CineWatch is a Flask web app that recommends films similar to one you pick.
How it works #
- Each film’s genres, keywords and plot overview are joined into one text.
- scikit-learn turns that text into TF-IDF vectors, capped at 5,000 features.
- The chosen film is compared with every other film by cosine similarity, and the closest matches are returned.
The data comes from the TMDB movie dataset (v11). The processed TF-IDF matrix is cached as a sparse matrix, so it is not rebuilt on every start.
Features #
- Search with suggestions.
- Filtering by genre.
- A watchlist.
- Film details from the TMDB API when an API key is set: cast, directors, age rating and trailer.
Running it on a free plan #
The hosted demo on Render loads a 1,000-film sample instead of the full dataset, to stay within the free plan’s memory.
What I learned #
Content-based filtering only understands what a film is about, not what people who liked it went on to watch. Matching on the TF-IDF similarity of genres, keywords and plot gives sensible neighbours, but it can’t surprise the way collaborative signals can, and a film with a thin overview gets thin recommendations. It was a clear lesson in fitting the method to what the data can actually support, and in where a content-based approach stops being enough.