You can check out the live demo of Gróa here.
Michael Gospodinoff | Gabe flomo | Jeff Rowe | Coop Williams | Eric Smith |
---|---|---|---|---|
|
|
|
|
|
Trello Board We use Trello as a quick wireframe tracker through the first stages of development. As the product moves past releases 1.1 and 1.2 we will transition away from Trello and into the git ecosystem entirely.
Product Canvas This notion document serves as a solid resource if you want to learn more about our motivations for creating this product and the general development direction it is taking.
The companies which have the resources to create an elegant Movie Recommendation Engine have a profit motive aligned with recommending high cost proprietary content rather than films their customers would genuinely enjoy.
Existing web sites geared towards providing a recommendation rely entirely on basic rating models which are weighted heavily towards popular films and generally do a poor job identifying unique outliers.
Groa combines the public data available on IMDb with tried-and-true recommendation techniques to provide a user-driven movie discovery experience. We use two similar language embedding models to acheive this.
-
We trained Word2Vec on positive user ratings histories to create a user-based collaborative filtering recommender. The algorithm embeds over 97,000 movie IDs into a 100-dimensional vector space according to their co-occurence in a user's positive ratings history. The ID for each movie is a key for its vector, which can be called from the model and compared with any other vector in that space for cosine-similarity. To provide recommendations given a new user's watch history, we simply find the vector average of the user's choice of "good movies" and find the top-n cosine-similar vectors from the model. We can improve the recommendations by subtracting a "bad movies" vector from the "good movies" vector before inferencing. Models trained in this way can be tested by treating a user's watchlist (unwatched movies saved for later) as a validation set.
-
The above model fulfills most requirements for a general-purpose movie recommender system. However, it is unable to make riskier recommendations for movies that a majority of reviewers do not enjoy (cult movies). To satisfy users who seek underrated movies, we also trained Doc2Vec on user review histories to create a review-based collaborative filtering model. This model does not recommend movies, but finds reviewers who write similarly to a new user. We then query the review database for positive reviews from these users, both in cases where the ratings count is 1k-10k (hidden gems), and where the reviewer rates a movie 3 stars more than the average.
The lightning-fast inferencing of the Word2Vec/Doc2vec algorithms allows us to incorporate user feedback into progressively updating recommendations. If the user elects to approve or disapprove of a movie, its corresponding vector is added to, or subtracted from, the user's overall taste vector. Weighting these feedback vectors by a factor like 1.2 increases the influence of that feedback on the user's taste vector, and this factor can be tweaked to change the effective "learning rate" of the re-recommendations process.
- EC2
- S3 Bucket
- RDS
- Elastic Beanstalk
Based on the user's movie ratings and reviews, provide recommendations for movies to watch that they have never before considered watching. We can do this by vectorizing the user's Letterboxd or IMDb reviews and finding cosine-similar matches from 22GB worth of movie reviews. Results can be filtered to remove movies the user has already watched, so long as they provide their data exported from one of those sites.
Our primary sources of data are the user reviews of IMDb.com and Letterboxd.com. This was collected from these websites with our own custom-built web scraper. In addition, we make use of IMDb's data files for summary information on films.