moviemap

What is moviemap?

• Visualizing and exploring data points is a great (fun) way to get insights about data.
• What if we can make individual movie into single representation?
• If the representation quality is good enough similar movies would be clusttered together, so that recommending movie would become a trivial thing.

How it works

• Scrape movie list from movie review web site, Watcha
  • Get 5200 + rated movies by critic Lee Dongjyn.
  • Features processed: ["title", "plot", "date", "genere", "director", "country", "user_rating", "critic_rating"]
• Make representation vector for each movie
  • Using open source Korean NLP model (KoBERT), make plot into single sentence vector
  • Features describing movies are used together [director, date, genere, country, user_rating]
  • Train movie rating prediction model (forward path: bottom to top)
    • [Loss] (output, critic rating)
    • [MLP]
    • [plot vector] + [genere embs concat] + [director emb] + [country emb] + [date emb] + [user rating]
    • [koBERT (partially freezed)]
    • [tokenized plot text]
  • After training, get representation vector from hidden dim of MLP model
  • Reduce dimensionality using t-SNE (hidden_dim -> 2 dim)
• Make interactive plot on Web
  • Used Nomic AI's deepscatter library (https://github.com/nomic-ai/deepscatter) for efficiency and speed.

Test

• You can see the result on this page (Still WIP)
• https://scatterfilm.web.app