It's a Tabelog ML Project. Explores Japanese restaurant data with collaborative filtering and Bayesian modeling.

This application uses item-item collaborative filtering to analyze and predict missing ratings from users on Tabelog (食べログ), a popular Japanese restaurant review platform.

Collaborative filtering is a technique used in recommendation systems. It works by finding patterns in how users rate different items or categories and uses them to predict missing ratings.

Bayesian modeling uses numerical features of restaurant ratings such as engagement and price ranges. Using the data, it uses prior evidence and updates the predictions accordingly, since it's used to find probability of a hypothesis.

1. Select a restaurant from our database or provide a Tabelog URL
2. Run collaborative filtering to impute missing ratings
3. Explore visualizations of the rating data
4. Get insights about what the restaurant is best rated for

Tabelog ratings include:

• 料理・味 (Food and taste): Quality and taste of the food
• サービス (Service): Quality of the service
• 雰囲気 (Atmosphere): Restaurant ambiance and setting
• CP (Cost Performance): Value for money
• 酒・ドリンク (Alcohol and drink): Quality of alcoholic and non-alcoholic beverages
• Overall Rating: Reviewer's overall rating of the restaurant (independent of all other categorical ratings)

1. Clone this repository and download all the packages from requirements.txt
2. Start the venv source /venv/bin/activate (path to your venv)
3. Run python3 -m streamlit run tabelog_cf_app.py
4. See console and copy the local host link and open it in your browser or let the app open automatically

Navigate to the "Restaurant Analysis" page to begin exploring restaurant ratings!

If you'd like to add your own restaurants, visit https://tabelog.com/en, and pick out restaurants. **Before adding them, remove the "en" part from the link. Example: https://tabelog.com/en/tokyo/... -----> https://tabelog.com/tokyo/...

Here are some examples restaurants you can try entering on the app:

• https://tabelog.com/hyogo/A2801/A280109/28071150/ - A ramen shop in Hyogo
• https://tabelog.com/tokyo/A1302/A130201/13215419/ - A brewery in Tokyo
• https://tabelog.com/okayama/A3301/A330101/33015797/ - A pizza place in Okayama

• Content-Based Filtering might've been a much better fit instead of (Item-Item) Collaborative Filtering since there were more than type of rating and there were daytime / nighttime prices.
• Collaborative Filtering tended to compute higher-than-expected ratings which didn't always reflect the overall consensus due missing values being imputed then used.
• In a different approach, better methods like Content-Based Filtering would be used to properly predict ratings and explore (Japanese) restaurant data.

• Automatically parse Tabelog URLs to get the Japanese webpages only (or adjust the scraper for the English site).
• Use Content-Based Filtering.
• Cleanup and document the data collection and processing.
• Could use an NLP-based solution such as Word2Vec (see https://qiita.com/toshiyuki_tsutsui/items/19590b464f15f845efcd)

• https://qiita.com/toshiyuki_tsutsui/items/f143946944a428ed105b
• https://www.takapy.work/entry/2019/01/03/134720
• https://myfrankblog.com/find_best_restaurant_from_tabelog/