the feature vector of the activity, and the number of other activities within the neighborhood is counted. The higher the number of other activities within the neighborhood, the higher the local density. The neighborhood is determined using a fixed radius or a fixed number of neighbors.

[0048] The similarity threshold is dynamically adjusted based on the local density of activity pairs. For activity pairs with high local density, a higher threshold is used because, in dense regions, highly similar activities may still be distinct activities, requiring a stricter judgment standard. For activity pairs with low local density, a lower threshold is used because, in sparse regions, less similar activities may be different descriptions of the same activity, requiring a more lenient judgment standard. The threshold adjustment formula is: the adaptive threshold equals the base threshold plus the local density adjustment factor multiplied by the normalized local density value. The base threshold is a preset fixed value, such as 0.75. The local density adjustment factor is a positive parameter, ranging from 0.1 to 0.3, optimized using training data. The normalized local density value is obtained by dividing the original local density by the maximum local density, ranging from zero to one. In this way, the higher the local density, the higher the adaptive threshold, ensuring a strict judgment standard in dense regions.

[0049] For each activity pair, its similarity is compared with an adaptive threshold. If the similarity is greater than the threshold, it is determined to be different descriptions of the same activity; if the similarity is less than or equal to the threshold, it is determined to be different activities. Through the adaptive threshold method, the system can flexibly adjust the judgment criteria according to the characteristic distribution of activities, thereby improving the accuracy of deduplication.

[0050] S33, based on an adaptive threshold, a hierarchical clustering algorithm is used to cluster activities to obtain activity clustering results; Hierarchical clustering algorithm is used to cluster all tourism activities, grouping similar activities into one category. Hierarchical clustering is a bottom-up clustering method. Initially, each activity is treated as an independent cluster, and then similar clusters are gradually merged until a stopping condition is met.

[0051] The clustering process is as follows: First, calculate the similarity between all activity pairs and their corresponding adaptive thresholds to construct a similarity matrix. Then, select the activity pairs with the highest similarity exceeding the adaptive threshold from the similarity matrix and merge these two activities into one cluster. After merging, update the similarity matrix and calculate the similarity between the new cluster and other clusters. The similarity between clusters is calculated using the average link method, i.e., the similarity between the new cluster and another cluster is equal to the average of the similarities between all activities in the new cluster and all activities in the other cluster. Repeat the above merging process until no activity pair has a similarity exceeding its adaptive threshold, at which point the clustering process stops.

[0052] After clustering, each cluster represents a description of the same activity on different platforms or at different times. Although the activities within a cluster have different textual descriptions, they all describe the same tourism activity. The clustering results provide a foundation for subsequent information fusion and ensure the accuracy of deduplication.

[0053] S34. Based on the activity clustering results, a multi-source information weighted fusion strategy is adopted to obtain the deduplicated and integrated unified activity information; For each cluster, activity information from multiple sources within the cluster needs to be merged into a unified activity representation. Information fusion requires comprehensive consideration of factors such as the authority, completeness, and timeliness of the information.

[0054] First, the quality of each activity information within the cluster is assessed. The assessment dimensions include: platform authority (assigning authority scores based on factors such as platform popularity, user base, and content moderation mechanisms, with official tourism websites receiving higher authority scores than social media platforms); information completeness (calculating an information completeness score based on factors such as the detail of the activity description and the number of information fields included, with more detailed and comprehensive descriptions receiving higher completeness scores); and text freshness (calculating a freshness score based on the information's collection or publication time, with more recent information receiving higher freshness scores, using a time decay function where the freshness score decays exponentially over time).

[0055] Based on the combined scores from the three dimensions mentioned above, a comprehensive weight is calculated for each piece of activity information. The comprehensive weight equals the platform authority score multiplied by the authority weight coefficient, plus the information completeness score multiplied by the completeness weight coefficient, plus the text freshness score multiplied by the freshness weight coefficient. The weight coefficients are set according to actual application needs; if more emphasis is placed on information authority, the authority weight coefficient is increased; if more emphasis is placed on information timeliness, the freshness weight coefficient is increased.

[0056] A weighted fusion method is employed to integrate activity information within clusters. For textual fields such as activity descriptions, text concatenation or summary extraction methods are used to integrate descriptive texts from multiple sources, generating a more comprehensive activity description. During text concatenation, texts from different sources are sorted according to their overall weight, with texts of higher weights displayed first. For summary extraction, an extractive summarization method is used to select the most important sentences from texts from multiple sources to form a summary; the importance of a sentence is calculated based on its frequency of occurrence and overall weight across multiple sources. For structured fields such as activity location, activity time, and cost range, voting or weighted averaging methods are used for fusion. For categorical fields, the categories given by different sources are statistically analyzed, and the category with the highest frequency or the largest overall weight is selected as the fusion result. For numerical fields, a weighted average of the numerical values ​​given by different sources is calculated, with the weight being the overall weight.

[0057] After fusion, unified activity information is generated, which combines the advantages of multiple sources and has higher accuracy and completeness. The fused activity information is stored in a unified activity database, replacing the duplicate information from multiple sources, thus achieving deduplication and integration of cross-platform information.

[0058] S4. Based on the tourism activity information set, a method combining transfer learning and collaborative filtering is used to obtain activity information with dynamic quality scores and confidence levels. This step aims to establish a dynamic activity quality assessment system. By analyzing user review texts, it tracks changes in activity quality in real time, providing quality assurance for recommendations. Specifically, it includes the following steps: S41. Based on user review text, sentiment analysis and key information extraction are used to obtain the sentiment tendency and quality-related features of the reviews. Collect user reviews of tourism activities, including user comments on tourism platforms, user shares on social media, and discussion posts on tourism forums. Perform preprocessing on the review texts.

[0059] A sentiment analysis model is used to classify the sentiment of evaluation texts. The model is fine-tuned based on a pre-trained language model and can classify evaluation texts into three categories: positive, negative, or neutral, and calculate the sentiment intensity.

[0060] Key information extraction techniques are employed to extract specific information related to activity quality from the evaluation text, including service quality, facility status, user experience, cost-effectiveness, and safety. Named entity recognition and relation extraction techniques are used to identify quality-related entities and attributes mentioned in the evaluation, and to construct a quality feature vector.

