MyAnimeTaste
Background
I've been using the anime-rating website MyAnimeList since I was in middle school, but I've always wanted a better way to view my friends' favorite anime. My project was my effort to build a website that would allow users to explore aspects of their friends' tastes better. This project involved two parts:
Researching the possible ways to generate recommendations for anime given a list of users.
Creating the front- to back-end of the website.
Creating the website
For the backend, I created a pipeline that fetches data from MyAnimeList’s REST API. For the frontend, I used React.js with Material UI. In progress code can be found here.
Generating recommendations
To explore methods for generating anime recommendations, I used a static dataset from Kaggle retrieved from the MAL API on a specific day.
The reviews matrix, where each value is the score a user (row) has given an anime (column)
I used Collaborative Filtering with Matrix Factorization to find new anime most similar to a user's existing list and return a list of recommendations. Matrix factorization is the idea of representing the above matrix as a product of two orthogonal matrices. It is used to predict matrices based on the idea that a user's preferences can be determined by a number of hidden features. Since certain features are shared between users and anime, matrix factorization is able to predict a user's rating of an anime with respect to its similarity to a user's preference and interactions with other anime.
After matrix factorization, all missing values are filled with predicted scores a user would have given the anime
The resulting recommendation algorithm outputs the anime, sorted from highest to lowest predicted score, as its recommended list. However, matrix factorization is slow on sparse and large matrices, and may predict ratings that are outside the range of values given (0-10).