An intelligent camping site recommendation system and method

Abstract

The application discloses an intelligent camping ground recommendation system and method, and relates to the technical field of artificial intelligence, wherein the data collection layer is responsible for acquiring and preprocessing user behavior, campsite multi-modal information and environmental data, and constructing a basic database. The semantic understanding layer is based on a pre-trained large language model, analyzes the intent and entity of the user's natural language query, and generates a user demand portrait. The feature fusion layer adopts a cross-attention mechanism, deeply fuses the structured, text and visual features of the campsite, and generates a unified campsite representation vector. The recommendation generation layer adopts a double-tower architecture, performs personalized sorting through the recommendation tower, and generates corresponding natural language explanations using the generation tower. The intelligent interaction layer supports multi-round dialogue, and based on the RAG enhanced explanation system, combines an authoritative knowledge base to generate a reliable recommendation explanation, effectively improving the accuracy, explainability and interactive experience of the camping ground recommendation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to an intelligent campsite recommendation system and method. Background Technology

[0002] With the continuous improvement of residents' consumption levels and the increasing demand for outdoor leisure experiences, camping tourism has gradually become one of the most important leisure and vacation options for the public. To meet users' needs for choosing campsites, existing technologies generally employ recommendation systems. These systems collect users' historical behavioral data and primarily rely on algorithms such as keyword matching and collaborative filtering to construct user profiles, thereby providing users with basic campsite recommendation services. Such methods can, to some extent, achieve preliminary personalized recommendations based on users' past preferences.

[0003] However, the aforementioned existing technical solutions still have significant limitations. Because they primarily rely on structured data and statistical patterns, the systems lack deep semantic understanding of users' natural language query intent, making it difficult to parse complex, ambiguous, or emotionally charged text descriptions. Furthermore, existing systems often fail to effectively integrate multimodal information related to campsites (such as environmental images, geographical features, facility photos, and text reviews), resulting in relatively singular recommendation criteria. This further leads to insufficient interpretability of recommendation results, a somewhat rigid user experience, and ultimately affects the accuracy of recommendations and user satisfaction.

[0004] Therefore, given the shortcomings of existing campsite recommendation systems in semantic understanding and multi-source information fusion, a more intelligent recommendation scheme is urgently needed. It is necessary to develop an intelligent campsite recommendation system and method that combines large language models with multi-dimensional data fusion to improve the ability to deeply understand complex user needs, enhance the accuracy and interpretability of recommendation results, and thus improve the overall user experience. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of existing technologies and provide an intelligent campsite recommendation system and method. By combining the semantic understanding of pre-trained large language models, the cross-attention mechanism of multimodal data fusion, the dual-tower recommendation generation architecture, and RAG enhanced interpretation technology, this application solves the technical problems of existing campsite recommendation systems that cannot accurately handle complex natural language queries from users, lack interpretability of recommendation results, and have a monotonous user experience.

[0006] The objective of this application is achieved through the following technical solution:

[0007] Firstly, this application proposes an intelligent campsite recommendation system, the system comprising:

[0008] The data acquisition layer is used to acquire user behavior data, campsite multimodal information, and environmental data, and to preprocess them to form a basic information database of the campsite and a user preference dataset.

[0009] The semantic understanding layer, connected to the data acquisition layer, is used to perform intent recognition, entity extraction, and semantic vectorization on user natural language queries based on a pre-trained large language model, thereby obtaining a profile of the user's camping needs.

[0010] The feature fusion layer, connected to the semantic understanding layer, is used to deeply fuse campsite structured data, text semantic features, and image visual features using a cross-attention mechanism to generate a unified campsite representation vector and construct a full-dimensional campsite feature map.

[0011] The recommendation generation layer, connected to the feature fusion layer, adopts a dual-tower architecture, including a recommendation tower and a generation tower. The recommendation tower uses a Transformer encoder to output personalized camp rankings based on camp representation vectors and user demand profiles, while the generation tower generates natural language explanations corresponding to the ranking results.

[0012] The intelligent interaction layer, connected to the recommendation generation layer, is used to interact with users based on a multi-turn dialogue mechanism. It generates credible recommendation explanations based on an authoritative knowledge base through the RAG enhanced explanation system, and outputs personalized intelligent campsite recommendation solutions.

[0013] In one possible implementation, the semantic understanding layer includes:

[0014] The intent recognition unit is built based on a pre-trained large language model and uses a neural network-based classifier to recognize the intent of natural language queries, identifying at least one query intent among campsite search, price inquiry, facility inquiry, location search, review viewing, booking inquiry, recommendation request, and camping inquiry.

[0015] The entity extraction unit employs sequence labeling technology based on a large language model and uses the BIO labeling system to extract key entity information such as location, price, time, and facilities from user queries.

[0016] The semantic encoding unit utilizes the representation learning capability of a pre-trained large language model to generate a high-dimensional semantic vector representation of the user query, wherein the semantic vector is calculated as follows: ,in For the semantic vector of the user query, For pre-trained language models, For pooling operations, For users to query text;

[0017] The query preprocessing unit performs text cleaning, word segmentation, and standardization on the natural language input by the user;

[0018] The semantic similarity calculation unit calculates the semantic similarity between the user query semantic vector and the candidate campsites, and uses the cosine similarity measurement method to perform query semantic matching, user preference matching, and historical camping behavior matching.

[0019] The query expansion unit expands related camping query terms based on the semantic vector of the user query using a word vector model;

[0020] The intent confidence assessment unit calculates the confidence score of the intent recognition result and triggers a clarification dialogue mechanism when the confidence score is lower than a preset threshold.

[0021] In one possible implementation, the feature fusion layer includes:

[0022] The structured feature encoding unit uses a multilayer perceptron network to process the numerical attributes of campsite price, rating, number of facilities, and geographical location. It generates structured feature vectors through batch normalization and nonlinear transformation. The batch normalization calculation method is as follows: ,in For batch normalization function, and All are learnable parameters. and These are the batch mean and standard deviation, respectively. Input data;

[0023] The text semantic encoding unit performs deep semantic encoding on user reviews, camp descriptions, and service introduction texts based on a pre-trained language model, and combines sentiment analysis to capture user experience feedback.

[0024] The image feature encoding unit uses a deep convolutional neural network to extract visual features from camp images and identify visual information such as environmental features, facility conditions, and spatial layout.

[0025] The cross-attention fusion unit uses a multi-head attention mechanism to calculate the correlation weights between features of different modalities, and generates a unified camp representation vector through weighted fusion.

[0026] The modality alignment unit maps features of different modalities to a unified vector space through projection transformation;

[0027] The feature importance evaluation unit evaluates the contribution of different modal features to the recommendation effect based on information gain theory and dynamically adjusts the fusion weights.

[0028] The multimodal consistency detection unit detects the consistency between different modal information and identifies and processes conflicting information.

[0029] In one possible implementation, the cross-attention fusion unit is computed in the following ways:

[0030] Calculate the attention weights for structured features and text features: ;

[0031] Calculate the attention weights for text features and image features: ;

[0032] Calculate the attention weights for image features and structured features: ;

[0033] A unified camp representation vector is generated through weighted fusion:

[0034]

[0035] , , These are structured, text, and image feature vectors, respectively. This indicates an element-wise multiplication operation. This represents a vector concatenation operation. , and All are learnable weight matrices. For normalization function, , and All of them are multilayer perceptrons.

[0036] In one possible implementation, the recommended generation layer includes:

[0037] The recommendation tower unit is built on a multi-layer Transformer encoder and uses a multi-task learning framework to simultaneously perform click-through rate prediction, rating prediction and ranking learning, and output personalized camp recommendation ranking results.