[0061] For each evaluation text, we obtain the sentiment tendency label, sentiment intensity value, and quality feature vector, which serve as input data for quality assessment.

[0062] S42, Based on activities with sufficient evaluation data, a transfer learning training quality assessment model is used to obtain a source domain quality assessment model; Tourism activities are grouped according to attributes such as type and region. For each group, activities with sufficient evaluation data are selected as source domain data. The criterion for sufficient evaluation data is that the number of evaluations exceeds a preset threshold, such as more than 100 evaluations. These activities with sufficient evaluation data can reliably calculate quality scores using statistical methods, serving as supervisory signals for training the quality assessment model.

[0063] For the source domain data, a quality assessment model is constructed. The model's input consists of the sentiment tendency, sentiment intensity, and quality feature vector of the evaluation; the output is the activity's quality score. The quality score is represented by a continuous value from zero to ten, with higher scores indicating better quality. The model is implemented using a regressive neural network, containing multiple fully connected layers and activation functions. It achieves quality assessment by learning the mapping relationship between evaluation features and quality scores.

[0064] The model is trained using a supervised learning method, with training data containing evaluation features and corresponding quality score labels. For activities with sufficient evaluation data, the quality score labels are obtained through statistical aggregation. Specifically, all evaluations of the activity are categorized according to sentiment: positive evaluations are assigned 8-10 points, negative evaluations 1-3 points, and neutral evaluations 4-7 points, with the specific scores adjusted within the range based on sentiment intensity. Then, a weighted average of all evaluation scores is calculated, with weights determined based on the freshness of the evaluation and the credibility of the evaluator, resulting in the overall quality score label for the activity. This statistical method reliably reflects activity quality when evaluation data is abundant, but it is not suitable for activities with sparse evaluation data. Therefore, a model needs to be trained to predict the quality of activities with sparse data. The training objective is to minimize the mean squared error between the model's predicted score and the statistically obtained quality score label, and the model parameters are optimized using a backpropagation algorithm.

[0065] After training, the model learns how to evaluate activity quality from evaluation features. This model, as the source domain model, provides the basis for transfer learning for evaluating activities with sparse data.

[0066] S43, Based on the source domain quality assessment model, the transfer learning method is used to assess the quality of activities with sparse evaluation data and obtain the transfer learning score. For activities with sparse evaluation data—that is, activities with fewer than a preset threshold of evaluations—transfer learning is used for quality assessment. First, the group to which the activity belongs is determined, and a source domain quality assessment model for the corresponding group is selected. Then, the parameters of the source domain model are used as initialization and fine-tuned using a small amount of evaluation data from the activity. During fine-tuning, a small learning rate is used to update the model parameters incrementally, avoiding overfitting to the limited data and losing the general knowledge learned from the source domain. The goal of fine-tuning is to adapt the model to the specific evaluation characteristics of the activity while preserving the quality assessment capabilities of the source domain model.

[0067] After fine-tuning, the fine-tuned model is used to assess the quality of the activity. The activity's evaluation features are input, and a quality score is output. This score, called the transfer learning score, reflects the quality estimate obtained based on transfer learning methods.

[0068] For new activities with no evaluation data, fine-tuning is not possible. The source domain model is used directly for quality assessment, outputting an initial quality score. However, the reliability of this initial quality score is low. As evaluation data accumulates for the activity, it is gradually fine-tuned and updated.

[0069] S44. Based on the feature similarity of activities, a collaborative filtering method is used to infer the activity quality score and obtain the collaborative filtering score. Collaborative filtering methods infer quality scores based on the similarity between activities. For a target activity with sparse evaluation data, other activities with similar features are retrieved from the activity database. Similarity calculation is based on the comprehensive feature vector obtained in step S2, using cosine similarity as a metric. Similarity values ​​range from zero to one, with higher similarity indicating greater similarity between the two activities.

[0070] Select the activities with the highest similarity as neighbor activities, setting the number of neighbors to a fixed value, such as selecting the 20 most similar activities. Obtain the quality scores of these neighbor activities, which are derived from the aggregation of their historical evaluation data or existing quality assessment results.

[0071] A weighted average method is used to infer the quality score of the target activity based on the quality scores of neighboring activities. The inference formula is: the collaborative filtering score equals the sum of the quality scores of all neighboring activities multiplied by their similarity weights, divided by the sum of the similarity weights. The similarity weights are determined based on the feature similarity between the target activity and its neighboring activities; the higher the similarity, the greater the weight of the neighboring activity and the greater its influence on the inference result.

[0072] Collaborative filtering scoring reflects quality inferences based on similar activities. This method is suitable for situations where evaluation data is sparse but similar activities exist. By drawing on the quality information of similar activities, it provides quality estimates for the target activity.

[0073] S45. Based on historical evaluation data of the activity, time series analysis is used to track the trend of quality changes and obtain the trend adjustment factor. For activities with a certain amount of accumulated evaluation data, time series analysis is used to track the trend of their quality scores over time. The evaluation data is arranged chronologically, and a sliding window mechanism is used to divide the time axis into multiple time windows, each containing evaluation data for a specific period. The length of the time window is set according to the density of the evaluation data; shorter time windows are used for activities with dense evaluation data, and longer time windows are used for activities with sparse evaluation data.

[0074] For each time window, the average quality score within that window is calculated, yielding time series data. This time series data reflects how activity quality changes over time. Moving averages or exponential smoothing methods are used to smooth the time series data, eliminating short-term fluctuations and highlighting long-term trends.

[0075] Analyze the smoothed time series to identify trends in quality changes. Trend analysis uses linear regression to fit a linear relationship between time and quality scores; the regression coefficient represents the rate of change of quality over time. A positive regression coefficient indicates an upward trend in quality; a negative regression coefficient indicates a downward trend in quality; and a regression coefficient close to zero indicates that quality remains stable.

[0076] A trend adjustment factor is calculated based on the quality trend. The trend adjustment factor reflects the impact of the quality trend on the current quality score. For activities with increasing quality, the trend adjustment factor is positive, improving the current quality score; for activities with decreasing quality, the trend adjustment factor is negative, decreasing the current quality score; for activities with stable quality, the trend adjustment factor is zero, and the current quality score is not adjusted. The magnitude of the trend adjustment factor is directly proportional to the rate of change; the greater the rate of change, the larger the absolute value of the adjustment factor.

