An AI that recommends anime based on personal history. You can use it here: https://discord.gg/h6nTTprMKd.

the files are ordered so that /train is for training files, /deployment is what you need to deploy the model as a discord bot and the root folder has the website and any miscelaneous files that are there by default

───────────────────── architecture ─────────────────────
Input (1D array of shape (1000)) - Conveys which anime has been watched (1 for true, 0 for false)
↓
Embedding (transforms 1D array to a 2D array of shape (1000, 10)) - Infers features about each anime
↓
Dense (transforms 2D array of shape (1000, 10) to 2D array of shape (1000, 32) with ReLU) - Infers features about each anime
↓
Flatten (transforms 2D array of shape (1000, 32) to 1D array of shape (3200) - This allows the next layer to work
↓
Dense (1000 long array of probabilities of each 1000 animes being liked) - Predicts what anime suits the user with softmax
───────────────────── training ─────────────────────
I got all animes a user has watched from the kaggle dataset https://www.kaggle.com/hernan4444/anime-recommendation-database-2020 then corrupted the data - this involved taking out one of the animes. The anime list remaining is the input. The target is the one that I had taken out. The batch is all of the possibilities of taking one out. This assumes every anime a user has finished was liked. This can be seen in the ratings given, with a mode of 8/10. Only the 1,000 most popular animes have been considered as the training time would increase with diminishing returns, in the future with better hardware I may allow more input features.

• showing more lower popularity shows (adjusting it)
• distributed training + UI to make it accesible. I do not have a windows system so I will only be able to make it for linux (ubuntu will be best due to tensorflow support)
• higher base encoding

email: [email protected]

Please do not use this code without crediting me.