[0038] The generation tower unit is built based on a pre-trained large language model and adopts an autoregressive generation method. It takes user queries, recommendation results and camp attribute information as input and generates natural language recommendation explanations.

[0039] The joint optimization unit is trained end-to-end using a joint loss function to ensure logical consistency between the recommendation results and the explanations. The total loss function is: ,in To recommend losses, To generate loss, For alignment loss, , , These are the weighting coefficients;

[0040] The weight adaptive unit dynamically adjusts the weights of each loss term based on gradient information during training.

[0041] A diversity control unit is recommended, which uses diversity regularization techniques to ensure the diversity of recommendation results;

[0042] The quality assessment unit uses a combination of automated assessment indicators and manual assessment to evaluate the quality and credibility of the generated explanations.

[0043] The personalized weight learning unit learns personalized feature weights and recommendation strategies based on users' historical behavior and preference features.

[0044] In one possible implementation, the intelligent interaction layer includes:

[0045] The dialogue state management unit uses a finite state machine to manage the dialogue process, including the initial state, query understanding, result display, clarification inquiry, result adjustment, and dialogue end state transition.

[0046] The ambiguity detection unit detects the degree of ambiguity in a user's query by calculating the intent distribution entropy. When the ambiguity exceeds a threshold, a clarification dialogue is triggered. The degree of ambiguity is defined as follows:

[0047]

[0048] For ambiguous scoring, Let i be the probability of the i-th intent given a query. This represents the summation of probabilities for all possible intent categories;

[0049] The feedback processing unit dynamically adjusts the recommendation strategy and weight parameters based on user satisfaction feedback and behavioral data using reinforcement learning algorithms.

[0050] The multi-round interaction unit allows users to refine their requirements, modify conditions, and filter results.

[0051] In one possible implementation, the intelligent interaction layer further includes:

[0052] The user intent tracking unit tracks changes in user intent and evolution of needs across multiple rounds of dialogue, and maintains the dialogue context state.

[0053] The Personalized Dialogue Strategy Unit develops personalized dialogue strategies and response styles based on user profiles and dialogue history.

[0054] The dialogue effectiveness evaluation unit assesses the performance of the dialogue system through dialogue success rate and user satisfaction indicators, and continuously optimizes dialogue strategies.

[0055] In one possible implementation, the RAG enhanced interpretation system includes:

[0056] The knowledge base building unit integrates authoritative knowledge from multiple sources, including official campsite information, expert guides, safety guidelines, industry reports, and user reviews, and uses a vectorized index storage method.

[0057] The knowledge retrieval unit employs dense retrieval technology to retrieve relevant knowledge fragments based on the semantic similarity between user queries and recommendation results. It also optimizes retrieval quality through a re-ranking model. The relevance score is calculated as follows: ,in and These are the vector representations of the query and the document, respectively, and ⊙ represents element-wise multiplication. It is a multilayer perceptron;

[0058] The knowledge enhancement generation unit constructs a structured prompt template by combining the retrieved knowledge fragments with the recommended information, and generates recommended explanations by integrating an attention mechanism;

[0059] The fact consistency detection unit uses a natural language reasoning model to evaluate the consistency between the generated explanation and the source knowledge. When the consistency score is lower than the threshold, a regeneration mechanism is triggered.

[0060] In one possible implementation, the fusion attention mechanism of the knowledge enhancement generation unit is as follows:

[0061]

[0062] For query vector, , Key-value pairs representing text features. , Key-value pairs representing knowledge features. For weighting;

[0063] The fusion weights are learned through a gating mechanism: ; Calculation of the importance weight of knowledge fragments: , For query representation vectors, For knowledge representation vectors, For the gated weight matrix, For gating bias terms, For activation function, For the k-th knowledge fragment, This represents the attention score.

[0064] Secondly, this application proposes a smart campsite recommendation method, the method comprising:

[0065] Step S1: Multi-source data collection and preprocessing. Based on the distributed crawler system and API interface, user behavior data, campsite multimodal information and environmental data are obtained. Data cleaning, formatting and quality control are performed to build a basic information database of campsites and a user preference dataset.

[0066] Step S2: Natural language query parsing, based on a pre-trained large language model, performs intent recognition, entity extraction and semantic encoding on user input to generate structured query representation and user demand profile;

[0067] Step S3: Multimodal feature fusion, using a cross-attention mechanism to integrate the structured features of the camp, textual semantic features, and image visual features to generate a unified camp representation vector;

[0068] Step S4: Recommendation calculation and interpretation generation. A dual-tower architecture is used to simultaneously perform personalized recommendation ranking and natural language interpretation generation, and end-to-end optimization is achieved through a joint loss function.

[0069] Step S5: Interactive recommendation optimization, supporting multi-turn dialogue interaction, dynamically adjusting recommendation strategies based on user feedback, and generating knowledge-enhanced recommendation explanations through RAG technology.

[0070] The main solution and its various further alternatives described above can be freely combined to form multiple solutions, all of which are solutions that can be adopted and are claimed in this application; furthermore, the (non-conflicting alternatives) can also be freely combined with each other and with other alternatives. Those skilled in the art, after understanding the solution of this application, will realize from the prior art and common general knowledge that there are many combinations, all of which are technical solutions to be protected in this application, and will not be exhaustively listed here.

[0071] This application discloses an intelligent campsite recommendation system and method, relating to the field of artificial intelligence technology. The data acquisition layer is responsible for acquiring and preprocessing user behavior, campsite multimodal information, and environmental data to construct a basic database. The semantic understanding layer, based on a pre-trained large language model, parses the intent and entities of user natural language queries to generate user demand profiles. The feature fusion layer employs a cross-attention mechanism to deeply fuse the structured, textual, and visual features of campsites, generating a unified campsite representation vector. The recommendation generation layer adopts a dual-tower architecture, using recommendation towers for personalized ranking and generating corresponding natural language explanations using the generation towers. The intelligent interaction layer supports multi-turn dialogue and, based on the RAG enhanced explanation system, combines an authoritative knowledge base to generate credible recommendation explanations, effectively improving the accuracy, explainability, and interactive experience of campsite recommendations. Attached Figure Description

[0072] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 A schematic diagram of the structure of an intelligent campsite recommendation system proposed in an embodiment of this application is shown.

[0074] Figure 2 A flowchart illustrating a smart campsite recommendation method proposed in an embodiment of this application is shown.

[0075] Figure 3 A diagram of the multimodal feature fusion network architecture proposed in an embodiment of this application is shown. Detailed Implementation

[0076] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0077] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0078] In existing technologies, user profiles are built based on keyword matching and simple collaborative filtering algorithms to achieve basic recommendation functions. While this can meet user needs to a certain extent, it lacks deep semantic understanding and multimodal information integration capabilities, making it unable to effectively handle complex natural language queries from users. This results in a lack of interpretability in recommendation results, a monotonous user experience, and ultimately, low recommendation accuracy and low user satisfaction.

[0079] Therefore, to address the technical problems in existing technologies where the lack of deep semantic understanding and multimodal information integration capabilities leads to ineffective handling of complex natural language queries, lack of interpretability in recommendation results, and monotonous user experience, resulting in insufficient recommendation accuracy and low user satisfaction, this application proposes an intelligent campsite recommendation system and method. This system fully utilizes the powerful semantic understanding capabilities of a pre-trained large language model to deeply analyze users' natural language queries, obtaining accurate user demand profiles. Simultaneously, it combines multimodal data fusion technology to integrate structured information, text reviews, images, and other multi-dimensional data from campsites to construct comprehensive campsite feature profiles. Based on the intelligent matching of user needs and campsite features, and through an interpretable recommendation generation mechanism, it provides users with more accurate, diverse, personalized, and intelligent campsite recommendation services.

