CN-122021976-A - Stroke planning system and method based on large model

CN122021976ACN 122021976 ACN122021976 ACN 122021976ACN-122021976-A

Abstract

The invention relates to the technical field of journey planning and discloses a journey planning system based on a large model, which comprises an involuntary behavior recognition and preference purification module, a dynamic user state characterization module, an enhanced user portrait construction module, a journey generation and multi-constraint feasibility verification module and a journey scheme interactive optimization and user feedback closed-loop module, wherein the involuntary behavior recognition and preference purification module outputs a purified behavior sequence reflecting real preference of a user, the dynamic user state characterization module maps high-dimensional heterogeneous user features to a low-dimensional continuous potential space, the enhanced user portrait construction module integrates regional culture priori knowledge and external ecological behavior signals, and the journey generation and multi-constraint feasibility verification module based on the large model is further included. The method and the system promote the authenticity and feasibility of the trip planning, enhance the user portrait accuracy by purifying the non-autonomous behavior data, integrate the regional culture and the external interest signals, promote the recommendation suitability, combine the dynamic budget and the real-time resource verification guarantee scheme, can be executed, support the user interaction optimization and feedback closed loop, and realize the personalized, practical and participatable intelligent trip service.

Inventors

JIANG LIHUI
LIU HAO
WANG QINTING
WANG ZAORONG

Assignees

浙江深大智能科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251224

Claims (10)

1. A large model-based trip planning system, comprising the following modules: the non-autonomous behavior recognition and preference purification module is used for carrying out fusion analysis on the multi-source context signals, recognizing the order records forced to be generated due to external constraint, correcting, reducing the weight or eliminating the trip identification according to the order records, and outputting a purified behavior sequence reflecting the real preference of the user; The dynamic user state characterization module is used for constructing dynamic characterization reflecting the current travel state of the user and mapping the high-dimensional heterogeneous user characteristics to a low-dimensional continuous potential space; The enhanced user portrait construction module is used for further integrating regional culture priori knowledge and external ecological behavior signals on the basis of the dynamic potential state output by the dynamic user state characterization module and outputting enhanced user portraits; the travel generation and multi-constraint feasibility verification module based on the large model takes the enhanced user image as a core context, drives the large language model subjected to field fine adjustment, and generates a complete travel scheme based on the structural prompt; And the journey scheme interactive optimization and user feedback closed loop module supports the user to carry out fine-granularity interactive adjustment on the journey scheme generated by the large model, and feeds back the explicit operation and implicit behavior of the user to the upstream modeling unit in real time.
2. The large model based trip planning system of claim 1, wherein the non-autonomous behavior recognition and preference cleansing module comprises: The multi-dimensional alternative path detection unit is used for constructing a cross-dimensional intention and intersection consistency measurement model and quantifying the deviation degree between the preference actively expressed by the user and the final intersection result; A context urgency recognition unit recognizing a forced reservation behavior under the combined action of time pressure and resource scarcity; the group behavior abnormality sensing unit is used for identifying resources which are avoided, canceled or selected only under alternative logic by a large number of users by utilizing platform-level group behavior statistical information and assisting in judging whether individual selection is non-autonomous or not; And the self-adaptive preference correction unit is used for merging the non-autonomous evidences output by the three units, introducing holiday context signals, calculating the non-autonomous probability of each record through a learnable probability model, and executing a layered purification strategy according to the non-autonomous probability.
3. The large model-based trip planning system according to claim 2, wherein in the multi-dimensional alternative path detection unit, an active behavior log of a user before order generation is extracted, mapped into vector representation through a semantic encoder to form distribution of the user in an intention space, and for structural decision dimensions strongly related to trip planning, inconsistent scores between intersection values and user intention distribution are calculated respectively to obtain alternative path intensities; the structured decision dimension includes destination type, hotel style, length of travel, number of peers, and traffic pattern.
4. The large model-based trip planning system according to claim 2, wherein in the context urgency recognition unit, two key features of a reservation advance period and a resource price standardization deviation degree are calculated, and feature vectors composed of the two are input into a pre-trained isolated forest model to output an anomaly score as a context urgency signal.
5. A large model based trip planning system according to claim 2, characterized in that in the group behaviour anomaly perception unit, the system identifies two anomaly metrics, comprising: The cancellation rate of the scenic spot under the specific space-time granularity is judged whether to be obviously higher than the average value of similar resources; And detecting from a system background log whether the alternative option is automatically recommended by the system due to the fact that the preferred resource is not available or not.
6. The large model based trip planning system of claim 2 wherein in the adaptive preference correction unit, the hierarchical purification strategy is: if the non-autonomous probability is greater than the high confidence rejection threshold, considering the order record as passive selection, and thoroughly rejecting the order record from the user history sequence; if the low confidence retention threshold value is less than or equal to the non-autonomous probability and less than or equal to the high confidence rejection threshold value, the recorded part is considered to be interfered, but still has reference value, and soft correction is executed; if the involuntary probability is less than or equal to the low confidence retention threshold, the record is considered to reflect the real preference, and the original tour identification is retained unchanged.
7. The large model based trip planning system of claim 1, wherein the dynamic user state characterization module comprises: The multi-source user state feature coding unit is responsible for fusing heterogeneous information from different data sources into a unified high-dimensional user state feature vector, and is used as an input basis of potential space modeling; The variation potential state space modeling unit is used for compressing the high-dimensional feature vector into a low-dimensional continuous potential space based on a variation self-encoder architecture; the sudden environment disturbance dynamic modulation unit solves the problem of the practical adaptability of the potential state in the open world; And the on-line state evolution and feedback calibration unit is used for deploying a lightweight on-line learning pipeline to enable the user state to continuously evolve along with the new behavior.
8. The large model based trip planning system of claim 1, wherein the enhanced user representation construction module comprises: The regional culture mode embedding unit encodes culture characteristics of administrative division levels into computable vector representations and carries out self-adaptive fusion with potential states of users; The cross-platform weak signal alignment unit is accessed to an aggregation level behavior statistical interface provided by a partner, acquires weak signal data of a passenger origin ground level on the premise of not acquiring individual identity information of a user, and fuses the weak signal data to a user potential state space to expand an interest perception boundary; and the multi-granularity interest evolution modeling unit is used for finely modeling the time dynamic characteristics of the user interests and distinguishing the long-term stability preference from the short-term fluctuation interests.
9. The large model based trip planning system of claim 8, wherein the large model based trip generation and multi-constraint feasibility verification module comprises: the structural prompt construction and dynamic budget injection unit decodes key trip intention from the enhanced user image, and converts user state, external environment and business constraint into a large-model resolvable structural prompt; The large model travel generation and semantic consistency control unit is used for calling a large language model subjected to fine adjustment in the travel field and outputting a standardized JSON format travel scheme; And the multisource real-time feasibility verification and dynamic rollback unit is used for calling an external tool interface and executing end-to-end feasibility verification on the generation scheme.
10. A large model based trip planning method based on a large model based trip planning system according to any one of claims 1-9, characterized by the following steps: the method comprises the steps that firstly, an involuntary behavior identification and preference purification module identifies and purifies involuntary selection behaviors in historical tour data of a user, and a purification behavior sequence reflecting real preferences is generated; Step two, a dynamic user state characterization module constructs dynamic potential state characterization of the user based on the purification behavior sequence; Step three, an enhanced user image construction module constructs an enhanced user image on the basis of dynamic potential state representation; Step four, based on the travel generation and multi-constraint feasibility verification module of the large model, taking the enhanced user image as a core context, driving the large language model to generate and verify a travel scheme; And fifthly, supporting the interactive optimization of the travel scheme by the user through the travel scheme interactive optimization and user feedback closed loop module, and returning feedback to the user state modeling unit in real time.

