CN-121980087-A - Intelligent travel recommendation method and system integrating knowledge graph and large model

CN121980087ACN 121980087 ACN121980087 ACN 121980087ACN-121980087-A

Abstract

The invention discloses an intelligent travel recommendation method and system integrating a knowledge graph and a large model, and relates to the technical field of intelligent recommendation. The system comprises a data source and processing layer, a core engine layer and an application service layer, wherein the core engine layer comprises a demand understanding engine and a travel knowledge graph. The data source and processing layer gathers multi-source data and preprocesses, the demand understanding engine converts the user natural language inquiry into a structured preference portrait by SFT and DPO fine adjustment, the travel knowledge atlas comprises four kinds of nodes and weight associated edges, and knowledge iteration is realized by multi-step construction and dynamic update. The core engine layer outputs a recommendation result with relevance and novelty and an interpretable reason through cooperative work of intention analysis, map weighted path optimization and context awareness filtering and combination of a exploratory recommendation mechanism. The method solves the problems of insufficient modeling of the deep motivation of the user, poor cold start effect and the like in the prior art, and improves recommendation accuracy, reliability and dynamic adaptability.

Inventors

LI CHUNXIAO
Bi Jianwu
CHEN WEIHONG
Dou Feipeng
LI HAO
CHEN SHOUSONG

Assignees

南开大学
嗯噢哇网络科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260206

Claims (10)

