AnimeRecBERT: BERT-Based Anime Recommendation System

AnimeRecBERT is a personalized anime recommendation system based on BERT transformer architecture. Adapted from https://github.com/jaywonchung/BERT4Rec-VAE-Pytorch, this project introduces customizations tailored for an anime recommendation system and inference. Main dataset has 1.77M users and 148M ratings, in this project we used a subset of 560K users and 54M ratings for efficiency.

Full dataset used in this repo: https://github.com/MRamazan/AnimeRecBERT-Hybrid

This project provides a solid foundation for further development in personalized anime recommendation using transformer-based models.

• 🕒 No positional encoding, since the dataset contains no temporal information, i removed positional encoding and it performed better.
• 🎌 Anime-specific user-item dataset
• 🖥️ GUI interface for real-time recommendations
• 🌐 Web Demo

Kaggle Inference & Interactive Recommender

In kaggle notebooks, full dataset used for training and inference instead of subset like in web demo.

Interactive Recommender notebook: https://www.kaggle.com/code/tavuksuzdurum/animerecommenderbert-inference

Train notebook: https://www.kaggle.com/code/tavuksuzdurum/animerecommenderbert-train

Web Demo Available 🌐

www.animerecbert.online
(may be down due to 512MB free ram limit) https://github.com/MRamazan/AnimeRecBertWeb

Metrics

Model trained on a large-scale dataset with 560,000 users and 54 million ratings. Below are the Top-K recommendation metrics:

Metric Value Recall@100 0.9998 NDCG@100 0.733 Recall@50 0.994 NDCG@50 0.732 Recall@20 0.967 NDCG@20 0.727 Recall@10 0.919 NDCG@10 0.715 Recall@5 0.841 NDCG@5 0.689 Recall@1 0.507 NDCG@1 0.507

Setup & Usage

Clone Repo

git clone https://github.com/MRamazan/AnimeRecBERT
cd AnimeRecBERT

Create & Activate venv

For Linux

python3 -m venv venv
source venv/bin/activate 

For Windows

python -m venv venv
venv\Scripts\activate 

Download Dataset & Pretrained Model

For Linux

curl -L -o Data/AnimeRatings54M/animeratings-mini-54m.zip \
     https://www.kaggle.com/api/v1/datasets/download/tavuksuzdurum/animeratings-mini-54m

unzip Data/AnimeRatings54M/animeratings-mini-54m.zip -d Data/AnimeRatings54M/

For Windows

curl -L -o Data\AnimeRatings54M\animeratings-mini-54m.zip `
     https://www.kaggle.com/api/v1/datasets/download/tavuksuzdurum/animeratings-mini-54m

Expand-Archive -Path 'Data\AnimeRatings54M\animeratings-mini-54m.zip' -DestinationPath 'Data\AnimeRatings54M'

Install Requirements

Install PyTorch from https://pytorch.org/get-started/locally/

pip install -r requirements.txt

All Done, Start Inference

Run PYQT App

python inference_gui.py \
  -c Data/AnimeRatings54M/pretrained_bert.pth \
  -d Data/preprocessed/AnimeRatings54M_min_rating7-min_uc10-min_sc10-splitleave_one_out/dataset.pkl \
  -a Data/AnimeRatings54M/animes.json \
  --inference True \
  --template train_bert        

Train Code (Not Required for inference)

you can set parameters in templates.py file

# This script will train, validate and test the model.
# Training not required for inference.
python main.py  --template train_bert             

GUI

Results

🌟 My Favorites (Input for Inference)

#	Anime Title
1	Youkoso Jitsuryoku Shijou Shugi no Kyoushitsu e
2	Giji Harem
3	Ijiranaide, Nagatoro-san
4	86 (Eighty-Six)
5	Mushoku Tensei: Isekai Ittara Honki Dasu
6	Made in Abyss
7	Shangri-La Frontier: Kusoge Hunter, Kamige ni Idoman to su
8	Vanitas no Karte
9	Jigokuraku

🌟 BERT Recommendations Based on My Favorites

Note: The position of favorites does not affect inference results, as the model uses only the presence of items (not sequence).

🏆 Top Anime Recommendations for Me

Rank	Anime Title	Score
#1	Yofukashi no Uta	14.1062
#2	Summertime Render	12.8611
#3	Mushoku Tensei II: Isekai Ittara Honki Dasu	12.8518
#4	Tengoku Daimakyou	12.7181
#5	Jujutsu Kaisen	12.3117
#6	Horimiya	12.0604
#7	Chainsaw Man	11.7838
#8	86 Part 2	11.7756
#9	Mushoku Tensei: Isekai Ittara Honki Dasu Part 2	11.6520
#10	Ore dake Level Up na Ken (Solo Leveling)	11.6396
#11	Kage no Jitsuryokusha ni Naritakute! 2nd Season	11.6270
#12	Youkoso Jitsuryoku Shijou Shugi no Kyoushitsu e 2nd Season	11.5398
#13	Sousou no Frieren	11.5376
#14	Tonikaku Kawaii (Tonikawa: Over the Moon for You)	11.5205
#15	Cyberpunk: Edgerunners	11.5020
#16	Tenki no Ko (Weathering With You)	11.5014
#17	Dandadan	11.4893
#18	Spy x Family	11.4581
#19	Make Heroine ga Oosugiru!	11.4468
#20	Boku no Kokoro no Yabai Yatsu	11.4122

✅ Evaluation: How Good Are the Recommendations?

Out of the Top 20 recommendations, 10 titles were already in my completed/favorites list — showing strong personalization performance.

Watched & Liked? ✅	Title
✅	Mushoku Tensei II: Isekai Ittara Honki Dasu
✅	Mushoku Tensei: Isekai Ittara Honki Dasu Part 2
✅	Youkoso Jitsuryoku Shijou Shugi no Kyoushitsu e 2nd Season
✅	Make Heroine ga Oosugiru!
✅	Spy x Family
✅	Dandadan
✅	86 Part 2
✅	Jujutsu Kaisen
✅	Chainsaw Man
✅	Cyberpunk: Edgerunners

I’m genuinely excited to watch the remaining anime as well — even with a quick glance, it’s clear they’re a great fit for my taste.