How It Works

Overview

Welcome! This page is a more technical look at how this recommender works. You may find it interesting or useful for understanding why certain shows place well or poorly, it will also cover how AI is used as I want to be transparent in its use.

With that out of the way, lets get into the actual details! The recommender compares anime across five main dimensions, these are Setting/Content, Atmosphere, Characters, Narrative and Themes. Most of those are then broken into even more specific components:

Setting/Content

World Setting - The physical world and environment the story takes place in

Core Content - The activities and events that happen in said world

Atmosphere

Tone - The broad emotions the anime tries to make you feel

Texture - The moment to moment emotions you experience while watching

Characters

Structure - The overall structure of the primary cast

Behavior - The personalities or behaviors of the main and important characters

Relationships - The important relationships between the primary cast

Narrative

Structure - The overall format of the plot/story.

Style - The way the story is actually delivered scene by scene.

Themes

Themes - Themes looks at the ideas/concepts a story explores, how central those ideas are to the anime, and how seriously or directly it engages with them.

Similarity Scores

Every dimension gets a "similarity score" which is a processed score based on each component added together. By default every dimension then gets equal weighting with a slight boost in the anime's "core dimensions." These are the dimensions that are most related to the core viewing experience of an anime. This final score becomes the final "similarity score" between the origin anime and the recommended anime.

These weights can be manually tuned for what you care about though. In the "Advanced Settings" you can alter the amount of weight you want each dimension to carry for the resulting recommendations.

For example, if you liked the Setting/Content of Jujutsu Kaisen, turning that weight up would push anime with similar supernatural settings and exorcist-style combat higher, like Tokyo Ravens, Blue Exorcist,Bleach, and Twin Star Exorcists. You can see an example here.

AI Usage

AI is used in a few parts of the process, so I want to be clear about exactly where it is used and what it does. The recommendations themselves are not generated by asking an LLM for similar anime. All AI generated data is created once ahead of time, stored, and then used by the recommender when you search for something.

LLMs (Large Language Models)

Profile Generation - Every anime in my database gets its own generated "profile" that contains natural language descriptions of each of the components mentioned above. Doing this manually myself for thousands of anime isnt possible for me alone so a LLM is used instead.

Embedding Models

Similarity - The profiles are transformed into embeddings for each component that then can be used by the recommender to estimate how similar two anime are in each dimension. Those similarities are then converted and processed to turn into the final scores and rankings you see on each recommendation page.

Contribution

This leads us to the final part of how the site works, user contributions! The human feedback from users of this site allows me to improve the overall recommendations. I use this data to train the embedding models to understand how to put similar anime closer together and very different ones farther apart. Generally, the more high quality data I have to use, the better the recommendations get for everyone.

If you would like to help, you can contribute your own ratings here!

Questions or Feedback?

If you have any further questions for me please feel free to reach out to me at [email protected] or using the feedback page!