Description

Stroke planning system and method based on large model Technical Field The invention relates to the technical field of journey planning, in particular to a journey planning system and method based on a large model. Background The travel planning system based on the large model is an intelligent decision system integrating historical behavior portraits of users, multi-source real-time environment data and large language model reasoning capacity, and aims to provide a travel scheme which is highly personalized, logically consistent and has practical feasibility for the users. Current large model-based trip planning systems generally assume that the user's historical trip record is a direct representation of its real preferences, however, this premise is severely distorted in reality in that the user's historical order often contains a large number of non-autonomous selection actions, i.e., sub-optimal schemes that are forced to accept by external constraints, rather than an expression of active willingness, typical scenarios include limiting or reserving full volume due to hot spots, the user having to turn to alternative destinations that would otherwise be uninteresting; the system is forced to accept accommodation options with significantly higher prices and disfavored positions or facilities due to the temporary full house or the rising price of a target hotel, or in holiday peak time, the number of the same party is regulated due to traffic and ticketing resources shortage, the journey time is shortened and the like, these passive decisions are represented as false signals in a system record mark, and because the existing system lacks a discrimination mechanism for autonomy selection, these contaminated data are directly used for constructing a user portrait, so that subsequent recommendation is seriously deviated from real intention. Disclosure of Invention The invention aims to provide a journey planning system and a journey planning method based on a large model, which are used for identifying and purifying non-autonomous selection behaviors in historical journey data of a user and constructing dynamic user state characterization reflecting real preference, so that the large model is driven to generate highly personalized and practical journey planning. In order to achieve the above purpose, the present invention adopts the following technical scheme: in one aspect, the invention provides a journey planning system based on a large model, which comprises the following modules: the non-autonomous behavior recognition and preference purification module is used for carrying out fusion analysis on the multi-source context signals, recognizing the order records forced to be generated due to external constraint, correcting, reducing the weight or eliminating the trip identification according to the order records, and outputting a purified behavior sequence reflecting the real preference of the user; The dynamic user state characterization module is used for constructing dynamic characterization reflecting the current travel state of the user and mapping the high-dimensional heterogeneous user characteristics to a low-dimensional continuous potential space; The enhanced user portrait construction module is used for further integrating regional culture priori knowledge and external ecological behavior signals on the basis of the dynamic potential state output by the dynamic user state characterization module and outputting enhanced user portraits; the travel generation and multi-constraint feasibility verification module based on the large model takes the enhanced user image as a core context, drives the large language model subjected to field fine adjustment, and generates a complete travel scheme based on the structural prompt; And the journey scheme interactive optimization and user feedback closed loop module supports the user to carry out fine-granularity interactive adjustment on the journey scheme generated by the large model, and feeds back the explicit operation and implicit behavior of the user to the upstream modeling unit in real time. As a preferred technical solution of the present invention, the non-autonomous behavior recognition and preference purification module includes the following units: The multi-dimensional alternative path detection unit is used for constructing a cross-dimensional intention and intersection consistency measurement model and quantifying the deviation degree between the preference actively expressed by the user and the final intersection result; A context urgency recognition unit recognizing a forced reservation behavior under the combined action of time pressure and resource scarcity; the group behavior abnormality sensing unit is used for identifying resources which are avoided, canceled or selected only under alternative logic by a large number of users by utilizing platform-level group behavior statistical information and assisting in judging whether individual sele