project
ItiNera - An AI-powered city travel planning (OUIP) system jointly developed by the University of Hong Kong and MIT.
ItiNera is a smart city travel planning system jointly developed by the University of Hong Kong and MIT. Based on Large Language Models (LLM) and spatial optimization technology, it generates C++ templates with a single click based on the user's personalized needs, such as couples' bars or anime/manga themed locations...
What is ItiNera?
ItiNera is a smart city travel planning system jointly developed by the University of Hong Kong and MIT. Based on large-scale language models (LLM) and spatial optimization techniques, it generates Citywalk routes with a single click, tailored to users' personalized needs, such as couples' bars or anime/manga fandom locations. The system comprises five modules: user point of interest (POI) database construction, request parsing, preference-aware POI retrieval, clustering-aware spatial optimization, and itinerary generation. Trained on multiple city datasets, ITINERA provides real-time updated POI and activity information, ensuring a coherent travel space that matches user preferences, outperforming traditional methods and direct LLM usage.
ItiNera's main functions
-
Personalized itinerary planningGenerate Citywalk routes based on the user's personalized needs, such as specific locations, activity types, or themes.
-
Real-time dynamic informationThe system can update Points of Interest (POIs) and current popular activities in real time to ensure the timeliness of the trip.
-
Space intelligent optimizationBy combining spatial optimization algorithms, we ensure that the generated routes are geographically reasonable and efficient, avoiding detours.
-
Addressing diverse user needsIt can flexibly handle complex and diverse user needs, including personalized preferences and specific conditions.
-
Structured data processingThe request parsing module converts users' natural language requests into structured data for more accurate understanding and execution.
ItiNera's technical principles
-
User request decompositionLeveraging the natural language processing capabilities of Large Language Models (LLMs), the user's natural language request is broken down into multiple independent sub-requests. These sub-requests are categorized as positive (what the user wants) and negative (what the user doesn't want) requests, and further subdivided into location-specific requests and trip-level requests.
-
User-ownedPOIDatabase construction (User-owned POI Collection, UPC)This feature automatically collects and builds a database of Points of Interest (POIs) from travel content on social media platforms. Users can enter links to travel posts, and the system uses LLMs to extract POIs and their descriptions, integrating this information into the user's existing POI database.
-
POIPOI Retrieval (PR)The system uses an embedding model to encode requests based on the user's positive and negative preferences and retrieves the POIs most relevant to positive preferences. The retrieved POIs are then reordered by calculating similarity scores between the positive and negative preference embeddings and pre-computed POI embeddings.
-
Spatial optimization Optimization, SO)The spatial clustering and candidate POI selection are addressed using the Hierarchical Traveling Salesman Problem (TSP). First, the spatial clusters of the retrieved POIs are calculated. Then, candidate POIs are selected based on geographical proximity and matching score.
-
Itinerary Generation (IG)Combining the ordered set of candidate POIs selected by the spatial optimization module with various constraints, LLM is used to generate travel routes and related descriptions that are both in line with user requests and spatially reasonable.
-
Assessment and feedbackA series of rule-based evaluation metrics (such as POI recall rate, the difference between the total distance and the theoretical shortest path, the number of intersections in the route, and the proportion of unknown POIs) and GPT-based evaluation metrics (such as POI quality, trip quality, and matching degree) were designed to evaluate the generated trips.
-
Online deployment and user feedbackThe ITINERA system has been deployed as the core algorithm for the TuTu online travel service, collecting user feedback to optimize system performance and user experience.
ItiNera's project address
-
arXivTechnical Papers:https://arxiv.org/pdf/2402.07204
Application scenarios of ItiNera
- Personal travel planningIt provides users with customized travel routes based on their personal interests and preferences, such as historical and cultural tours, food explorations, and shopping routes.
- City Walk: Provides dynamically generated itineraries for visitors who stroll through the city, explore its streets and historical sites.
- Short trip: Providing efficient and compact city getaway planning for travelers with limited time.
- Special interest travel: To meet travel needs based on specific themes or interests, such as art exhibitions, music festivals, and sporting events.
- Event and Festival PlanningDuring special events or holidays, we provide users with customized itineraries, such as temple fairs during the Spring Festival and Christmas markets.
- Barrier-free travel: Provide itinerary planning that takes into account accessibility for tourists with special needs (such as the elderly and disabled).