1. The intelligent travel recommendation system integrating the knowledge graph and the large model is characterized by comprising a data source and processing layer, a core engine layer and an application service layer, wherein the core engine layer is respectively in communication connection with the data source and processing layer and the application service layer; the data source and processing layer is used for converging multi-source data and preprocessing, wherein the multi-source data comprises user interaction data, a domain knowledge base, an external content source and real-time situation information, the preprocessing comprises the steps of finishing structured data mapping and importing through an ETL script, finishing entity and relation extraction of unstructured data through an NLP (non-line protocol) pipeline, and finishing user behavior feedback attribution through stream or batch processing; The core engine layer comprises a demand understanding engine and a travel knowledge graph, wherein the demand understanding engine is used for converting a natural language query of a user into a structured preference portrait based on a fine-tuned Large Language Model (LLM), the travel knowledge graph is constructed based on a graph database and comprises four types of nodes including Motivation (motivation), itemFeature (commodity characteristics), item (commodity) and Metadata) and corresponding relation edges, the relation edges comprise TRIGGERSPREFERENCEFOR relation edges between Motivation (motivation) and ItemFeature (commodity characteristics), CONTAINSITEM relation edges between ItemFeature (commodity characteristics) and Item (commodity), HASMETADATA relation edges between Item (commodity) and Metadata, and the TRIGGERSPREFERENCEFOR relation edges and CONTAINSITEM relation edges are provided with weight values for representing relation strength; The application service layer is used for packaging the processing result of the core engine layer into standardized services, including personalized recommendation API, trip planning service and operation analysis platform, and realizing the output of full-link travel recommendation.
2. The intelligent travel recommendation system integrating knowledge graph and large model according to claim 1, wherein the construction and update of the travel knowledge graph comprises: the structured data injection, namely mapping and importing structured data such as an internal POI database, a large model label and the like into a graph database through an ETL script, and generating Item nodes and associated Metadata nodes in batches; Unstructured data mining, namely extracting ItemFeature (commodity characteristics) and Item (commodity) association relations from text data such as UGC content, wandering, attack and the like by using NER, phrase mining, syntactic analysis and pre-training models, mining potential association of Motivation (motivation) and ItemFeature (commodity characteristics) by using a statistical method or a large language model questioning mode, and giving initial weight; user behavior data association, namely dynamically adjusting the weight of a relation edge in the map based on behavior logs such as clicking, collecting, booking, skipping and the like of a user; And (3) knowledge augmentation and life cycle management, namely discovering new entities and relations from an external new text through an NLP technology, extracting temporal information to update nodes or relation attributes, and finishing knowledge denoising, expired knowledge archiving and initial knowledge verification by combining expert auditing.
3. The intelligent travel recommendation system integrating knowledge graph and large model according to claim 1, wherein the constructing of the demand understanding engine comprises: The model training process comprises inputting travel field data and a Prompt template into a basic Large Language Model (LLM) through instruction fine tuning (SFT) to enable the model to learn and generate a structured preference image conforming to a preset Schema, and optimizing the fit degree of a model generation result and human real preference by using a preference data set comprising input Prompt, a better answer (Chosen) and a worse answer (Rejected) through Direct Preference Optimization (DPO); The input and output design is that the input is a context rich instruction Prompt containing user original inquiry, motivation distribution, user portrait and real-time situation, and the output is a structural preference portrait in a JSON format, and the output comprises scenic spot preference (attraction _preferences), accommodation preference (accommodation_preferences), catering preference (resultant_preferences) and preference summary (preference_summary) fields, and each preference field is attached with a feature tag and a corresponding reasoning reason.
4. The intelligent travel recommendation system integrating knowledge graph and large model according to claim 1, wherein the cooperative working mechanism of the core engine layer comprises: the method comprises the steps of intent analysis and preference generation, wherein a demand understanding engine receives natural language query of a user, generates a structured preference portrait by combining multi-source context information, and obtains fact information by querying a travel knowledge graph through a retrieval enhancement generation (RAG) mechanism in the generation process, so that model illusion is reduced; The travel knowledge graph takes motivation and feature labels in the preference portrait as starting points, carries out weighted path optimization along TRIGGERSPREFERENCEFOR, CONTAINSITEM relation sides, calculates the accumulated score of Item nodes, and recalls a candidate commodity set; Context aware filtering, namely combining the real-time context information of the user with Metadata tags of items (commodities), and dynamically rearranging candidate commodity sets through time filtering, context matching, accurate feature matching and LBS sequencing.
5. An intelligent travel recommendation method integrating a knowledge graph with a large model, which is suitable for the intelligent travel recommendation system integrating the knowledge graph with the large model according to any one of claims 1-4, and is characterized by comprising the following steps: S1, data aggregation and preprocessing, namely aggregating user interaction data, a domain knowledge base, external content sources and real-time situation information, and respectively finishing structured data import, unstructured data entity relation extraction and user behavior feedback attribution through ETL scripts, NLP pipelines and stream/batch processing; S2, constructing and updating a travel knowledge graph, constructing the travel knowledge graph comprising Motivation (motivation), itemFeature (commodity characteristics), item (commodity) and Metadata (Metadata) nodes and weight associated sides based on the preprocessed data through structured data injection, unstructured data mining, user behavior data association and knowledge augmentation and life cycle management, and dynamically updating; s3, training a demand understanding engine, enabling a Large Language Model (LLM) to learn through instruction fine tuning (SFT) to generate a structured preference portrait, and obtaining a fine-tuned demand understanding engine through the matching degree of Direct Preference Optimization (DPO) optimization model output and human preference; S4, full-link recommendation execution is carried out, user natural language query is received, a structured preference portrait is generated through a demand understanding engine, a travel knowledge graph carries out weighted path optimization and candidate commodity recall based on the portrait, candidate set filtering rearrangement is completed by combining real-time context information, and finally recommendation results and interpretable recommendation reasons are output through an application service layer.
6. The intelligent travel recommendation method integrating knowledge graph and large model according to claim 5, wherein the weighted path optimizing in S4 comprises the following sub-steps: S4.1, traversing along TRIGGERSPREFERENCEFOR relation sides by taking a user Motivation (motivation) identified in the preference portrait as a starting point to obtain associated ItemFeature (commodity characteristics) nodes and corresponding weights; S4.2, starting from each ItemFeature (commodity characteristic) node, traversing along CONTAINSITEM relation sides to obtain associated Item (commodity) nodes and corresponding weights; s4.3, calculating a path accumulated score of each Item (commodity) node, wherein the score is equal to the product of the edge weights of all relations on the paths, and if a plurality of paths exist in the Item (commodity) node, accumulating the score; And S4.4, correcting the accumulated score by combining Metadata information (heat and score) of Item nodes to generate a candidate commodity set.
7. The intelligent travel recommendation method integrating knowledge graph and large model according to claim 5, wherein the generation mode of the interpretable recommendation reason in S4 is that the reasoning path from Motivation (motivation) to ItemFeature (commodity characteristics) to Item (commodity) in the travel knowledge graph is converted into natural language expression, and the natural language expression comprises association logic of user Motivation (motivation), itemFeature (commodity characteristics) characteristics and Item (commodity) commodity information.
8. The intelligent recommendation method for travel integrating knowledge graph and large model according to claim 5, wherein the dynamically updating in S2 comprises: feedback-driven weight self-learning, namely attributing user behavior feedback to a specific recommended path through batch tasks every day, lifting corresponding side weights for a positive feedback path, and reducing corresponding side weights for a negative feedback path; the map augmentation driven by knowledge is Zhou Du or month is to mine new entity, new relation and time information from the external new text by NLP technology, and supplement the new entity, new relation and time information to the map; expert governance, in which an expert edits and examines initial nodes, edges and weights, periodically examines time knowledge, cleans low confidence knowledge and files expired knowledge.
9. The intelligent travel recommendation method integrating knowledge graphs and large models according to claim 5, wherein S4 further comprises a exploratory recommendation mechanism, wherein a random walk algorithm is introduced in the weighted path optimizing process, and the association relation with lower exploration weight and longer path is explored to generate a recommendation result with relevance and novelty.
10. The intelligent travel recommendation method integrating knowledge graphs and large models according to claim 5, wherein the construction mode of the preference data set in S3 comprises the steps of collecting preference images with higher conversion rate in an A/B test as 'better answers', constructing a preference pair output by the model through manual sequencing, and marking the preference image corresponding to the recommended travel adopted by a user as 'better answer'.

