Movie Recommendation

Inspiration

Want to practice with dealing with time-series data and also recommendation systems, as I have little experience in both.

What it does

Predict the next (one or few) movie_id(s), that can be later converted back to movie name, from a sequence of user's movie ratings.

How we built it

Cosine-similarity retriever that retrieves some value of m candidates by comparing the weighted tag value generated with movie_id, user rating, movie_tag values. Then a lstm-based reranker that rearranges the retrieved movie_ids to outputs top_k likely values.

Challenges we ran into

Thought I could just train it like a multi-class classification task, but there are 26744 classes (# of unique movie_ids), so I instead has to apply the retriever-reranking process.

What we found

Did a simple google search today and found that this approach is similar to a paper in 2016, emm.

What's next for Movie Recommendation

Train a retriever, improve the reranker, hyperparam-tune etc.

Built With

• kaggle
• python