[0077] S46, based on transfer learning scoring, collaborative filtering scoring, and trend adjustment factors, employs a weighted fusion strategy to obtain dynamic quality scores and confidence levels. The transfer learning score obtained in step S43, the collaborative filtering score obtained in step S44, and the trend adjustment factor obtained in step S45 are combined, and a weighted fusion strategy is used to calculate the final dynamic quality score of the activity.

[0078] First, the weights of each rating source are determined. These weights are based on the reliability and data sufficiency of the rating source. For activities with sufficient evaluation data, transfer learning ratings have higher weights because they are based on ample data and therefore highly reliable. For activities with sparse evaluation data, collaborative filtering ratings have higher weights because they draw on information from similar activities, compensating for insufficient data. The weights employ an adaptive adjustment strategy, dynamically adjusting based on the number of evaluations for the activity. The more evaluations, the higher the weight of transfer learning ratings and the lower the weight of collaborative filtering ratings; conversely, the fewer evaluations, the lower the weight of transfer learning ratings and the higher the weight of collaborative filtering ratings.

[0079] The fusion formula is: Dynamic Quality Score equals Transfer Learning Score multiplied by Transfer Learning Weights plus Collaborative Filtering Score multiplied by Collaborative Filtering Weights plus Trend Adjustment Factor. Through weighted fusion, Dynamic Quality Score integrates the advantages of multiple evaluation methods, considering not only the activity's own evaluation data but also information from similar activities, and reflecting the trend of quality changes, thus possessing high accuracy and timeliness.

[0080] The confidence score of a dynamic quality rating is calculated, reflecting the reliability of the rating result. The confidence score calculation considers the following factors: the quantity of evaluation data (more evaluations result in higher confidence); the temporal distribution of the evaluation data (more recent evaluations result in higher confidence); the consistency between the transfer learning score and the collaborative filtering score (closer scores result in higher confidence); and the similarity of neighbor activities (higher similarity results in higher confidence for the collaborative filtering score). A weighted calculation method is used to obtain the confidence score, which ranges from zero to one, with higher values ​​indicating more reliable ratings.

[0081] The dynamic quality score and confidence level are linked to the activity information to update the activity database. The dynamic quality score serves as the quality attribute of the activity and is used for subsequent recommendation ranking and quality screening; the confidence level serves as the reliability indicator of the score and is used for risk control and priority determination of manual review.

[0082] S5 receives activity information with dynamic quality scores and confidence levels, as well as feature vectors of tourism activities. It then uses a hierarchical sampling strategy to construct an activity association graph and performs graph neural network embedding learning to obtain the activity association graph and activity embedding representation. This step aims to uncover multi-dimensional relationships between tourism activities, providing association information for activity combination recommendations by constructing activity association graphs and graph neural network embedding learning; specifically, it includes the following steps: S51, based on the quality score, popularity and timeliness of the activity, adopts a hierarchical sampling strategy to obtain a hierarchical set of activity nodes; With a large number of activities, including all activities in the association graph would result in an excessively large graph size, leading to excessively high computational complexity. A hierarchical sampling strategy is adopted, stratifying activities based on their importance. More important activities retain more association information, while less important activities retain only key associations or are removed from the graph altogether.

[0083] The importance of an activity is assessed by considering three dimensions: quality score, popularity, and timeliness. The quality score is derived from the dynamic quality score obtained in step S4; activities with higher quality scores are more important. Popularity is calculated based on user behavior data such as activity views, favorites, and reviews; activities with high user attention have higher popularity and therefore higher importance. Timeliness is calculated based on the activity's update time or its time-related attributes; recently updated activities or activities suitable for the current season have higher timeliness and importance.

[0084] The importance score of an activity is calculated by combining three dimensions: quality score multiplied by quality weight, popularity score multiplied by popularity weight, and timeliness score multiplied by timeliness weight. Based on the importance score, activities are divided into multiple levels, such as high importance, medium importance, and low importance. All activities in the high importance level are included in the association graph, and complete relationships are constructed for them. Activities in the medium importance level are selectively included in the association graph, retaining only relationships with high importance activities. Activities in the low importance level are not included in the association graph or only a very small number of key relationships are retained.

[0085] By employing a hierarchical sampling strategy, the size of the association graph is controlled to ensure that the activities included in the graph are high-quality, high-popularity, or high-timeliness activities. These activities are the key targets for recommendations, and constructing association graphs for them can effectively support recommendation tasks.

[0086] S52, based on the comprehensive feature vector of the activities, adopts a multi-dimensional association calculation method to obtain the association edges and association weights between activities; For activities included in the association graph, multi-dimensional relationships between activities are calculated, including thematic associations, geographical associations, temporal associations, and user behavior associations.

[0087] Theme association reflects the similarity of activities in terms of theme content. Theme-related feature dimensions are extracted from the comprehensive feature vectors of the activities, and the cosine similarity of the two activities in terms of theme features is calculated. If the theme similarity exceeds a preset threshold, a theme association edge is established between the two activities, with the weight of the edge being the theme similarity value.

[0088] Geographic association reflects the proximity of activities in terms of geographical location. The geographical distance between two activities is calculated. If the geographical distance is less than a preset threshold, such as 10 kilometers, a geographical association edge is established between the two activities. The weight of the edge is inversely proportional to the distance.

[0089] Temporal association reflects the complementarity or sequence of activities in terms of temporal attributes. By analyzing the temporal attributes of two activities, if the two activities are suitable to be carried out in the same season or have a sequential relationship, a temporal association edge is established, and the weight is determined according to the degree of matching of temporal attributes.

[0090] User behavior associations reflect the co-occurrence of activities in user choices. By analyzing historical user behavior data, combinations of activities participated in by the same user are statistically analyzed, and association rule mining algorithms are used to discover frequently occurring activity combination patterns. For frequently co-occurring activity pairs, their support and confidence are calculated. If both exceed a preset threshold, user behavior association edges are established, with the edge weight being a weighted combination of support and confidence.

