SASRec-Pytorch

PyTorch implementation of SASRec (Self-Attentive Sequential Recommendation) based on the paper for studying sequential recommender system:

Self-Attentive Sequential Recommendation (SASRec) Wang-Cheng Kang, Julian McAuley (ICDM 2018) https://arxiv.org/abs/1808.09781

---

📂 Project Structure

SASRec-Pytorch/
│
├── model/
│   └── sasrec.py        # SASRec model, evaluation, trainer
│
├── src/
│   ├── data.py          # data loader, sampler, batch generator
│   ├── path.py          # project path manager
│   ├── utils.py         # yaml loader, seed, misc utilities
│   └── main.py          # main training script
│
├── data/                # dataset directory (raw + processed)
│   └── ratings.csv      # Movielens input file
│
├── config.yaml          # all hyperparameters & experiment settings
└── README.md

---

All hyperparameters are centralized in config.yaml:

data:
  dataset: "movielens"
  max_len: 50
  num_workers: 4

train:
  num_epochs: 100
  batch_size: 128
  num_batches_per_epoch: 1000
  device: "cuda"
  best_model_path: "./checkpoints/best_model.pth"

early_stopping:
  patience: 5
  min_delta: 0.0

model:
  hidden_dim: 64
  num_blocks: 2
  num_heads: 1
  dropout: 0.2

optimizer:
  name: "adam"
  lr: 0.001
  weight_decay: 0.0

loss:
  name: "bce_with_logits"

---

Run Training

1. Place ratings.csv inside data/

data/
└── ratings.csv

2. Run main script:

uv run main.py

During training you’ll see:

[Epoch 10] Train Loss: 0.3421
  Val   - NDCG@10: 0.4821, Hit@10: 0.7013
  Test  - NDCG@10: 0.4751, Hit@10: 0.6942
  ** Best model updated and saved to './checkpoints/best_model.pth' **

---

Evaluation Metrics

The repository implements the common metrics in sequential recommendation:

• NDCG@10
• Hit Ratio@10

Evaluation uses the standard:

• 1 positive item
• 100 negative samples

as in the original SASRec paper.

---

Typical Results (Movielens-1M, example)

(Values vary depending on hyperparameters)

Metric	Score
NDCG@10	~0.48
Hit@10	~0.70

---

Acknowledgements

Inspired by the official SASRec implementations from:

• https://github.com/pmixer/SASRec.pytorch/tree/main