[0080] Please refer to Figure 1 , Figure 1 This application provides a schematic diagram of the structure of an intelligent campsite recommendation system, comprising:

[0081] The data acquisition layer is used to acquire user behavior data, campsite multimodal information, and environmental data, and to preprocess them to form a basic information database of the campsite and a user preference dataset.

[0082] The semantic understanding layer, connected to the data acquisition layer, is used to perform intent recognition, entity extraction, and semantic vectorization on user natural language queries based on a pre-trained large language model, thereby obtaining a profile of the user's camping needs.

[0083] The feature fusion layer, connected to the semantic understanding layer, is used to deeply fuse campsite structured data, text semantic features, and image visual features using a cross-attention mechanism to generate a unified campsite representation vector and construct a full-dimensional campsite feature map.

[0084] The recommendation generation layer, connected to the feature fusion layer, adopts a dual-tower architecture, including a recommendation tower and a generation tower. The recommendation tower uses a Transformer encoder to output personalized camp rankings based on camp representation vectors and user demand profiles, while the generation tower generates natural language explanations corresponding to the ranking results.

[0085] The intelligent interaction layer, connected to the recommendation generation layer, is used to interact with users based on a multi-turn dialogue mechanism. It generates credible recommendation explanations based on an authoritative knowledge base through the RAG enhanced explanation system, and outputs personalized intelligent campsite recommendation solutions.

[0086] The data acquisition layer collects three core types of data in real time or periodically through distributed web crawlers, application programming interfaces (APIs), and a logging system: user behavior data, campsite multimodal information, and environmental data. The collected raw data is cleaned, deduplicated, standardized in format, and assessed for quality by the data preprocessing unit, ultimately forming a high-quality campsite basic information database and user preference dataset.

[0087] The data acquisition layer includes:

[0088] The user behavior data collection unit gathers user behavior data such as campsite browsing history, click behavior, booking history, and evaluation feedback to construct user camping preference trajectories. The campsite multimodal information collection unit acquires multimodal data such as structured attribute information, text descriptions, user reviews, campsite real-world images, and video materials to establish a comprehensive campsite information archive. The environmental data collection unit collects environmental background data such as weather information, traffic conditions, attraction distribution, and safety index around the campsite to construct a campsite environmental feature database. The data preprocessing unit cleans, denoises, standardizes, and unifies the format of the collected heterogeneous data to ensure the quality and consistency of campsite recommendation data.

[0089] The data acquisition layer also includes: a data quality monitoring unit, which monitors the integrity and accuracy of campsite data collection in real time, identifies and processes abnormal data, and ensures the reliability of basic data; a privacy protection unit, which uses data desensitization and anonymization technologies to protect user privacy information and ensure the compliance of camping user data usage; and a data storage management unit, which adopts a distributed storage architecture to establish an efficient campsite data retrieval and indexing mechanism, supporting rapid access to large-scale campsite information.

[0090] The semantic understanding layer connects with the data acquisition layer to deeply analyze users' camping needs expressed in natural language. This layer is built based on a pre-trained large language model, employing few-shot learning and prompting engineering techniques. When a user inputs a query, this layer first uses an intent recognition unit to determine the fundamental purpose of the query, and then uses an entity extraction unit to extract key constraint entities using sequence labeling technology. Finally, the semantic encoding unit transforms the entire query into a high-dimensional semantic vector, thereby generating a structured profile of the user's camping needs.

[0091] The feature fusion layer is connected to the semantic understanding layer. Addressing the multi-source and heterogeneous nature of campsite information, a multimodal fusion network based on a cross-attention mechanism is designed. Specifically, the structured feature encoding unit processes numerical attributes such as campsite price and rating; the text semantic encoding unit analyzes text such as campsite descriptions and reviews, incorporating sentiment analysis; and the image feature encoding unit extracts visual features from campsite images. The crucial cross-attention fusion unit dynamically weights and fuses these features by calculating the correlation weights between feature vectors of different modalities, ultimately generating a unified campsite representation vector containing full-dimensional information, which is then used to construct a full-dimensional campsite feature map.

[0092] The recommendation generation layer is connected to the feature fusion layer, employing an innovative dual-tower architecture to simultaneously complete recommendation ranking and explanation generation. The recommendation tower, built on a Transformer encoder, performs deep matching and inference between the semantic vectors of the user's camping needs profile and the campsite representation vectors, outputting a personalized campsite ranking list through multi-task learning. Simultaneously, the generation tower, based on another pre-trained large language model, takes the user query, recommendation results, and campsite attributes as input, and uses an autoregressive approach to generate fluent and natural text explanations explaining the reasons for the recommendations.

[0093] The intelligent interaction layer connects with the recommendation generation layer, providing a dynamic and trustworthy interactive experience. At its core is a multi-turn dialogue management engine that tracks dialogue state, manages context, and proactively clarifies when ambiguous user queries are detected. More importantly, this layer integrates a Retrieval Enhanced Generation (RAG) system. When generating or supplementing recommendation explanations, the RAG system retrieves the most relevant knowledge fragments from a pre-built authoritative knowledge base and combines them with the generation capabilities of a large language model to output professional, accurate, and verifiable recommendation explanations. Ultimately, this layer outputs a personalized campsite intelligent recommendation solution that integrates precise ranking, natural language interpretation, and professional knowledge citations.

[0094] The semantic understanding layer includes:

[0095] The intent recognition unit is built based on a pre-trained large language model and uses a neural network-based classifier to recognize the intent of natural language queries, identifying at least one query intent among campsite search, price inquiry, facility inquiry, location search, review viewing, booking inquiry, recommendation request, and camping inquiry.

[0096] The entity extraction unit employs sequence labeling technology based on a large language model and uses the BIO labeling system to extract key entity information such as location, price, time, and facilities from user queries.

[0097] The semantic encoding unit utilizes the representation learning capability of a pre-trained large language model to generate a high-dimensional semantic vector representation of the user query, wherein the semantic vector is calculated as follows: ,in For the semantic vector of the user query, For pre-trained language models, For pooling operations, For users to query text;

[0098] The query preprocessing unit performs text cleaning, word segmentation, and standardization on the natural language input by the user;

[0099] The semantic similarity calculation unit calculates the semantic similarity between the user query semantic vector and the candidate campsites, and uses the cosine similarity measurement method to perform query semantic matching, user preference matching, and historical camping behavior matching.

[0100] The query expansion unit expands related camping query terms based on the semantic vector of the user query using a word vector model;

[0101] The intent confidence assessment unit calculates the confidence score of the intent recognition result and triggers a clarification dialogue mechanism when the confidence score is lower than a preset threshold.

[0102] First, the query preprocessing unit cleanses, segments, and standardizes the raw query input by the user, removing irrelevant characters, correcting spelling errors, and converting it into a unified format, laying the foundation for subsequent deep analysis. Then, the intent recognition unit, built on a pre-trained Large Language Model (LLM), leverages its powerful contextual understanding capabilities and combines it with a neural network-based classifier to accurately classify the intent of the pre-processed query. This unit can identify various query intents, including but not limited to campsite searches, price inquiries, facility searches, location lookups, review viewing, booking inquiries, recommendation requests, and other camping inquiries, thus clarifying the user's core purpose. Simultaneously, the entity extraction unit uses sequence labeling technology based on the same LLM and follows the BIO labeling system to accurately identify and extract key camping entity information such as location, price, time, and facilities from the query text. These entities constitute the specific constraints of the user's needs. To condense the semantic information of the entire query into a computable mathematical form, the semantic encoding unit utilizes the deep representation learning capabilities of the pre-trained LLM to encode the user query text into a high-dimensional, dense semantic vector.