[0091] Based on the four types of relationships described above, an activity relationship graph is constructed. Nodes in the graph represent activities, edges represent relationships, and each edge has a type label and a weight value.

[0092] S53, based on the activity association graph, uses a graph neural network for embedding learning to obtain the graph embedding representation of the activity; A graph neural network model is constructed, consisting of multiple graph convolutional layers. Each graph convolutional layer collects and aggregates the feature vectors of each active node in the graph from its neighboring nodes. The aggregation methods include summation, averaging, or weighted summation, with the weights being the weights of the associated edges. The aggregated neighbor features are then combined with the node's own features, and the node's feature representation is updated through linear transformations and nonlinear activation functions.

[0093] Graph neural networks consist of multiple layers, with each layer performing message passing and feature updates. This multi-layered structure allows nodes to aggregate information from distant neighbors, capturing higher-order relationships within the graph. During message passing, an attention mechanism is introduced, assigning different weights to different types of relationships. These attention weights are learned through the neural network.

[0094] Graph neural networks are trained using unsupervised or semi-supervised learning methods. Unsupervised learning makes connected nodes in the graph closer in the embedding space, while semi-supervised learning uses the label information of some nodes for training.

[0095] After training, the graph neural network generates an embedding representation for each activity node. This embedding representation integrates the activity's own features and its structural information in the association graph, providing a foundation of association information for subsequent activity combination recommendations.

[0096] S54. Based on the dynamic changes of the activity association graph, an incremental update mechanism is adopted to obtain the updated activity association graph and embedded representation; Tourism activity information is dynamic, with new activities constantly being added, old activity information continuously updated, and the relationships between activities changing accordingly. Reconstructing the entire relationship graph and retraining the graph neural network for each change would be computationally too costly. An incremental update mechanism, updating only the affected local subgraphs, significantly reduces update costs.

[0097] When a new activity is added, its importance score is first determined. If the importance score meets the criteria for inclusion in the association graph, the activity is added as a new node. The association relationships between the new activity and existing activities are calculated. Following the method in step S52, thematic associations, geographical associations, temporal associations, and user behavior associations are calculated to establish association edges for the new activity. The addition of a new activity only affects the embedding representations of its neighboring nodes. A local update strategy is adopted, updating only the embeddings of the new activity and its neighboring nodes, while the embeddings of other nodes remain unchanged.

[0098] When existing activity information is updated, such as changes in the activity's quality score or feature attributes, the association between the activity and its neighboring activities is recalculated, and the weights of the associated edges are updated. If the association changes significantly, such as the addition or deletion of associated edges, the affected nodes are embedded and updated.

[0099] Incremental updates employ an online learning approach, fine-tuning the graph neural network with new data. Only the embedding parameters of affected nodes are updated, while the embeddings of other nodes remain unchanged. Through this incremental update mechanism, the system can respond to changes in activity information in real time, maintaining the timeliness of the association graph and embedding representations, while avoiding the high computational cost of full reconstruction.

[0100] S6. Based on the user demand text, activity association graph and feature vector of tourism activities, a multi-objective optimization algorithm is used to obtain a tourism activity sequence recommendation scheme. This step aims to select suitable activities from the activity database based on the user's personalized needs and combine them into a coherent sequence of activities to achieve accurate personalized recommendations; specifically, it includes the following steps: S61, based on the user requirement text, uses multi-level feature extraction and constraint parsing to obtain the user requirement feature vector and constraint set; Users' travel needs are expressed in natural language text, such as wanting to experience traditional ethnic minority festivals in Yunnan, suitable for traveling with children, with a budget under 5000 yuan, and the trip scheduled for summer vacation. The user's text is analyzed in depth to extract its characteristics and constraints.

[0101] The multi-level feature extraction framework in step S2 is used to extract features from the demand text. The demand text is input into a pre-trained language model for the tourism domain, and semantic representations are obtained through lexical, sentence, and document-level encoding. Explicit demand features, such as activity type preferences, location preferences, and participant types, are extracted from the semantic representation. Named entity recognition technology is used to identify key entities in the demand text, such as the location entity "Yunnan," the activity type entity "traditional festival activities," and the participant entity "children." Intent recognition technology is used to identify the user's deeper demand intent; for example, experiencing ethnic minority culture reflects the user's preference for cultural experiences, and "suitable for bringing children" reflects the user's concern for parent-child compatibility and safety.

[0102] Construct a user demand feature vector, which shares the same dimensions and semantic space as the activity feature vector, facilitating subsequent matching calculations. The demand feature vector includes feature values ​​from multiple dimensions, such as activity type preference, regional preference, experience preference, and participant suitability.

[0103] The constraints in the requirement text are analyzed. These constraints are the user's hard requirements for the recommendation results and must be met. Constraints include time constraints, budget constraints, geographical constraints, and participant constraints. A combination of rule matching and semantic understanding is used to identify the type and specific value of the constraints. For example, a budget constraint is extracted from the budget of 5000 yuan, with an upper limit of 5000 yuan; a time constraint is extracted from the summer vacation schedule, with the time range being the summer vacation period. The constraints are formally represented, constructing a constraint set. Each constraint includes a constraint type, constraint value, and constraint relationship such as less than, equal to, or belong to.

[0104] S62, based on user demand feature vectors and activity feature vectors, uses a multi-dimensional weighted matching algorithm to obtain the demand matching score of the activity; Calculate the matching degree between the user demand feature vector and each activity feature vector. The matching degree reflects the extent to which the activity meets the user's needs.

[0105] A multi-dimensional weighted matching algorithm is employed. This algorithm does not simply calculate the overall similarity of feature vectors, but instead calculates the matching degree for features in different dimensions separately, and then performs weighted aggregation. For each feature dimension, the similarity or matching degree between the demand feature value and the activity feature value is calculated. For categorical features, the matching degree is 1 if the demand feature value and the activity feature value are the same, and 0 otherwise. For numerical features, the difference between the demand feature value and the activity feature value is calculated; the smaller the difference, the higher the matching degree. A Gaussian function or linear function is used to map the difference to a matching degree value between zero and one. For vector features, the cosine similarity between the demand feature vector and the activity feature vector is calculated, and the similarity value is used as the matching degree.

