Watchat

Chat with Watchat!
Explore new movies!
Request similar movies!

Inspiration

Not knowing what to watch on your streaming provider is an everyday problem, despite many thousands of choices. Scrolling through hundreds of movies and series can be tiering and overwhelming. We want to tackle this problem by implementing an interactive, minimalistic and fun interface with a chat-like environment. Our highest goal for this project was to make the users experience as effortless as possible.

What it does

Watchat recommends movies based on free text inputs and small user choices. The text inputs are extracted from the chat-like format, where Watchat asks the user a question, some of which seemingly not related to movies at all. It is also possible to send text by yourself. After analysing the response, Watchat will pose some movies as a recommendation. The user is now able to look up details on the movies, redirect to a movie or select a movie and ask for similar ones. When doing the last, Watchat will pose improved recommendations by combining past responses of the user with the new input. By repeating this process in a completly free manner, Watchat can improve its recommendation further while still being responsive towards change and open for exploration.

How we built it

The core of Watchat is made possible by combining NLP and Neareast Neighbour Approximization. First we use NLP to form numerical vectors, consisting of similarity values of the movie description and different tags, and some additional information. After precomputing this vector for each movie, we save the movies in a graph-like manner, where neighbours to a movie are the nearest movies available. Now we form the same vector by analysing the users text input. Next we can approximate the nearest neighbour to this vector by traversing through the graph. The approximization can futher improve by considering movie choices and refining the distance function.

Challenges we ran into

We ran into many kinds of challenges, but we felt that coming up with the right representation for movies, preparing the data and refining our classification and matching criteria were especially hard.