[0103] The semantic understanding layer also integrates: a semantic similarity calculation unit, based on the generated user query semantic vector, uses cosine similarity and other measurement methods to calculate its multi-dimensional semantic relevance with candidate campsites. This calculation includes not only direct query semantic matching but also user long-term preference matching and historical camping behavior matching, thus achieving a more comprehensive relevance assessment. The query expansion unit utilizes a word vector model to automatically associate and add semantically related camping-related vocabulary based on the current query's semantic vector, expanding and enriching the query expression to compensate for implicit needs that users may have omitted or incompletely expressed. Finally, the intent confidence assessment unit quantifies and scores the certainty of the intent recognition result. When the calculated confidence score is lower than a preset reliability threshold, this unit triggers the system's clarification dialogue mechanism, guiding the user to further clarify their intent, thereby ensuring the accuracy of understanding the user's camping needs from the source.

[0104] The feature fusion layer includes:

[0105] The structured feature encoding unit uses a multilayer perceptron network to process the numerical attributes of campsite price, rating, number of facilities, and geographical location. It generates structured feature vectors through batch normalization and nonlinear transformation. The batch normalization calculation method is as follows: ,in For batch normalization function, and All are learnable parameters. and These are the batch mean and standard deviation, respectively. Input data;

[0106] The text semantic encoding unit performs deep semantic encoding on user reviews, camp descriptions, and service introduction texts based on a pre-trained language model, and combines sentiment analysis to capture user experience feedback.

[0107] The image feature encoding unit uses a deep convolutional neural network to extract visual features from camp images and identify visual information such as environmental features, facility conditions, and spatial layout.

[0108] The cross-attention fusion unit uses a multi-head attention mechanism to calculate the correlation weights between features of different modalities, and generates a unified camp representation vector through weighted fusion.

[0109] The modality alignment unit maps features of different modalities to a unified vector space through projection transformation;

[0110] The feature importance evaluation unit evaluates the contribution of different modal features to the recommendation effect based on information gain theory and dynamically adjusts the fusion weights.

[0111] The multimodal consistency detection unit detects the consistency between different modal information and identifies and processes conflicting information.

[0112] The feature fusion layer first processes different types of data through three independent encoding channels: the structured feature encoding unit is responsible for processing the campsite's numerical attributes, such as price, rating, number of facilities, and geographic coordinates. This unit typically uses a multilayer perceptron network and introduces batch normalization technology to stabilize training and accelerate convergence. After batch normalization and subsequent nonlinear transformations, structured feature vectors representing the campsite's objective attributes are generated.

[0113] The text semantic encoding unit focuses on processing unstructured text information, such as user reviews, official camp descriptions, and service introductions. Based on a pre-trained language model, this unit performs deep semantic encoding on this text, transforming it into vectors rich in semantic information. Simultaneously, the unit incorporates sentiment analysis technology to automatically identify positive, negative, or neutral sentiment tendencies from the reviews, thereby accurately capturing user experience feedback and subjective evaluations.

[0114] The image feature encoding unit aims to extract valuable visual information from real-world images of the camp. This unit employs a deep convolutional neural network to automatically learn and identify high-level features in the images, such as the beauty of the natural environment, the condition of facilities, and the rationality of the campsite layout, ultimately outputting image feature vectors that represent the visual content.

[0115] The core of the cross-attention fusion unit is a multi-head attention mechanism, capable of dynamically calculating the correlation weights between each pair of structured features, textual semantic features, and image visual features. By performing weighted summation and interaction on feature vectors from different modalities based on attention weights, this unit ultimately generates a unified campsite representation vector that integrates information from all dimensions. This vector forms the data foundation for the system's full-dimensional campsite feature map. The calculation method of the cross-attention fusion unit includes:

[0116] Calculate the attention weights for structured features and text features: ;

[0117] Calculate the attention weights for text features and image features: ;

[0118] Calculate the attention weights for image features and structured features: ;

[0119] A unified camp representation vector is generated through weighted fusion:

[0120]

[0121] , , These are structured, text, and image feature vectors, respectively. This indicates an element-wise multiplication operation. This represents a vector concatenation operation. , and All are learnable weight matrices. For normalization function, , and All of them are multilayer perceptrons.

[0122] The modality alignment unit uses projection transformation to map feature vectors from different encoders into a shared, unified vector space, resolving the issue of feature scale and distribution differences caused by different sources. The feature importance evaluation unit, based on theories such as information gain, automatically analyzes the influence of different modal features on the final recommendation result and dynamically adjusts their weights during fusion, enabling the model to adaptively focus on more important information. The multimodal consistency detection unit is responsible for verifying whether there are logical conflicts between information from different sources. By identifying and handling such conflicts, it significantly improves the reliability and credibility of the final feature representation.

[0123] Recommended generation layers include:

[0124] The recommendation tower unit is built on a multi-layer Transformer encoder and uses a multi-task learning framework to simultaneously perform click-through rate prediction, rating prediction and ranking learning, and output personalized camp recommendation ranking results.

[0125] The generation tower unit is built based on a pre-trained large language model and adopts an autoregressive generation method. It takes user queries, recommendation results and camp attribute information as input and generates natural language recommendation explanations.

[0126] The joint optimization unit is trained end-to-end using a joint loss function to ensure logical consistency between the recommendation results and the explanations. The total loss function is: ,in To recommend losses, To generate loss, For alignment loss, , , These are the weighting coefficients;

[0127] The weight adaptive unit dynamically adjusts the weights of each loss term based on gradient information during training.

[0128] A diversity control unit is recommended, which uses diversity regularization techniques to ensure the diversity of recommendation results;

[0129] The quality assessment unit uses a combination of automated assessment indicators and manual assessment to evaluate the quality and credibility of the generated explanations.

[0130] The personalized weight learning unit learns personalized feature weights and recommendation strategies based on users' historical behavior and preference features.

[0131] The recommendation generation layer comprises a recommendation tower unit and a generation tower unit. The recommendation tower unit, built upon a multi-layer Transformer encoder, receives a unified campsite representation vector from the feature fusion layer and a user demand profile from the semantic understanding layer. Through its powerful sequence modeling and attention mechanisms, this tower deeply analyzes the complex matching relationships between user needs and the features of each campsite. It employs a multi-task learning framework, simultaneously optimizing multiple objectives such as click-through rate prediction, rating prediction, and ranking learning, ultimately outputting a list of campsite recommendations ranked according to personalized matching degrees. Working in parallel, the generation tower unit is based on a pre-trained Large Language Model (LLM). It takes the user's original query, the ranking results generated by the recommendation tower, and specific attribute information of the campsites as input.

[0132] The joint optimization unit performs end-to-end joint training of the two towers using a joint loss function. The weight adaptive unit intelligently adjusts the weight coefficients dynamically based on gradient information during training to seek the optimal balance of overall model performance. The recommendation diversity control unit introduces techniques such as diversity regularization to prevent excessive homogenization of recommendation lists, ensuring that the results cover different types of campsites and meet users' exploratory needs. The explanation quality assessment unit uses a combination of automatic evaluation metrics and manual sampling evaluation to continuously monitor the fluency, information content, and credibility of the generated explanations, ensuring the professionalism of the output content. The personalized weight learning unit dynamically adjusts the emphasis on different features within the recommendation model based on each user's unique historical camping behavior and preference profile, achieving deeper personalized recommendations and thus comprehensively improving the accuracy of recommendations and user satisfaction.