[0106] The matching scores across different dimensions are weighted and aggregated, with each weight reflecting the importance of that dimension's feature to the user's needs. The weights are dynamically adjusted based on the features of each dimension in the requirement text. An attention mechanism is employed to analyze the expression intensity and frequency of different features in the requirement text, assigning greater weight to features with higher expression intensity and frequency. For example, if a user repeatedly emphasizes cultural experience in their requirement text, the cultural attribute dimension has a higher weight; if the user explicitly mentions suitability for children, the participant suitability dimension has a higher weight.

[0107] The demand matching score is equal to the sum of the matching scores of all dimensions multiplied by their respective weights. The demand matching score ranges from zero to one; a higher score indicates that the activity better meets user needs. Demand matching scores are calculated for all activities to provide a basis for subsequent activity selection.

[0108] S63, Based on the set of constraints, perform constraint screening on the activities to obtain a set of candidate activities that meet the constraints; Based on the set of constraints obtained in step S61, activities are filtered out to remove those that do not meet the constraints.

[0109] For time constraints, check the activity's time attributes, including suitable season, suitable month, and activity duration, to determine if the activity is feasible within the user-specified time frame. If the activity's suitable time does not match the user's time constraints, exclude the activity. For budget constraints, check the activity's cost range to determine if the cost is within the user's budget. If the cost exceeds the user's budget limit, exclude the activity. For geographical constraints, check the activity's geographical location to determine if the activity is within the user-specified geographic area. If the activity's location is outside the specified area, exclude the activity. For participant constraints, check the activity's participant attributes, such as suitable age group and suitable demographic type, to determine if the activity is suitable for the user-specified participants. If the activity is unsuitable for the specified participants, exclude the activity.

[0110] After constraint filtering, a set of candidate activities that satisfy all constraints is obtained. All activities in the candidate activity set are feasible, meeting the user's hard requirements, and providing a candidate space for subsequent activity combination optimization.

[0111] S64. Based on the candidate activity set and activity association graph, a genetic algorithm is used to solve the multi-objective optimization problem to obtain the Pareto optimal activity sequence set; The goal of recommendations is not simply to select the single activity with the highest degree of matching of needs, but to build a sequence of activities that contains multiple activities with logical connections between them, which together meet the user's needs and optimize multiple objectives.

[0112] Define a multi-objective optimization problem with the following objectives: maximizing demand matching degree (maximizing the sum of demand matching degree scores for all activities in the activity sequence); maximizing activity quality score (maximizing the sum of dynamic quality scores for all activities in the activity sequence); maximizing activity sequence coherence (maximizing the sum of connection strengths between adjacent activities in the association graph); and minimizing time and space costs (minimizing the total time and total space costs of the activity sequence). Time cost includes the duration of activities and travel time between activities, while space cost includes geographical distance between activities and transportation costs.

[0113] Genetic algorithms are used to solve multi-objective optimization problems. A genetic algorithm is a heuristic search algorithm that searches for the optimal solution in a candidate solution space by simulating the process of biological evolution.

[0114] Activity sequences are encoded and represented as chromosomes. A chromosome is a sequence of activity identifiers; its length represents the number of activities, and each element in the sequence represents an activity. The population is initialized by randomly generating several chromosomes, each representing a candidate activity sequence. During initialization, activities are randomly selected from the candidate activity set to form a sequence, with the sequence length within a preset range, such as 3 to 7 activities. During initialization, constraints are checked on each generated chromosome, calculating the total cost, total duration, and geographical scope of the activity sequence. If the user's constraints are not met, a new chromosome is generated to ensure that all chromosomes in the initial population are feasible solutions.

[0115] The fitness of each chromosome is evaluated, calculated based on multiple optimization objectives. For each chromosome, its total demand matching score, total quality score, coherence score, time cost, and space cost are calculated. The coherence score is calculated based on the weights of the edges connecting adjacent activities in the activity sequence within the association graph. If an edge exists between adjacent activities, the coherence score increases by the edge weight; otherwise, it does not. The time cost is calculated by estimating travel time based on the activity duration and the geographical distance between activities. The space cost is calculated based on geographical distance.

[0116] A non-dominated sorting method is used to sort the chromosomes in the population. The non-dominated sorting divides the chromosomes into multiple front layers. The first front layer contains Pareto optimal solutions, i.e., solutions that are not dominated by other solutions on any objective. A solution dominates another solution if and only if the solution is not worse than the other solution on any objective and is better than the other solution on at least one objective.

[0117] Genetic operations include selection, crossover, and mutation. Selection selects superior chromosomes for the next generation based on their non-dominated status and crowding distance. Crowding distance reflects the distribution density of chromosomes in the target space; chromosomes with larger crowding distances are prioritized to maintain solution diversity. Crossover randomly selects two chromosomes, exchanging partial gene segments to generate a new chromosome. During crossover, the structural information of the activity association graph is considered, prioritizing the exchange of closely connected activity segments to improve crossover efficiency. After crossover, the newly generated chromosome undergoes constraint checks. If constraints are violated, repair operations are performed, such as deleting activities exceeding the budget or replacing them with similar activities with lower costs, ensuring the generated chromosome remains a feasible solution. Mutation randomly selects a chromosome, randomly replacing one of its activities or adjusting the order of activities to introduce new variations and avoid the algorithm getting trapped in local optima. During mutation, a replacement activity is selected from the candidate activity set, ensuring the replaced chromosome still satisfies the constraints.

[0118] The process of repeated evaluation, sorting, selection, crossover, and mutation is carried out over multiple generations, allowing the population to gradually evolve. Superior chromosomes are preserved and propagated, while inferior chromosomes are eliminated. The iteration terminates when a preset number of iterations is reached or the population's fitness no longer significantly improves.

[0119] After iteration, chromosomes from the first frontier layer are extracted from the final population. These chromosomes represent Pareto optimal activity sequences, i.e., activity sequences that achieve the best balance among multiple optimization objectives. The Pareto optimal solution set contains multiple candidate solutions, each with different emphases on different objectives.

[0120] S65, based on the Pareto optimal activity sequence set and user preference weights, uses a weighted scoring method to obtain the final recommended activity sequence scheme; The final recommended activity sequence is selected from the set of Pareto optimal solutions. Since Pareto optimal solutions have varying advantages and disadvantages for different objectives, the selection needs to be based on the user's preference weights.

