ITINERA Revolutionizing Citywalk Itinerary Planning with Large Models

Nowadays, large models can act as personal tour guides and plan Citywalk routes for you -

HKU and MIT jointly launched ITINERA, combining LLM with spatial optimization to achieve personalized open-domain urban 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 generates a route with several locations and provides corresponding introduction texts. Even personalized needs such as a bar suitable for couples, anime and manga holy land, and passing by internet-famous check-in points, ITINERA can also understand and meet.

Just looking at it might not feel much, let's directly compare the routes generated by ITINERA (left side) and GPT-4CoT.

Same prompt: I want an artistic route that passes through bridges and ferries.

It can be seen that the itinerary generated by ITINERA will pass through several bridges along the Suzhou River and the Huangpu River ferry, and end at the artistic Duoyun Bookstore, and the route is more reasonable, concentrating the locations in two spatial clusters.

While in the right picture, the POIs (points of interest) selected by GPT are not quite in line with the user's requirement of bridges and ferries, and there are also situations such as detours and POIs being too far apart. In addition to this example, GPT sometimes has hallucinations and generates non-existent POIs.

In general, ITINERA has the following features:

Dynamic information: Real-time update of POIs and current popular activities

Personalized customization: Giving priority to personal preferences rather than just considering popular attractions

Diverse constraints: Flexibly responding to complex and diverse user needs

Spatial intelligence: Combining with spatial optimization algorithms to ensure a reasonable and efficient route

ITINERA was trained and evaluated on the travel itinerary dataset of four cities (1233 popular city routes and 7578 POIs). The results show that it can generate better effects than traditional itinerary planning and directly using LLM and other methods.

Currently, the relevant paper has been included in the KDDUrbanComputingWorkshop (UrbComp) 2024.

Five Modules Composing ITINERA

The next question is: How does it achieve that?

As shown in the figure below, ITINERA is composed of five modules driven by a large model.

First, the User-ownedPOIDatabaseConstruction (UPC) module collects and constructs a database of user points of interest from the travel content on social platforms.

In order to plan an itinerary that conforms to the user's request, the RequestDecomposition (RD) module interprets and organizes the user's preferences and converts them into structured data forms.

The Preference-awarePOIRetrieval (PPR) module will retrieve the most relevant several points of interest according to the user's preferences.

In order to ensure the itinerary is spatially coherent, the author adopted the Cluster-aware SpatialOptimization (CSO) module to spatially filter and sort the retrieved points of interest by solving the hierarchical traveling salesman problem.

Finally, the ItineraryGeneration (IG) module combines the candidate point of interest set with multiple constraints and uses a large model to generate a travel route and related description that is both spatially reasonable and in line with the user's request.

After understanding the principle, how does ITINERA perform in practice?

In order to figure this out, the author collected the travel itinerary dataset from four cities, including user requests, the corresponding urban itinerary routes, and detailed point of interest (POI) data.

Through objective indicators such as the recall rate of POIs (RR), the difference between the total distance and the theoretical shortest path (AM), the number of intersection points in the route (OL), and the proportion of unknown POIs (FR), to evaluate the accuracy of personalized recommendation of POIs, the matching degree with the user's request, and the spatial reasonableness of the generated route.

Even to address the unquantifiable issues such as the attractiveness of points of interest and the matching degree of user requests, the author also used LLM to automatically evaluate the quality of POIs, the quality of the route, and the matching degree between the itinerary and the user's request.

It can be seen that compared with other methods such as GPT-3.5, GPT-4, and GPT-4CoT, the ITINERA system performs better in all indicators.

In user and expert evaluation of the quality of POIs (POIQuality), the quality of the itinerary (ItineraryQuality), and the matching degree (Match), the ITINERA system also obtained higher scores.

In general, ITINERA can directly generate personalized and spatially coherent Citywalk itineraries from natural language requests, not only exploring the open-domain itinerary planning problem in the era of large models, but also providing ideas for using large models to solve complex spatially related problems in urban applications.

For more details of the method and experimental result s, welcome to read the original text.

Paper:

https://arxiv.org/abs/2402.07204