ProxySR

This is a PyTorch Implementation of ProxySR, proposed in "Unsupervised Proxy Selection for Session-based Recommender Systems", SIGIR'21.

Requirements

• Python2
• PyTorch 1.0.1 or higher

Configuation

Settings for training and evaluating ProxySR. Please refer to our paper for detailed description of each configuration.

• --dataset: Dataset. ex) diginetica
• --batch_size: Mini-batch size for training.
• --val_batch_size: Mini-batch size for evaluation.
• --embed_dim: Embedding size.
• --lr: Learning rate.
• --k: Number of proxies.
• --dropout_rate: Dropout rate.
• --margin: Margin for the marginal loss.
• --lambda_dist: Regularization coefficient for distance regularizer.
• --lambda_orthog: Regularization coefficient for orthogonality regularizer.
• --E: Number of annealing epoch.
• --patience: Number of epoches to wait for learning to end after no improvement.
• --max_position: Maximum length of input sequence.
• --t0: Initial temperature.
• --te: Final temperature.
• --num_epoch: Maximum number of training epoches.
• --repetitive: (True) Next item recommendation with repetitive consumption or (False) Next unseen item recommendation.

How to train and test a model on Diginetica dataset (next unseen item recommendation)

python main.py --dataset=diginetica

How to train and test a model on Diginetica dataset (item recommendation with repetitive consumption)

python main.py --dataset=diginetica --repetitive=True

Reference

• Unsupervised Proxy Selection for Session-based Recommender Systems (SIGIR'21)

  • Junsu Cho, SeongKu Kang, Dongmin Hyun, Hwanjo Yu