[0121] User preference weights reflect the degree of importance users place on different optimization goals. If users prioritize activity quality, the quality score will have a higher weight; if users prioritize the compactness of the itinerary, the time cost will have a higher weight; if users prioritize the continuity of the activity, the continuity score will have a higher weight. User preference weights can be inferred from the requirement text, learned from users' historical behavior data, or set by directly asking users.

[0122] For each activity sequence in the Pareto optimal solution set, its comprehensive score is calculated. The comprehensive score equals the total score of demand matching multiplied by the demand matching weight, plus the total score of quality multiplied by the quality weight, plus the coherence score multiplied by the coherence weight, minus the time cost multiplied by the time cost weight, minus the space cost multiplied by the space cost weight. Through weighted calculation, multiple optimization objectives are integrated into a single comprehensive score. The higher the score, the more the solution meets the user's comprehensive needs and preferences.

[0123] The activity sequence with the highest overall score is selected as the final recommended option. Simultaneously, multiple alternative options with top overall scores are provided to increase the diversity of recommendations and give users more choices. Each recommended option includes information such as the activity sequence, detailed information about the activities, the relationships between the activities, estimated time and space costs, and the overall score, presented to the user in a visual manner to help them understand and make a selection.

[0124] In one embodiment of the present invention, in order to establish a continuous learning and optimization mechanism for the model, enabling the system to continuously learn from new data, adapt to changes in the tourism market, and improve model performance, the method further includes: S7. Based on newly collected tourism activity texts and user feedback data, an active learning strategy is used to continuously optimize the model, resulting in an optimized feature extraction model and recommendation model. This step is an optimization process executed after the system completes the initial construction and is put into operation in steps S1 to S6, entering the continuous operation phase. It forms a complete closed loop with the previous steps: data collection - model training - system application - continuous optimization. Specifically, it includes the following steps: S71. Based on newly collected tourism activity texts, model prediction uncertainty assessment is used to obtain a set of high-value samples to be labeled. During operation, the system continuously collects new activity text data from various tourism platforms. The newly collected data undergoes the same preprocessing process as in step S1, and is then input into the currently running feature automatic annotation model for processing.

[0125] For newly collected activity text, the current automatic feature annotation model is used for feature annotation, and the model outputs the annotation results and the confidence level of the annotations. An uncertainty assessment method is used to quantify the model's prediction uncertainty for each sample, including entropy calculation based on probability distribution and prediction variance calculation based on ensemble model.

[0126] Based on the uncertainty assessment results, the samples with the highest uncertainty are selected as high-value samples to be labeled. These are samples that the model currently has difficulty in accurately handling. Manually labeling them and using them for model training can maximize the improvement of model performance.

[0127] S72, based on high-value unlabeled samples, uses manual annotation and quality review to obtain high-quality labeled data; High-value samples to be labeled are submitted to domain experts for manual annotation. The annotation content includes various dimensions such as activity type, spatiotemporal attributes, participant attributes, and experience attributes. The manual annotation results undergo quality review, including multi-person annotation consistency checks, expert review, and comparison of annotation results with model predictions. After quality review, high-quality labeled data is obtained, which is used as new training samples for incremental training and optimization of the model.

[0128] S73, based on high-quality labeled data, uses incremental learning to optimize the feature extraction model, resulting in an updated feature extraction model; An incremental learning approach is employed to optimize the automatic feature annotation model using new, high-quality labeled data. The new labeled data is mixed with a portion of historical labeled data to construct an incremental training dataset; the historical data is selected using an importance sampling strategy.

[0129] The model was fine-tuned using an incremental training dataset with a small learning rate. Regularization techniques, such as elastic weight consolidation, were employed to add a regularization term to the loss function, protecting the model's memory of old knowledge and preventing catastrophic forgetting.

[0130] After fine-tuning, the model's performance is evaluated on the validation dataset. If the model's performance improves on the new data while remaining stable on the old data, the incremental learning is considered successful. The optimized feature extraction model can more accurately handle new types of activity text and adapt to new changes in the tourism market.

[0131] S74. Based on user feedback data, reinforcement learning is used to optimize the recommendation strategy and obtain the updated recommendation model. Collect user feedback data on the recommendation results, including explicit feedback such as user ratings and whether they accept the recommendation, as well as implicit feedback such as browsing time, click behavior, collection and sharing.

[0132] The recommendation process is modeled as a Markov decision process, where the state represents the user's needs and the current recommended activity sequence, the action is the activity of selecting the next recommendation, and the reward is the user's feedback signal. A deep Q-network method is used to optimize the recommendation strategy. A Q-network is constructed where the input is the state representation and the output is the Q-value for each action.

[0133] The Q-network is trained using user feedback data, with training data consisting of a quadruple of state, action, reward, and next state. The training process employs an experience replay mechanism and a target network mechanism to improve training stability. After training, the Q-network learns the optimal recommendation strategy and can continuously adjust based on actual user feedback, improving recommendation accuracy and user satisfaction.

[0134] S75, based on model performance monitoring results, employs automatic rollback and diagnostic mechanisms to obtain a stable and reliable model version; Establish a model performance monitoring mechanism to continuously monitor the model's performance on new data. Monitoring metrics include feature annotation accuracy, recommendation click-through rate, and user satisfaction rating. When a monitoring metric falls below a threshold, a performance alert is triggered.

[0135] When a performance degradation of the model is detected, an automatic rollback mechanism is activated. The system maintains multiple historical model versions and automatically rolls back to the previous stable version to avoid the impact of performance degradation on user experience.

[0136] Simultaneously, a model diagnostic process is initiated to analyze the reasons for performance degradation. Diagnostic methods include data distribution analysis, error sample analysis, and model parameter analysis. Based on the diagnostic results, a model optimization plan is formulated. Possible optimization measures include adding annotations for specific types of data, adjusting the model structure, adjusting training hyperparameters, and improving data preprocessing methods.

[0137] After implementing the optimization scheme, the model is retrained, and its performance is evaluated on the validation dataset. If the performance recovers or improves, the new model is deployed online. Through automatic rollback and diagnostic mechanisms, the system can quickly respond to model performance issues, maintaining system stability and reliability.