[0133] The intelligent interaction layer includes:

[0134] The dialogue state management unit uses a finite state machine to manage the dialogue process, including the initial state, query understanding, result display, clarification inquiry, result adjustment, and dialogue end state transition.

[0135] The ambiguity detection unit detects the degree of ambiguity in a user's query by calculating the intent distribution entropy. When the ambiguity exceeds a threshold, a clarification dialogue is triggered. The degree of ambiguity is defined as follows:

[0136]

[0137] For ambiguous scoring, Let i be the probability of the i-th intent given a query. This represents the summation of probabilities for all possible intent categories;

[0138] The feedback processing unit dynamically adjusts the recommendation strategy and weight parameters based on user satisfaction feedback and behavioral data using reinforcement learning algorithms.

[0139] The multi-round interaction unit allows users to refine their requirements, modify conditions, and filter results.

[0140] The dialogue state management unit employs a finite state machine model to clearly define and manage each stage and its transition logic during a complete camping consultation dialogue. The states include: initial state, query and understanding state, result display state, clarification and inquiry state, result adjustment state, and conversation end state. This unit ensures that the dialogue proceeds in an orderly manner according to the preset logical path.

[0141] The ambiguity detection unit is responsible for identifying the ambiguity of user needs in the early stages of interaction to prevent bias in subsequent recommendations. This unit quantifies the degree of ambiguity by calculating the entropy value of the intent probability distribution output by the semantic understanding layer. When the calculated ambiguity score exceeds a preset threshold, it indicates that the user's query may point to multiple intents. At this point, the unit will trigger a clarification dialogue mechanism to guide the user to clarify their specific needs.

[0142] The feedback processing unit collects explicit feedback and implicit behavioral data generated by users during interaction. Based on these feedback signals, this unit uses reinforcement learning algorithms to dynamically adjust the policy parameters or feature weights in the recommendation generation layer, enabling the system's recommendation behavior to continuously adapt to changes in user preferences and achieve increasingly intelligent personalized services. The multi-turn interaction unit supports users in conducting multiple rounds of dialogue within a single session. This unit maintains the dialogue history context to ensure the system understands the connection between each new request and previous conversations.

[0143] The intelligent interaction layer also includes:

[0144] The user intent tracking unit tracks changes in user intent and evolution of needs across multiple rounds of dialogue, and maintains the dialogue context state.

[0145] The Personalized Dialogue Strategy Unit develops personalized dialogue strategies and response styles based on user profiles and dialogue history.

[0146] The dialogue effectiveness evaluation unit assesses the performance of the dialogue system through dialogue success rate and user satisfaction indicators, and continuously optimizes dialogue strategies.

[0147] The intelligent interaction layer also includes:

[0148] The user intent tracking unit tracks changes in user camping intent and the evolution of needs across multiple rounds of dialogue, maintaining the context of camping consultation conversations. The personalized dialogue strategy unit, based on user profiles and dialogue history, develops personalized camping consultation dialogue strategies and response styles. The dialogue effectiveness evaluation unit assesses the performance of the camping dialogue system through metrics such as dialogue success rate and user satisfaction, continuously optimizing dialogue strategies. The group analysis module, based on user insights, categorizes users by analyzing camping frequency and location preferences, generating camping user group segmentation results. The campsite rating module, based on group segmentation results, extracts and analyzes the historical campsites preferred by each group, performing popularity and satisfaction analysis to generate campsite satisfaction metrics. The campsite selection module, based on campsite satisfaction metrics, sets and adjusts the attractiveness index of campsites, generating adjusted campsite rankings. The recommendation optimization module, based on the adjusted campsite rankings, analyzes users' immediate query intent and preferences, reorders the recommendation list, matches users' actual camping needs, and generates personalized intelligent campsite recommendation solutions.

[0149] RAG enhanced interpretation systems include:

[0150] The knowledge base building unit integrates authoritative knowledge from multiple sources, including official campsite information, expert guides, safety guidelines, industry reports, and user reviews, and uses a vectorized index storage method.

[0151] The knowledge retrieval unit employs dense retrieval technology to retrieve relevant knowledge fragments based on the semantic similarity between user queries and recommendation results. It also optimizes retrieval quality through a re-ranking model. The relevance score is calculated as follows: ,in and These are the vector representations of the query and the document, respectively, and ⊙ represents element-wise multiplication. It is a multilayer perceptron;

[0152] The knowledge enhancement generation unit constructs a structured prompt template by combining the retrieved knowledge fragments with the recommended information, and generates recommended explanations by integrating an attention mechanism;

[0153] The fact consistency detection unit uses a natural language reasoning model to evaluate the consistency between the generated explanation and the source knowledge. When the consistency score is lower than the threshold, a regeneration mechanism is triggered.

[0154] The knowledge base construction unit is responsible for systematically aggregating, cleaning, and structuring authoritative knowledge from multiple sources in the camping field. These knowledge sources include infrastructure and pricing information officially released by campgrounds, professional guides and safety guidelines written by industry associations or experienced practitioners, macro-level industry analysis reports, and verified user reviews. All knowledge documents are transformed into high-dimensional vectors through an embedding model and indexed and stored using an efficient vector database, thereby constructing a professional camping knowledge base that can be quickly semantically retrieved by machines.

[0155] The knowledge retrieval unit is activated when an explanation is needed for a specific recommendation. This unit employs dense retrieval technology, encoding the user's current query or the features of the camp to be explained into vectors, and performs a preliminary approximate nearest neighbor search in the vector knowledge base to recall a batch of potentially relevant knowledge fragments. To further improve retrieval accuracy, this unit introduces a re-ranking model, which uses a neural network to perform deeper interaction modeling of the query and candidate document vectors, calculating a precise relevance score.

[0156] Subsequently, the knowledge enhancement generation unit organizes the retrieved authoritative knowledge fragments, the original recommendation results, and detailed camp attribute information into a structured prompt template, which is then input into the pre-trained large language model. By incorporating an attention mechanism, the model can simultaneously consider the original general knowledge, the user-specific query, and the authoritative evidence retrieved externally when generating explanatory text, thereby generating professional, accurate, and context-appropriate recommendation explanations.

[0157] Finally, to ensure the absolute reliability of the output results, the system includes a factual consistency detection unit. This unit uses a natural language reasoning model to compare and evaluate the logical consistency between the generated explanation and the retrieved knowledge fragments that served as the source. When the consistency score obtained from the evaluation is lower than a preset credibility threshold, this unit will automatically trigger a regeneration mechanism to prevent content that is inconsistent with the facts from being output in the final output, thereby fundamentally ensuring the authority and credibility of the explanation content.

[0158] The fusion attention mechanism of the knowledge enhancement generation unit is as follows:

[0159]

[0160] For query vector, , Key-value pairs representing text features. , Key-value pairs representing knowledge features. For weighting;

[0161] The fusion weights are learned through a gating mechanism: ; Calculation of the importance weight of knowledge fragments: , For query representation vectors, For knowledge representation vectors, For the gated weight matrix, For gating bias terms, For activation function, For the k-th knowledge fragment, This represents the attention score.

[0162] Figure 2 The diagram illustrates a flowchart of a smart campsite recommendation method proposed in an embodiment of this application. This method is applied to the aforementioned smart campsite recommendation system and includes:

[0163] Step S1: Multi-source data collection and preprocessing. Based on a distributed crawler system and API interface, user behavior data, campsite multimodal information and environmental data are obtained. Data cleaning, formatting and quality control are performed to build a basic information database of campsites and a user preference dataset.

