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.

• ⚙️ Hybrid Model with Genre Embeddings — Added genre-based embeddings to enrich BERT inputs with auxiliary item metadata.
• ⚙️ No Positional Encoding — Removed positional encoding due to lack of temporal signals in the dataset.
• 🎌 Anime-Specific User-Item Dataset — Built on a large-scale dataset tailored for anime recommendations.
• Kaggle Train and Inference Notebooks available
• Huggingface Space available

https://www.kaggle.com/code/ramazanturann/animerecommenderbert-inference

https://www.kaggle.com/code/ramazanturann/animerecommenderbert-train

https://huggingface.co/spaces/mramazan/AnimeRecBERT

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

The model was trained on a large-scale dataset containing 1.77 million users and 148 million ratings. Although positional encoding was removed, the results remain very close to those of the original BERT4Rec repository. Below are the Top-K recommendation metrics:

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

python3 -m venv venv
source venv/bin/activate 

python -m venv venv
venv\Scripts\activate 

mkdir -p Data/AnimeRatings

curl -L -o Data/AnimeRatings/animeratings.zip \
     https://www.kaggle.com/api/v1/datasets/download/ramazanturann/user-animelist-dataset

unzip Data/AnimeRatings/animeratings.zip -d Data/AnimeRatings/

mkdir Data\AnimeRatings

kaggle datasets download -d ramazanturann/user-animelist-dataset -p Data\AnimeRatings

Expand-Archive -Path "Data\AnimeRatings\user-animelist-dataset.zip" -DestinationPath "Data\AnimeRatings" -Force

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

pip install -r requirements.txt

python main_local.py \
  --checkpoint-path Data/AnimeRatings/pretrained_bert.pth \
  --dataset-path Data/preprocessed/AnimeRatings_min_rating7-min_uc10-min_sc10-splitleave_one_out/dataset.pkl \
  --animes-path Data/animes.json \
  --images-path Data/id_to_url.json \
  --mal-urls-path Data/anime_to_malurl.json \
  --type-seq-path Data/anime_to_typenseq.json \
  --genres-path Data/id_to_genres.json
 

You can set parameters in templates.py file

# Preprocess takes up to 30 minutes
# To avoid this, this script will move the dataset.pkl and negative samples file to preprocessed folder
# You can run this if you didn't change the data 
python move_files.py         

# This script will train, validate and test the model.
python main.py  --template train_bert             

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

Out of the Top 20 recommendations, 11 titles were already in my completed/favorites list — showing strong personalization performance. Most of the recommendations were sequels to my favorites, which shows BERT's capacity for data integrity and contextual awareness.