Mine & Dine Combining Collaborative Filtering and Social Media to Make Better Restaurant Recommendations

Currently, to find a restaurant recommendation, a user may go online and visit one of many popular review-aggregation websites (Yelp and TripAdvisor are two popular examples). Rather than explicitly make recommendations, however, these sites present the user with an expansive list of local options and their average ratings- calculated from members ’ reviews. These sites lack personalization of restaurant ratings for in-dividual users, and the sites ’ review pools are often small, noisy, and unrepresentative of the true quality of the restau-rants they describe. We describe a system that seeks to correct these shortcom-ings. Mine & Dine is a restaurant recommendation engine that mines web-based social media to make unique recom-mendations based on other restaurants the user has already visited. The user enters as input a variable number of restau-rants he or she has already visited in the area, and then ranks each one on a real number scale from 0 to 5. The system parses through web logs (or “blogs ” for short) and tags those that qualify as reviews of local restaurants. From these re-views, item-based collaborative filtering is employed to pre-dict the user’s ratings of other restaurants in the area and rank them. The highest-ranked restaurants are then recom-mended to the user. To evaluate our recommendations, we calculate the mean ab-solute percentage error (MAPE) of a users ’ predicted restau-rant ratings in test data. Comparing Mine & Dine’s per-formance to recommendations based on average restaurant ratings (the baseline), we have achieved a significant 10% reduction in MAPE