Improving Out-of-Vocabulary Handling in Recommendation Systems

This repository contains the code for the paper "Improving Out-of-Vocabulary Handling in Recommendation Systems".

Setup

First, install our local copy of RecBole. You will need to uninstall the original RecBole first (if already installed). (This framework is a major extension to RecBole https://github.com/RUCAIBox/RecBole. Recbole provides the basic framework to evaluate on recommendation system models with standard datasets. This framework enables inductive training and evaluation, as well as datasets.)

pip uninstall recbole
cd RecBole
pip install -e .

Then, you can run the code:

cd ../src
python run_recbole.py --dataset [DATASET]

The run_recbole.py script accepts the following general parameters:

• --dataset: Mandatory. The dataset name to train and evaluate on.
• --model: Mandatory. The model type. The currently supported types are BPR, DirectAU, DCNV2, WideDeep, and xDeepFM.
• --checkpoint_dir: Directory to store the model weights.
• --embedding_size: The embedding size to use.
• --train_batch_size: The training batch size.
• --eval_batch_size: The evaluation batch size.
• --weight_decay: Weight decay to use during training (if any).
• --gcs_bucket_name: The name of the GCS bucket you would like to write model weights to, if applicable. Note that you need write permissions on the bucket and should already be authenticated.
• --learning_rate: The model learning rate. Defaults to
• --log_wandb: Whether or not to log results to Weights and Biases.
• --model_eval_type: The model evaluation metrics type - either retrieval or ranking.

There are also inductive-specific parameters that can be passed in:

• --add_oov_buckets: Whether or not to add OOV buckets to the model. This should be true if using any trainable OOV methods.
• --inductive_embedder: The inductive embedder to use. The currently supported types are lsh, slsh, knn, dnn, dhe, fdhe, zero, and mean.
• --inductive_mapper: The inductive mapper to use. The only currently supported type is random.
• --inductive_eval: Whether or not to perform inductive evaluation (versus training only)
• --user_oov_buckets: The number of user OOV buckets to use.
• --item_oov_buckets: The number of item OOV buckets to use.
• --oov_feature_mask_rate: The rate at which to mask OOV features during training.
• --oov_freeze_embedding: Whether or not to freeze IV embeddings during OOV training.
• --oov_freeze_skip_optim: Whether or not to also freeze the optimizer parameters.
• --train_oov: Whether or not to train OOV embeddings at all. Should be true for all embedders except for zero and mean.
• --oov_only_epoch: Whether or not to split OOV samples out into its own epoch.
• --oov_train_ratio: Ratio of IV samples used for OOV training at every epoch.
• --oov_normalization_type: Feature normalization type. Can be one of three options: per-feature, global, none. Not implemented for all OOV embedders.

Note that any empty parameters will use the default parameters found in RecBole/recbole/properties/overall.yaml unless overriden by model-specific or dataset-specific configuration files. You can also pass in any model-specific hyperparameters here, which can be found in the RecBole/recbole/properties/model/MODEL_NAME.yaml files.

Citation

If you use this code in your research, please cite the following paper:

@article{shiao2024improving,
  title={Improving Out-of-Vocabulary Handling in Recommendation Systems},
  author={Shiao, William and Ju, Mingxuan and Guo, Zhichun and Chen, Xin and Papalexakis, Evangelos and Zhao, Tong and Shah, Neil and Liu, Yozen},
  journal={arXiv preprint arXiv:2403.18280},
  year={2024}
}