2024/08/02 13:40
The large model becomes a private tour guide, one-click planning Citywalk, jointly produced by HKU MIT
2024/08/02 13:40
Now, the big model can be used as a private tour guide to plan your Citywalk route for you.
HKU MIT and other units jointly launched ITINERA, which combines LLM with space optimization to achieve personalized open-domain urban itinerary planning.
For example, the user typed in "Plan for me a citywalk route that includes 'Huge Wealth' and ends at Jing'an Temple".
The ITINERA system immediately generates a route with several locations and provides the corresponding introductory text.
Even the personalized needs of "bars suitable for couples to go to", "two-dimensional holy places", and "online celebrity check-in points on the way" can be understood and met by ITINERA.
It may not feel like it when you look at it alone, but let's directly compare the routes generated by ITINERA (left) and GPT-4 CoT.
The same prompt: "I want a literary route, to cross bridges and ferries." ”
It can be seen that the itinerary generated by ITINERA will pass several bridges along the Suzhou Creek and the Huangpu River ferry, and end at the Duoyun Bookstore of Literature and Art, and the route is more reasonable, concentrating the location in two spatial clusters.
However, the POI (personal preference point of interest) selected by GPT in the figure on the right does not match the bridge and ferry required by the user, and there are also detours and POI distances that are too far. In addition to this example, GPT sometimes hallucinosizes, generating POIs that don't exist.
In a nutshell, ITINERA has the following features:
ITINERA conducted training evaluations on a dataset of travel itineraries (1,233 popular city routes, 7,578 POIs) in four cities.
The results show that it can produce better results than traditional trip planning, direct use of LLMs, etc.
The paper has been included in the KDD Urban Computing Workshop (UrbComp) 2024.
The next question is: How?
As shown in the figure below, ITINERA consists of five modules driven by a large model.
First of all, the User-owned POI Database Construction (UPC) module collects and builds a database of user points of interest from the tourism content on the social platform.
In order to plan a trip that meets the user's request, the Request Decomposition (RD) module interprets and organizes the user's preferences and converts them into a structured data form.
The Preference-aware POI Retrieval (PPR) module will search for the most relevant points of interest based on the user's preferences.
To ensure that the itinerary is spatially coherent, the authors employ the Cluster-aware Spatial Optimization (CSO) module to spatially filter and arrange the retrieved points of interest by solving the hierarchical traveling salesman problem.
Finally, the Itinerary Generation (IG) module combines a set of candidate points of interest with multiple constraints to generate a travel itinerary and related descriptions that are spatially reasonable and meet the user's request using a large model.
Now that the principle is clear, how does ITINERA actually perform?
To find out this question, the authors collected a dataset of travel itineraries from four cities, including user requests, corresponding city itineraries, and detailed point-of-interest (POI) data.
Through objective indicators such as the recall rate (RR) of POI, 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), the accuracy of personalized recommendation POIs, the matching degree with user requests, and the spatial rationality of the generated route were evaluated.
Even in order to solve the problem that the attractiveness of points of interest and the matching degree of user requests cannot be quantified, the authors also use LLMs to automatically evaluate the quality of POIs, the quality of routes, and the degree of matching of trips and user requests.
It can be seen that compared to other methods such as GPT-3.5, GPT-4, and GPT-4 CoT, the ITINERA system performs better on all indicators.
The ITINERA system also received higher scores in terms of POI Quality, Itinerary Quality, and Match as assessed by users and experts.
In general, ITINERA can directly generate personalized and spatially coherent citywalk itineraries from natural language requests, which not only explores the problem of open-domain itinerary planning in the era of large models, but also provides ideas for using large models to solve complex space-related problems in urban applications.
For more details of the method and experimental results, please read the original article.
Thesis:
https://arxiv.org/abs/2402.07204
This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com