README

This is the official implementation of the paper DART: Resolving Training Biases for Recommender Systems via Data Discarding and Relabeling .

Hierarchy

.
├── torchfm # TorchFM library with some modifications
│   ├── dataset
│   │   ├── __init__.py
│   │   ├── avazu.py    # Modified code for Avazu -- Setting 1
│   │   ├── criteo.py   # Modified code for Criteo -- Setting 1
│   │   ├── custom.py   # Code for custom dataset (all datasets except MovieLens, Avazu and Criteo) -- Setting 2
│   │   ├── movielens_1.py  # Modified code for MovieLens -- Setting 1
│   │   ├── movielens_2.py  # Modified code for MovieLens -- Setting 2
│   ├── model
│   │   ├── __init__.py
│   │   ├── fm.py   # FM model
│   │   ├── dfm.py  # DFM model
│   │   ├── ...
├── test
│   ├── test_layers.py  # Test for layers
├── examples
│   ├── arxiv   # Arxiv deprecated code
│   ├── chkpt   # Checkpoints
│   ├── data    # Datasets
│   │   ├── adressa
│   │   ├── Jester
│   │   ├── book_crossing
│   │   ├── ml-100k
│   │   ├── ml-1m
│   │   ├── ...
│   ├── logs    # Logs, Processing Codes and Results
│   │   ├── [2023-05-06]
│   │   ├── ...
│   ├── config.py   # Configuration file to set random seed
│   ├── main_1.py   # Main function for Setting 1
│   ├── main_2.py   # Main function for Setting 2
├── README.md # This file

Requirements

Here are the requirements to run the code:

python==3.9.7
torch==1.10.1+cu102
tqdm==4.62.3
numpy==1.21.5

Dataset

We use the following datasets in our experiments, which are all publicly available:

The datasets should be stored in the examples/data folder separately. For several of them, you need to run the preprocessing code in each folder to process the raw data into the required format.

Setting

Setting 1

For the first setting, we use the MovieLens, Avazu and Criteo datasets. Here we simulate false-positive interactions by randomly sampling noise.

Setting 2

For the second setting, we use the Jester, Book-Crossing, Adressa, Yelp, Amazon and Gowalla datasets. Here we use a more reasonable method to model false-positive interactions as described in the paper.

Method

We implement the following methods in our experiments:

Base (none)
DART (mixed)
ADT-CE-T (discard)
ADT-CE-R (reweight)
LCD (lcd)
LCD-Re (lcd-re)
Bootstrap-Soft (bootstrap_soft)
Bootstrap-Hard (bootstrap_hard)

Usage

Run

To run the code, you can use the following command in /examples folder:

usage: main_2.py [-h] [--dataset_name DATASET_NAME] [--model_name MODEL_NAME] [--epoch EPOCH]
                 [--learning_rate LEARNING_RATE] [--batch_size BATCH_SIZE]
                 [--weight_decay WEIGHT_DECAY] [--device DEVICE] [--save_dir SAVE_DIR]
                 [--method METHOD]

optional arguments:
  -h, --help            show this help message and exit
  --dataset_name DATASET_NAME
                        criteo, avazu, movielens100K, movielens1M, or adressa, amazon-book,
                        yelp, book-crossing, jester
  --model_name MODEL_NAME
  --epoch EPOCH
  --learning_rate LEARNING_RATE
  --batch_size BATCH_SIZE
  --weight_decay WEIGHT_DECAY
  --device DEVICE
  --save_dir SAVE_DIR
  --method METHOD       none, mixed, discard, reweight, lcd, lcd-re, bootstrap_soft,
                        or bootstrap_hard

Hyper-parameters

Our four hyper-parameters $T_i, T_m, \tau_m, r$ are set in the train() method of main_2.py

Ti, Tm, taum, r = 0, 20000, 0.04, 0

It is worth noting that each experimental result may be adapted to different hyper-parameters to obtain the best results, depending on the specific dataset. Please refer to the Sensitivity Analysis section in the paper for details.

Citation

If you find this code useful, please cite our paper:

@inproceedings{,
  title={},
  author={},
  booktitle={},
  year={},
  organization={}
}

Contact

If you have any questions, please contact Ningzhi Tang via email: tangnz2019@mail.sustech.edu.cn.

Acknowledgement

We thank the authors of TorchFM for their open-source code.

TTangNingzhi/CS490-DART