RandomAnime x Rekku

This site proudly uses Rekku's machine learning API. I've had the pleasure of working with the gentlemen listed below that put in the work to make an anime recommendation engine for everyone to love.

Christopher Kok Ignacio Raposo Oliver De Jesús Rosario Reyes Joey Wong Joel Wong

What is Rekku?

Rekku is an anime recommendation system based on a variety of different machine learning algorithms. The goal is to provide understanding to the user in terms of how their recommendations are produced using data driven models. It's definitely a work in progress and we're developing on features that will allow for automated learning from your participation as we speak!

Recommendation Types

—Best

We train a linear regression model using multiple different factors (genre similarity, popularity, ratings, etc.) to score and rank every other anime in terms of their similarity based on manual reviews by other humans on MyAnimeList.

—Characters

This uses the very detailed character tags from Anime-Planet to make recommendations. We use the tags of an anime's main character and finds similar characters in other shows which it returns as the recommendations.

—Reviews

This takes the top reviews of an anime and uses the Skip-Gram model implemented in the Gensim library to create vector representations of all the words in the review, then it averages them to create a single vector representation of the reviews of the anime. To make a recommendation, we take the review embedding of an anime and use the cosine distance metric to find it's closest neighbors which are returned as the recommendations.

—Trending

This uses anime with similar genres and ranks them by their release date using a unsupervised soft clustering algorithm. Hence, anime in the latest seasons would be prioritized first.

—Score

This takes anime with similar genres and ranks them by their MyAnimeList scores using a unsupervised soft clustering algorithm. Hence, anime with the higher ratings overall would be prioritized first.

—Synopsis

With the synopsis of an anime, we use the bert-base-nli-mean-tokens model from the SentenceTransformers library to create a sentence embedding for the synopsis of the anime. To make a recommendation, we take the synopsis embedding and use the cosine distance metric to find it's closest neighbors which are returned as the recommendations.

Questions?