news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Now, the big model can be a private tour guide for youPlan your Citywalk route——

Jointly launched by HKU, MIT and other institutionsITINERA, combining LLM with space optimization to achieve personalizedOpen Domain City Itinerary Planning。

For example, the user inputs "Plan a citywalk route for me that includes 'Ju Fu Chang' and ends at Jing'an Temple".

The ITINERA system immediately generated a route including several locations and provided corresponding introductory texts.

ITINERA can even understand and meet personalized demands such as "a bar suitable for couples to go together", "a holy place for the second dimension", and "a popular check-in spot along the way".

You may not feel it when you look at it alone, let's compare ITINERA (left) andGPT-4 CoTGenerated routes.

Same prompt: "I want a literary route, passing by bridges and ferries."

can be seen,ITINERA generated itineraryIt will pass several bridges along the Suzhou River and the Huangpu River ferry, and end at the literary Duoyun Bookstore. The route is relatively reasonable, concentrating the locations in two spatial clusters.

In the right figure, the POIs (personal preferred points of interest) selected by GPT are consistent with the bridge and ferry requested by the user.Not quite consistentIn addition to this example, GPT sometimes generates hallucinations and generates non-existent POIs.

In summary, ITINERA hasThe following features：

• Dynamic information: real-time updates of POI and current popular activities
• Personalization: Prioritize personal preferences rather than just popular attractions
• Diverse constraints: Flexibly respond to complex and diverse user needs
• Spatial intelligence: Combined with spatial optimization algorithms to ensure reasonable and efficient routes

ITINERAFour citiesThe training and evaluation were performed on a travel itinerary dataset (1233 popular city routes, 7578 POIs).

The results show that it can produce better results than traditional itinerary planning, direct use of LLM and other methods.

Currently, the relevant papers have been included in the KDD Urban Computing Workshop (UrbComp) 2024.

Five modules make up ITINERA

The next question is: How is this done?

As shown in the figure below, ITINERA consists ofFive modules driven by big modelscomposition.

First, User-owned POI Database Construction（UPC）The module collects and constructs travel content from social platformsUser interest point database。

In order to plan a trip that meets the user's request, Request Decomposition（RD）The module interprets and organizes user preferences and converts them into structured data form.

Preference-aware POI Retrieval（PPR）The module will search based on user preferences and obtain the most relevant points of interest.

To ensure that the itinerary is spatially coherent, the author uses Cluster-aware Spatial Optimization（CSO）Module,spatially filters and ranks the retrieved POIs by solving the hierarchical,traveling salesman problem.

Finally, Itinerary Generation（IG）The module combines a set of candidate POIs with multiple constraints and uses a large model to generate travel routes and related descriptions that are both spatially reasonable and meet user requests.

Now that we understand the principle, how does ITINERA actually perform?

To clarify this issue, the authors collected a travel itinerary dataset from four cities, including user requests, corresponding city itinerary routes, and detailed point of interest (POI) data.

The evaluation is carried out through objective indicators such as POI recall rate (RR), the difference between the total distance and the theoretical shortest path (AM), the number of intersections in the route (OL), and the proportion of unknown POIs (FR).Accuracy of personalized POI recommendations、Matching degree with user request,as well asSpatial rationality of generated routes。

Even to solve the problems of attractiveness of points of interest, matching of user requests, etc.UnquantifiableTo solve the problem of traffic congestion, the authors also used LLM to automatically evaluate the quality of POIs, the quality of routes, and the degree of match between itineraries and user requests.

It can be seen that compared with other methods such as GPT-3.5, GPT-4 and GPT-4 CoT, the ITINERA system has a higher performance in all indicators.Better performance。

existUser and expert evaluationThe ITINERA system also received higher scores in terms of POI Quality, Itinerary Quality and Match.

In generalITINERA can generate personalized and spatially coherent citywalk itineraries directly from natural language requests. It not only explores the open domain itinerary planning problem in the era of big models, but also provides ideas for using big models to solve complex space-related problems in urban applications.