Description

Intelligent travel recommendation method and system integrating knowledge graph and large model Technical Field The invention relates to the technical field of intelligent recommendation, in particular to a travel intelligent recommendation method and system integrating a knowledge graph and a large model. Background In the intelligent tourism field, the personalized recommendation system is a core technical support for improving user experience and business transformation efficiency. With the iterative development of big data and artificial intelligence technology, a recommendation system gets rid of the simple range of early rule-based and collaborative filtering, and gradually evolves into a deep learning model architecture capable of deeply mining implicit characteristics of users and articles. The current mainstream technical scheme is represented by a double-tower model, the model maps user side multi-element characteristics (including user ID embedded vectors, historical interaction behavior sequences, demographic attributes and the like) and article side core characteristics (including travel product ID embedded vectors, category attribute tags, content semantic embedded vectors and the like) to a low-dimensional dense vector space by constructing mutually independent 'user tower' and 'article tower' neural networks, and finally calculates and completes recommendation matching by means of similarity among vectors. The method is technically characterized in that end-to-end training is carried out by relying on massive user-object historical interaction data, so that a user behavior mode is accurately fitted. One of the core defects in the prior art is that the representation of the user demands only stays at the shallow behavior level and lacks modeling capability of deep psychological motivation for driving travel decisions. The mainstream model can only construct user portraits based on explicit interaction data such as user historical clicks, reservations and the like or demographic attributes, is difficult to penetrate behavior appearances, and reaches essential psychological causes of induced travel behaviors, such as seeking identity acceptance, relieving time anxiety and realizing emotion compensation and other core requirements. In the face of fuzzy intention expression such as 'want to find places to relax', or data scarcity scenes such as cold start of new users and cold start of new travel products, the problem that recommendation accuracy is greatly reduced is very easy to occur due to the fact that an existing model lacks an efficient semantic analysis and intention reasoning mechanism and excessively depends on historical interaction data, and potential real requirements of users which are not explicitly expressed are difficult to accurately capture and meet. Another key drawback is represented by the unexplainability of the recommendation logic, and the limitations of knowledge fusion in the travel field. The deep learning scheme represented by the double-tower model has the advantages that the internal decision process is based on similarity operation of an implicit vector space, belongs to a typical black box mechanism, cannot provide explanatory recommendation reasons strongly related to knowledge in the travelling field for users or platform operators, and seriously weakens transparency and user trust of the system. Meanwhile, the model belongs to a pure data-driven architecture, is difficult to systematically integrate with the structured field knowledge of the tourism industry in the technical aspect, and cannot establish the deep association of 'user psychological motivation-tourism product core characteristics-recommendation decision logic', so that the reasoning capacity of the model in a data sparse scene is obviously limited. In addition, the user preference and the domain knowledge learned by the model are both solidified in the network parameters, so that the model is difficult to carry out agile adjustment according to the real-time feedback of the user or the trend change of the external tourism market, the model iteration can be completed only by means of periodic batch retraining, and the response agility to the market and the user demand change is seriously insufficient. Based on the above, the scheme provides a full-link intelligent travel recommendation system and method integrating a knowledge graph and a Large Language Model (LLM) so as to solve the technical pain point in a targeted manner. Disclosure of Invention In view of the above, the technical problem to be solved by the present invention is to provide a method and a system for intelligent recommendation of travel by integrating knowledge maps and large models, so as to solve the following technical problems: (1) The method solves the problem that modeling of deep psychological motivation and essential requirements of users is insufficient in the prior art, an existing double-tower model only builds implici