[0164] Step S1 involves acquiring multi-source heterogeneous data, including user behavior data, campsite multimodal information, and environmental data, and performing data cleaning, formatting, and quality control. Based on a distributed web crawler system and API interface, user behavior data, campsite information data, and environmental data are collected to obtain a multi-source heterogeneous data set. The user behavior data includes records of actions such as clicks, browsing, searching, and booking; the campsite information data includes structured data such as campsite name, location, price, facilities, and ratings, as well as unstructured data such as campsite descriptions, user reviews, and campsite images. A corresponding data preprocessing workflow is established according to the data types to clean, format, and control the quality of the multi-source heterogeneous data, resulting in a standardized data set.

[0165] Using data quality assessment methods, we construct data quality evaluation indicators, calculate quality indicators such as data integrity, consistency, and accuracy, clean and repair low-quality data, and obtain a high-quality training dataset.

[0166]

[0167] Score the data quality. As a completeness indicator, As a consistency indicator, For accuracy indicators, , , These are the weighting coefficients.

[0168] Data analysis and statistics are performed based on high-quality training datasets. A data dictionary and feature engineering process are constructed, using multi-source heterogeneous data as input and standardized dataset collaboration as output to obtain a data preprocessing model.

[0169] As a preferred implementation, for the processing of various data types, a preprocessing workflow corresponding to multi-source heterogeneous data is established by utilizing text cleaning, image preprocessing, and data standardization techniques. After extracting the corresponding data features, they are input into the subsequent processing module. The data update frequency and storage strategy are determined according to the timeliness and importance of the data, unifying various data types into a standard format to achieve standardized data management and improve the data processing effect of subsequent models. Since camping needs are highly susceptible to external environmental and seasonal factors, time and geographical location factors need to be considered during data preprocessing. Furthermore, to ensure the accuracy of the recommendation results, the feature extraction process focuses on key feature parameters that influence camping choices, avoiding inaccurate output results due to excessive data noise.

[0170] Step S2: Natural language query parsing. Based on a pre-trained large language model, the user input is subjected to intent recognition, entity extraction, and semantic encoding to generate a structured query representation and user demand profile.

[0171] Step S2 uses a pre-trained large language model to perform intent recognition, entity extraction, and semantic encoding on user input, generating a structured query representation. First, the natural language query input from the user is cleaned, segmented, and standardized. Then, a pre-trained large language model is used for intent recognition, identifying query intents such as campsite search, price inquiry, and facility inquiry. This intent recognition includes intent classification, confidence calculation, and ambiguity detection. Sequence labeling technology is used to extract key entity information from the user query, including entity types such as location, price, time, and facility requirements. Based on the user query intent and entity information, a structured query representation is constructed, and the query semantic vector is calculated. ,in Given the probability distribution of the intent of the query, For the intention classification weight matrix, For pre-trained language models.

[0172] Based on the semantic vector of the user query, semantic similarity calculation methods are used to match relevant campsites, and query expansion techniques are employed to enhance the completeness of the query expression. Based on intent recognition and entity extraction results, a user demand model is established to calculate the matching degree between the user query and candidate campsites, using the following method: ,in To match the rating, The cosine similarity function is used. To query semantic vectors, For the camp feature vector, This is the weighting factor.

[0173] A high-dimensional semantic vector representation of user queries is generated through a pre-trained large language model. This representation is then combined with user historical behavior and preference information to construct a personalized query representation. Dialogue state tracking technology is used to manage changes in user intent across multiple rounds of dialogue, generating system responses based on dialogue history and the current query. Query importance weights are calculated based on the semantic vector representation and personalized information to generate the final structured query representation. User queries are categorized according to their complexity, resulting in different processing strategies for simple and complex queries. To ensure timely query parsing, user query history is updated according to time windows to guarantee the accuracy of intent recognition.

[0174] Step S3: Multimodal feature fusion. A cross-attention mechanism is used to integrate the structured features of the camp, textual semantic features, and image visual features to generate a unified camp representation vector.

[0175] Figure 3 This paper illustrates the multimodal feature fusion network architecture proposed in this application. The structured feature encoder uses a multilayer perceptron network to encode features based on structured data such as campsite prices, ratings, facility quantity, and geographical location, generating a structured feature vector. This structured feature encoding includes data standardization, feature transformation, and dimensionality mapping. The text semantic encoder uses a pre-trained language model for deep semantic encoding based on text content such as user reviews, campsite descriptions, and service introductions, combined with sentiment analysis techniques to capture user experience feedback. The image feature encoder uses a deep convolutional neural network to extract visual features based on campsite image data, recognizing visual information such as environmental features, facility conditions, and spatial layout. The cross-attention fusion layer utilizes a cross-attention mechanism to construct a multi-head attention function, calculating the correlation weights between structured features, text features, and image features, performing cross-modal feature fusion to obtain a unified campsite representation vector. Notably, the cross-attention mechanism achieves deep fusion of features from different modalities, generating a campsite representation sentence containing multidimensional information.

[0176] Feature fusion is performed based on a multi-head attention mechanism. A cross-attention fusion network is trained, taking multimodal features as input and a unified camp representation vector as output to obtain a multimodal feature fusion model. Modality alignment technology maps features from different modalities to a unified vector space, ensuring consistency in feature dimensions. The contribution of different modal features to recommendation performance is evaluated based on information gain theory, and the fusion weights are dynamically adjusted. A multimodal consistency detection method is used to identify and handle conflicts between different modalities, generating the final unified camp representation vector. Weights are assigned according to the importance of camp features to obtain high-quality camp feature representations.

[0177] By deeply fusing multimodal data, since the features of different modalities have different expressive power and information content, the correlation weights between modalities are calculated based on the cross-attention mechanism to achieve effective feature integration. Furthermore, modal alignment and consistency detection techniques are used to ensure the quality and reliability of the fused features, thereby improving the comprehensiveness and accuracy of camp representation.

[0178] Step S4: Recommendation calculation and interpretation generation. A dual-tower architecture is used to simultaneously perform personalized recommendation ranking and natural language interpretation generation, and end-to-end optimization is achieved through a joint loss function.

[0179] A dual-tower architecture is employed to simultaneously perform personalized recommendation ranking and natural language interpretation generation, ensuring the accuracy and interpretability of the recommendation results. Based on the recommendation tower module, a recommendation network is constructed using a multi-layer Transformer encoder, employing a multi-task learning framework to simultaneously perform click-through rate prediction, rating prediction, and ranking learning; the recommendation calculation includes: user preference modeling, campsite matching calculation, and recommendation ranking generation. Based on the generation tower module, an interpretation generation network is constructed using a pre-trained large language model, employing an autoregressive generation approach to produce natural language recommendation interpretations. Based on user queries, recommendation results, and campsite attribute information, a structured input representation is constructed to generate detailed recommendation interpretation content.

[0180] By utilizing a joint loss function, an end-to-end optimization objective is constructed to simultaneously optimize recommendation accuracy and explanation quality, ensuring logical consistency between recommendations and explanations. Based on a weight adaptive adjustment mechanism, the weights of each loss term are dynamically adjusted according to gradient information during training to optimize overall model performance. A dual-tower architecture is used for joint training of the recommendation generation network, taking user queries and camp features as input, and recommendation ranking and explanation text as output, to obtain the recommendation generation model.