[0138] This invention focuses on the intelligent operation and application of customized tourism service platforms. A customized tourism service platform uses the method of this invention to provide customers with personalized tourism activity recommendation services.

[0139] The platform collects tourism activity information from 20 tourism platforms, including social sharing platforms, travel guide platforms, online booking platforms, local life service platforms, and official tourism websites. The total number of activities collected reaches 50,000, covering various types such as cultural experiences, outdoor adventures, leisure and entertainment, and educational study tours. Examples of the collected data are shown in Table 1.

[0140] Table 1: Examples of collected tourism activity data

[0141] The system automatically labels the collected activity data with features, and examples of the labeling results are shown in Table 2.

[0142] Table 2: Examples of Automatic Feature Annotation Results

[0143] The system deduplicatively integrates the same activities from different platforms. For example, Yunnan Bai ethnic tie-dyeing experience and Dali Bai ethnic tie-dyeing workshop are identified as different descriptions of the same activity. After information fusion, unified activity information is generated. The fused description is more comprehensive and includes the advantages of both platforms.

[0144] The system collects user feedback data on activities and conducts dynamic quality assessments. For example, the Yunnan Bai ethnic group's tie-dyeing experience activity collected 120 user reviews in the past three months, of which 85% were positive, 5% were negative, and 10% were neutral. The system analyzes the review text, extracts quality-related features, and calculates a dynamic quality score of 8.5 with a confidence level of 0.88.

[0145] A client submitted a request: They want to experience traditional ethnic minority cultural activities in Yunnan, suitable for a 10-year-old child, with strong interactivity and educational value, a budget of less than 500 yuan per person, and the trip should be scheduled for mid-July during the summer vacation, lasting 3 days.

[0146] The system parses the requirement text and extracts its characteristics: regional preference is Yunnan, activity type preference is cultural experience - traditional culture, participants are parents and children, participant age is 10 years old, high interactivity requirement, and high educational significance requirement. The system also extracts constraints: budget constraint is within 500 yuan per person, time constraint is mid-July, and trip duration is 3 days.

[0147] The system calculates the matching degree between candidate activities and requirements, and selects candidate activities that meet the constraints, including 10 activities such as Yunnan Bai ethnic tie-dyeing experience, Yunnan Dai ethnic water splashing festival experience, and Yunnan Naxi ethnic Dongba culture experience.

[0148] The system employs a multi-objective optimization algorithm to generate recommended activity sequences. The final recommended sequence includes four activities: Day 1 morning - Yunnan Bai ethnic tie-dyeing experience; Day 1 afternoon - Dali Ancient City cultural exploration; Day 2 - Yunnan Naxi Dongba culture experience; Day 3 - Yunnan ethnic minority song and dance performance. The sequence has a demand matching score of 0.91, a total quality score of 33.2, an activity sequence coherence score of 2.8, an estimated total cost of 1350 yuan per person, and a total duration of 3 days.

[0149] The client was satisfied with the recommended plan, accepted it, and actually participated in the activity. After the activity, the client gave a positive review, rating it 9 out of 10, and commented that the activity was well-organized, the child enjoyed the tie-dyeing experience, learned a lot about ethnic culture, and the itinerary was smooth and not tiring; the client was very satisfied.

[0150] The system collects customer feedback data as training samples for reinforcement learning to optimize recommendation strategies. Simultaneously, the system continuously gathers new activity information, uses active learning methods to select high-value samples for manual annotation, and regularly updates the feature extraction and recommendation models to maintain system performance and timeliness.

[0151] Through the method of this invention, the customized tourism service platform achieves intelligent collection and analysis of tourism activity information, providing customers with accurate and personalized recommendation services, improving customer satisfaction, and significantly enhancing the platform's operational efficiency and service quality. The method of this invention has broad application value, not only applicable to customized tourism service platforms, but also extend to online tourism platforms, tourism information aggregation platforms, tourism consulting services, and other fields, providing technical support for the intelligent development of the tourism industry.

[0152] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for collecting and analyzing tourism activity information based on text features, characterized in that: Includes the following steps: S1. Collect the original tourism activity text dataset and use a self-supervised learning method to build an automatic feature annotation model to obtain tourism activity data; S2, acquire tourism activity data, and use a multi-level feature extraction framework and feature fusion method to obtain the feature vector of tourism activities; S3 receives the feature vectors of tourism activities and uses an adaptive threshold clustering algorithm and information fusion strategy to obtain a set of tourism activity information. S4, based on the tourism activity information set, uses a fusion of transfer learning and collaborative filtering to obtain activity information with dynamic quality scores and confidence levels; including: A sentiment analysis model was used to classify the sentiment tendencies of user review texts, and key information extraction technology was used to extract quality-related features from the review texts. Tourism activities are grouped according to type and regional attributes. Activities with more than a preset threshold of evaluations are selected as source domain data. A quality assessment model is constructed and trained by learning the mapping relationship between evaluation features and quality scores to obtain the source domain quality assessment model. For activities with fewer than a preset threshold of evaluations, the parameters of the source domain model are used as initialization for fine-tuning to obtain a transfer learning score; neighbor activities are retrieved based on the cosine similarity of the comprehensive feature vector, and a weighted average method is used to infer the collaborative filtering score of the target activity; a time series analysis method is used to track the trend of quality changes to obtain a trend adjustment factor. The transfer learning score, collaborative filtering score, and trend adjustment factor are weighted and fused to obtain the dynamic quality score and confidence level. S5 receives activity information with dynamic quality scores and confidence levels, as well as feature vectors of tourism activities. It then uses a hierarchical sampling strategy to construct an activity association graph and performs graph neural network embedding learning to obtain the activity association graph and activity embedding representation. S6. Based on the user demand text, activity association graph, and feature vector of tourism activities, a multi-objective optimization algorithm is used to obtain a tourism activity sequence recommendation scheme.

