Tabelog ML

For my Machine Learning 2 course, my partner and I had to show two different modeling approaches, and we built both to evaluate data from Tabelog, Japan’s restaurant-review platform.

Data collection and datasets

I developed a program that scraped per-restaurant reviews from Tabelog and saved data across six rating categories. The six categories are food and taste, service, atmosphere, cost performance, alcohol and drink, and an overall rating that is independent of the other five. We also assembled a curated set of nearly 3,000 restaurants with features such as review count, photo count, day and night price ranges, and like and bookmark counts.

Recommendations with Bayesian models

We fit a Bayesian linear model and a Beta regression in R. Given a user-written review of the features they want, the Beta model returns the five restaurants most similar to that request. This was the model and approach that recommended places based on a user’s review.

Filling in missing ratings with collaborative filtering

The other approach to the data was item-item collaborative filtering, which filled in missing category ratings. It is pure math with no training or special layers. This approach worked but the main issue was that cosine similarity over the rating dimensions tends to over-inflate the imputed values since many of the ratings across the five categories were missing. This happened because the overall rating is the only required rating on Tabelog. Since we needed to demonstrate our understanding of the concepts, we used what we learned to fill in the missing ratings. We flagged that other approaches, such as content-based filtering or an NLP-based solution, might’ve been better in the writeup instead of hiding it.

Interactive application

The collaborative filtering code was shipped as a Streamlit app that accepts a Tabelog URL, scrapes the restaurant’s reviews, and shows whether the place is recommended along with charts. The app also saves restaurants to use later. A lot of the effort went into the scraping.

The live demo rebuilds the collaborative filtering step in the browser over a static snapshot of the data. There’s no scraping nor server involved.