[0181] Diversity control techniques are employed to ensure the diversity of recommendation results and prevent excessive concentration of recommendations. Based on an explanation quality assessment method, a combination of automated and manual evaluation is used to assess the quality and credibility of the generated explanations. Personalized feature weights and recommendation strategies are learned based on users' historical behavior and preference characteristics to generate the final recommendation results. The recommendations are then sorted according to their relevance and diversity to obtain a high-quality personalized recommendation list.

[0182] Through an innovative dual-tower architecture design, since recommendation accuracy and explanation quality are interrelated, end-to-end optimization based on a joint loss function ensures the logical consistency between recommendation results and explanation content. Furthermore, through multi-task learning and adaptive weight adjustment mechanisms, recommendation accuracy and explanation quality are improved simultaneously, thereby achieving synergistic optimization of recommendation and explanation.

[0183] Step S5: Interactive recommendation optimization, supporting multi-turn dialogue interaction, dynamically adjusting recommendation strategies based on user feedback, and generating knowledge-enhanced recommendation explanations through RAG technology.

[0184] Step S5, based on dialogue state management technology, employs a finite state machine to manage the dialogue process, including state transitions such as initial state, query understanding, result display, clarification inquiry, result adjustment, and session end. The dialogue management includes state tracking, intent updating, and response generation. Based on an ambiguity detection method, the degree of ambiguity in the user's query is detected by calculating the intent distribution entropy; when the ambiguity exceeds a threshold, a clarification dialogue mechanism is triggered. Based on user satisfaction feedback and behavioral data, a reinforcement learning algorithm is used to dynamically adjust the recommendation strategy and weight parameters. Utilizing the RAG enhanced explanation system, a multi-source authoritative knowledge base is constructed, integrating knowledge sources such as official camp information, expert guides, and safety guidelines, and stored using a vectorized indexing method.

[0185] Official campsite information provides accurate basic data, including actual facilities, pricing, and opening hours, ensuring the factual accuracy of the recommendations. Expert guides offer professional camping advice, including best camping times for different seasons, criteria for family-friendly campsites, and camping equipment recommendations, enhancing the professionalism of the recommendations. Safety guidelines provide authoritative safety information, including measures for dealing with severe weather, key points for child safety, and emergency procedures, increasing user trust in the recommendations. The knowledge source uses a vectorized index for storage, enabling fast and accurate knowledge retrieval. When a user queries for family camping options, the system can quickly locate relevant family camping expertise.

[0186] Based on dense retrieval technology, relevant knowledge fragments are retrieved according to the semantic similarity between user queries and recommendation results, and the retrieval quality is optimized through a re-ranking model. Based on a knowledge augmentation generation method, the retrieved knowledge fragments and recommendation information are combined to construct a structured prompt template. A professional and accurate recommendation explanation is generated by fusing an attention mechanism. The RAG augmentation generation network is trained, using queries and knowledge fragments as input and augmented explanations as output, to obtain a knowledge augmentation explanation model.

[0187] By employing a factual consistency detection method and a natural language reasoning model, the consistency between the generated explanation and the source knowledge is evaluated, avoiding the AI ​​illusion problem. Based on user intent tracking technology, the system tracks changes in user intent and evolution of needs across multiple rounds of dialogue, maintaining the dialogue context state. Personalized dialogue strategies and response styles are formulated based on user profiles and dialogue history, generating the final intelligent interaction result. The performance of the dialogue system is evaluated using metrics such as dialogue success rate and user satisfaction, continuously optimizing the dialogue strategy.

[0188] As a preferred implementation, since the RAG enhanced explanation system can generate credible recommendation explanations based on an authoritative knowledge base, avoiding the illusion problem that is prone to occur in traditional generative models, the use of knowledge-enhanced generation technology to update the quality and credibility of recommendation explanations effectively improves the professionalism and user trust of the camping recommendation system.

[0189] Compared with the prior art, the embodiments of this application have the following beneficial effects:

[0190] First, the semantic understanding layer built through a pre-trained large language model can deeply analyze the complex needs expressed by users in natural language, which contain multiple constraints or ambiguous descriptions, overcoming the shortcomings of traditional keyword matching and fixed label filtering methods in terms of insufficient understanding. Simultaneously, by utilizing a cross-attention mechanism to deeply fuse multimodal information such as the camp's structured attributes, review text, and real-world images, a full-dimensional camp feature map is constructed.

[0191] Secondly, by supporting a multi-turn dialogue interaction mechanism, the system can dynamically clarify and refine user needs, flexibly adapting to the personalized preferences of different groups such as families, experienced players, and beginners. The innovative dual-tower architecture combined with retrieval-enhanced generation technology enables the system to provide professional and reliable reasons for each recommendation based on an authoritative camping knowledge base when generating explanations. This not only enhances the interactivity and user-friendliness of the recommendation system but also significantly improves user satisfaction and trust in the recommendation results, assisting users in making efficient decisions.

[0192] Third, through systematic architecture optimization, the entire process from query understanding to result generation is ensured to operate efficiently, keeping the average response time within an acceptable range for real-time interaction. The unique dual-tower architecture design allows recommendation ranking and explanation generation to share the same underlying feature representation and optimization objectives, thus fundamentally guaranteeing a high degree of logical consistency between the final recommendation result and its textual explanation. This solution effectively balances system response speed and output reliability, particularly meeting the dual needs of camping planning scenarios for immediate information acquisition and reliable decision support.

[0193] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A smart campsite recommendation system, characterized in that, The system includes: The data acquisition layer is used to acquire user behavior data, campsite multimodal information, and environmental data, and to preprocess them to form a basic information database of the campsite and a user preference dataset. The semantic understanding layer, connected to the data acquisition layer, is used to perform intent recognition, entity extraction, and semantic vectorization on user natural language queries based on a pre-trained large language model, thereby obtaining a profile of the user's camping needs. The feature fusion layer, connected to the semantic understanding layer, is used to deeply fuse campsite structured data, text semantic features, and image visual features using a cross-attention mechanism to generate a unified campsite representation vector and construct a full-dimensional campsite feature map. The recommendation generation layer, connected to the feature fusion layer, adopts a dual-tower architecture, including a recommendation tower and a generation tower. The recommendation tower uses a Transformer encoder to output personalized camp rankings based on camp representation vectors and user demand profiles, while the generation tower generates natural language explanations corresponding to the ranking results. Recommended generation layers include: The recommendation tower unit is built on a multi-layer Transformer encoder and uses a multi-task learning framework to simultaneously perform click-through rate prediction, rating prediction and ranking learning, and output personalized camp recommendation ranking results. The generation tower unit is built based on a pre-trained large language model and adopts an autoregressive generation method. It takes user queries, recommendation results and camp attribute information as input and generates natural language recommendation explanations. The joint optimization unit is trained end-to-end using a joint loss function to ensure logical consistency between the recommendation results and the explanations. The total loss function is: ,in To recommend losses, To generate loss, For alignment loss, , , These are the weighting coefficients; The weight adaptive unit dynamically adjusts the weights of each loss term based on gradient information during training. A diversity control unit is recommended, which uses diversity regularization techniques to ensure the diversity of recommendation results; The quality assessment unit uses a combination of automated assessment indicators and manual assessment to evaluate the quality and credibility of the generated explanations. Personalized weight learning unit learns personalized feature weights and recommendation strategies based on users' historical behavior and preference features; The intelligent interaction layer, connected to the recommendation generation layer, is used to interact with users based on a multi-turn dialogue mechanism. It generates credible recommendation explanations based on an authoritative knowledge base through the RAG enhanced explanation system, and outputs personalized intelligent campsite recommendation solutions.