2. The method for collecting and analyzing tourism activity information based on text features according to claim 1, characterized in that, S1 includes: Based on the original tourism activity text dataset, data cleaning and preprocessing operations were performed to obtain standardized text data; Based on standardized text data, a masked language model is constructed to mask some words in the text. By predicting the masked words, a self-supervised pre-training is performed to obtain a deep semantic representation of the text. Based on deep semantic representation of text, a sentence order prediction task is designed to learn the logical structure of activity descriptions. A contrastive learning task is designed to maximize the feature similarity of different descriptions of the same activity and minimize the feature similarity of different activity descriptions. The mask language model task, sentence order prediction task and contrastive learning task are jointly trained to obtain enhanced discriminative feature representations. Based on enhanced discriminative feature representation and a small amount of manually labeled seed data, a multi-dimensional feature labeling model is constructed, which includes a shared feature extraction layer and multiple task-specific classification layers. The model is then fine-tuned using transfer learning. The text of tourism activities is labeled with features based on a multi-dimensional feature automatic labeling model. The highest probability value of the categorized features is used as the confidence level. Labeling results with confidence levels lower than a preset threshold are marked as pending manual review, resulting in tourism activity data with multi-dimensional feature labels and confidence levels.

3. The method for collecting and analyzing tourism activity information based on text features according to claim 1, characterized in that, S2 includes: Based on large-scale tourism-related texts, domain pre-training was performed using masked language modeling tasks and next sentence prediction tasks to obtain a tourism domain pre-trained language model. Based on a pre-trained language model in the tourism field, a multi-level feature extraction framework is constructed, which includes a lexical encoder, a sentence encoder, and a document encoder, to obtain feature representations at the lexical, sentence, and document levels. Explicit features are extracted from multi-dimensional feature annotation results, and implicit features are extracted from multi-level feature representations. A gating fusion mechanism is designed to dynamically adjust the fusion weights of explicit and implicit features based on the confidence level of feature annotations, thereby obtaining a comprehensive feature vector.

4. The method for collecting and analyzing tourism activity information based on text features according to claim 1, characterized in that, S3 includes: Design a platform feature encoder, learn a feature transformation matrix for each platform, and train it through a contrastive learning method to map the feature vectors of different platforms to a unified semantic space; Design an adaptive threshold determination method based on local density awareness to calculate the local density of each activity in the feature space; A hierarchical clustering algorithm is used to select activity pairs with similarity exceeding an adaptive threshold from the similarity matrix and merge them until no activity pair has a similarity exceeding its adaptive threshold. Each activity information within a cluster is evaluated for quality and its overall weight is calculated. A weighted fusion method is then used to merge the activity information within the clusters to obtain a unified activity information after deduplication and integration.

5. The method for collecting and analyzing tourism activity information based on text features according to claim 1, characterized in that, S5 includes: The importance score of an activity is calculated by combining quality score, popularity, and timeliness. Based on the importance score, the activity is divided into multiple levels and incorporated into a relationship diagram. Calculate the thematic, geographical, temporal, and user behavior relationships between activities, and construct an activity relationship graph; A graph neural network model is constructed, and the feature vectors of the neighboring nodes of each active node are collected and aggregated to obtain the graph embedding representation of the activity.

6. The method for collecting and analyzing tourism activity information based on text features according to claim 1, characterized in that, S6 includes: A multi-level feature extraction framework is used to extract features from user requirement text, named entity recognition technology is used to identify key entities, and user requirement feature vectors are constructed. A combination of rule matching and semantic understanding is used to parse constraints and construct a set of constraints. A multi-dimensional weighted matching algorithm is used to calculate the matching degree of features in different dimensions and perform weighted aggregation to obtain the activity's demand matching degree score. Activities are filtered based on the set of constraints to obtain a set of candidate activities that meet the constraints.

7. The method for collecting and analyzing tourism activity information based on text features according to claim 6, characterized in that, S6 further includes: Define a multi-objective optimization problem, with optimization objectives including maximizing demand matching degree, maximizing activity quality score, maximizing activity sequence coherence, and minimizing time and space costs; A genetic algorithm is used to solve a multi-objective optimization problem. The activity sequences are encoded as chromosomes. The chromosomes in the population are sorted using a non-dominated sorting method. Genetic operations such as selection, crossover, and mutation are performed to obtain a Pareto optimal set of activity sequences. Calculate a comprehensive score for each activity sequence in the Pareto optimal solution set based on user preference weights, and select the activity sequence with the highest comprehensive score as the final recommendation.

8. The method for collecting and analyzing tourism activity information based on text features according to claim 1, characterized in that, Also includes: S7, based on newly collected tourism activity texts and user feedback data, employs an active learning strategy for continuous model optimization; specifically including: The newly collected activity texts are labeled with features using an automatic feature labeling model. An uncertainty assessment method is used to quantify the model’s prediction uncertainty for each sample, and the sample with the highest uncertainty is selected as a high-value sample to be labeled. High-value unlabeled samples are submitted to domain experts for manual annotation and quality review to obtain high-quality labeled data. An incremental learning method is used to fine-tune the model by mixing new labeled data with some historical labeled data, resulting in an updated feature extraction model. User feedback data is collected, and the recommendation process is modeled as a Markov decision process. A deep Q-network method is used to optimize the recommendation strategy, resulting in an updated recommendation model.

9. A text-feature-based tourism activity information collection and analysis system, used to perform the steps in the text-feature-based tourism activity information collection and analysis method as described in any one of claims 1-8, characterized in that, include: The feature annotation module is used to collect raw tourism activity text datasets and use a self-supervised learning method to build an automatic feature annotation model to obtain tourism activity data. The feature extraction module is used to acquire tourism activity data. It adopts a multi-level feature extraction framework and feature fusion method to obtain the feature vector of tourism activities. The information integration module receives feature vectors of tourism activities and uses an adaptive threshold clustering algorithm and information fusion strategy to obtain a set of tourism activity information. The quality assessment module is used to obtain activity information with dynamic quality scores and confidence levels by using a combination of transfer learning and collaborative filtering based on a set of tourism activity information. The graph embedding module is used to receive activity information with dynamic quality scores and confidence levels, as well as feature vectors of tourism activities. It uses a hierarchical sampling strategy to construct an activity association graph and performs graph neural network embedding learning to obtain the activity association graph and activity embedding representation. The recommendation generation module is used to generate a tourism activity sequence recommendation scheme based on user demand text, activity association graph and feature vector of tourism activities, using a multi-objective optimization algorithm.

Images