Anime Recommender

Overview

An anime recommendation system built on collaborative filtering — it recommends titles to a user based on the preferences of other users with similar taste, using the MyAnimeList dataset on Kaggle (24,325,191 ratings).

How it's built

The training notebook filters the ratings down to users with more than 200 ratings and anime with at least 50 ratings from those users, builds an anime-by-user pivot table (5523 anime by 41,734 users), and computes item-item cosine similarity across it. The result is pickled into three files the Streamlit app loads at start-up.

The app looks up the typed anime title in the pivot table's index, takes the 5 most similar titles, and renders a card for each with its name, Japanese name, synopsis, genres, and cover image.

Using it

• Type the anime's original English title exactly as released (matching spacing and case, as listed on sites such as hianime.to) — the lookup is an exact match on the title.
• The trained model files are too large for the repository and are hosted in a Google Drive folder.

Status and limitations

Per the README: a misspelled or differently-cased title returns "not found" rather than a fuzzy match, and no evaluation metrics are recorded beyond a single sample recommendation shown in the notebook.

• Demo video