2. The intelligent campsite recommendation system as described in claim 1, characterized in that, The semantic understanding layer includes: The intent recognition unit is built based on a pre-trained large language model and uses a neural network-based classifier to recognize the intent of natural language queries, identifying at least one query intent among campsite search, price inquiry, facility inquiry, location search, review viewing, booking inquiry, recommendation request, and camping inquiry. The entity extraction unit employs sequence labeling technology based on a large language model and uses the BIO labeling system to extract key entity information such as location, price, time, and facilities from user queries. The semantic encoding unit utilizes the representation learning capability of a pre-trained large language model to generate a high-dimensional semantic vector representation of the user query, wherein the semantic vector is calculated as follows: ,in For the semantic vector of the user query, For pre-trained language models, For pooling operations, For users to query text; The query preprocessing unit performs text cleaning, word segmentation, and standardization on the natural language input by the user; The semantic similarity calculation unit calculates the semantic similarity between the user query semantic vector and the candidate campsites, and uses the cosine similarity measurement method to perform query semantic matching, user preference matching, and historical camping behavior matching. The query expansion unit expands related camping query terms based on the semantic vector of the user query using a word vector model; The intent confidence assessment unit calculates the confidence score of the intent recognition result and triggers a clarification dialogue mechanism when the confidence score is lower than a preset threshold.

3. The intelligent campsite recommendation system as described in claim 1, characterized in that, The feature fusion layer includes: The structured feature encoding unit uses a multilayer perceptron network to process the numerical attributes of campsite price, rating, number of facilities, and geographical location. It generates structured feature vectors through batch normalization and nonlinear transformation. The batch normalization calculation method is as follows: ,in For batch normalization function, and All are learnable parameters. and These are the batch mean and standard deviation, respectively. Input data; The text semantic encoding unit performs deep semantic encoding on user reviews, camp descriptions, and service introduction texts based on a pre-trained language model, and combines sentiment analysis to capture user experience feedback. The image feature encoding unit uses a deep convolutional neural network to extract visual features from camp images and identify visual information such as environmental features, facility conditions, and spatial layout. The cross-attention fusion unit uses a multi-head attention mechanism to calculate the correlation weights between features of different modalities, and generates a unified camp representation vector through weighted fusion. The modality alignment unit maps features of different modalities to a unified vector space through projection transformation; The feature importance evaluation unit evaluates the contribution of different modal features to the recommendation effect based on information gain theory and dynamically adjusts the fusion weights. The multimodal consistency detection unit detects the consistency between different modal information and identifies and processes conflicting information.

4. The intelligent campsite recommendation system as described in claim 3, characterized in that, The computation methods for cross-attention fusion units include: Calculate the attention weights for structured features and text features: ; Calculate the attention weights for text features and image features: ; Calculate the attention weights for image features and structured features: ; A unified camp representation vector is generated through weighted fusion: , , These are structured, text, and image feature vectors, respectively. This indicates an element-wise multiplication operation. This represents a vector concatenation operation. , and All are learnable weight matrices. For normalization function, , and All of them are multilayer perceptrons.

5. The intelligent campsite recommendation system as described in claim 1, characterized in that, The intelligent interaction layer includes: The dialogue state management unit uses a finite state machine to manage the dialogue process, including the initial state, query understanding, result display, clarification inquiry, result adjustment, and dialogue end state transition. The ambiguity detection unit detects the degree of ambiguity in a user's query by calculating the intent distribution entropy. When the ambiguity exceeds a threshold, a clarification dialogue is triggered. The degree of ambiguity is defined as follows: For ambiguous scoring, Let i be the probability of the i-th intent given a query. This represents the summation of probabilities for all possible intent categories; The feedback processing unit dynamically adjusts the recommendation strategy and weight parameters based on user satisfaction feedback and behavioral data using reinforcement learning algorithms. The multi-round interaction unit allows users to refine their requirements, modify conditions, and filter results.

6. The intelligent campsite recommendation system as described in claim 5, characterized in that, The intelligent interaction layer also includes: The user intent tracking unit tracks changes in user intent and evolution of needs across multiple rounds of dialogue, and maintains the dialogue context state. The Personalized Dialogue Strategy Unit develops personalized dialogue strategies and response styles based on user profiles and dialogue history. The dialogue effectiveness evaluation unit assesses the performance of the dialogue system through dialogue success rate and user satisfaction indicators, and continuously optimizes dialogue strategies.

7. The intelligent campsite recommendation system as described in claim 1, characterized in that, RAG enhanced interpretation systems include: The knowledge base building unit integrates authoritative knowledge from multiple sources, including official campsite information, expert guides, safety guidelines, industry reports, and user reviews, and uses a vectorized index storage method. The knowledge retrieval unit employs dense retrieval technology to retrieve relevant knowledge fragments based on the semantic similarity between user queries and recommendation results. It also optimizes retrieval quality through a re-ranking model. The relevance score is calculated as follows: ,in and These are the vector representations of the query and the document, respectively, and ⊙ represents element-wise multiplication. It is a multilayer perceptron; The knowledge enhancement generation unit constructs a structured prompt template by combining the retrieved knowledge fragments with the recommended information, and generates recommended explanations by integrating an attention mechanism; The fact consistency detection unit uses a natural language reasoning model to evaluate the consistency between the generated explanation and the source knowledge. When the consistency score is lower than the threshold, a regeneration mechanism is triggered.

8. The intelligent campsite recommendation system as described in claim 1, characterized in that, The fusion attention mechanism of the knowledge enhancement generation unit is as follows: For query vector, , Key-value pairs representing text features. , Key-value pairs representing knowledge features. To integrate weights, For standard attention functions; The fusion weights are learned through a gating mechanism: ; Calculation of the importance weight of knowledge fragments: , For query representation vectors, For knowledge representation vectors, For the gated weight matrix, For gating bias terms, For activation function, For the k-th knowledge fragment, This represents the attention score.

9. A method for recommending intelligent campsites, characterized in that, The method includes: Step S1: Multi-source data collection and preprocessing. Based on the distributed crawler system and API interface, user behavior data, campsite multimodal information and environmental data are obtained. Data cleaning, formatting and quality control are performed to build a basic information database of campsites and a user preference dataset. Step S2: Natural language query parsing, based on a pre-trained large language model, performs intent recognition, entity extraction and semantic encoding on user input to generate structured query representation and user demand profile; Step S3: Multimodal feature fusion, using a cross-attention mechanism to integrate the structured features of the camp, textual semantic features, and image visual features to generate a unified camp representation vector; Step S4: Recommendation calculation and interpretation generation. A dual-tower architecture is used to simultaneously perform personalized recommendation ranking and natural language interpretation generation, and end-to-end optimization is achieved through a joint loss function. A multi-task learning framework is used to simultaneously perform click-through rate prediction, rating prediction, and ranking learning, and output personalized camp recommendation ranking results. An autoregressive generation method is used to generate natural language recommendation explanations by taking user queries, recommendation results, and campsite attribute information as input. End-to-end training is performed using a joint loss function to ensure logical consistency between the recommendation results and the explanations. The total loss function is: ,in To recommend losses, To generate loss, For alignment loss, , , These are the weighting coefficients; The weights of each loss term are dynamically adjusted based on gradient information obtained during training. Diversity regularization techniques are used to ensure the diversity of recommendation results; The quality and credibility of the generated explanations are evaluated using a combination of automated evaluation metrics and manual evaluation. Personalized feature weights and recommendation strategies are learned based on users' historical behavior and preference characteristics; Step S5: Interactive recommendation optimization, supporting multi-turn dialogue interaction, dynamically adjusting recommendation strategies based on user feedback, and generating knowledge-enhanced recommendation explanations through RAG technology.

Images