ML · 2025
TripCompassSF — Personalized Itinerary Planner
Generates budget- and time-aware daily itineraries for San Francisco by pairing a learned user–POI satisfaction model with a beam-search planner.
Role: Team of 2
Overview
Generic "top sights" itineraries ignore what a specific traveler actually wants and rarely respect real time and budget limits. TripCompassSF addresses this in two parts: a feed-forward model learns to predict user–point-of-interest satisfaction from engineered features (interest match, popularity, pace × duration), and a beam-search planner uses those predictions to assemble day-by-day San Francisco itineraries that stay within time, budget, and opening-hour constraints. We built a curated SF point-of-interest catalog and a synthetic user–POI interaction dataset to train and evaluate the satisfaction model.
Highlights
- Trained a feed-forward user–POI satisfaction model over engineered features (interest match, popularity, pace × duration).
- Integrated the model into a beam-search planner that assembles feasible itineraries under time, budget, and opening-hour constraints.
- Built a curated San Francisco POI catalog and a synthetic user–POI interaction dataset for training and evaluation.
Tech & topics
- Machine Learning
- Beam Search
- Python
- Recommender Systems