AutoLossGen

This repository includes the implementation for paper "AutoLossGen: Automatic Loss Function Generation for Recommender Systems".

Zelong Li, Jianchao Ji, Yingqiang Ge, Yongfeng Zhang. 2022. AutoLossGen: Automatic Loss Function Generation for Recommender Systems. In the Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '22)

This package is mainly contributed by Zelong Li (zelong.li@rutgers.edu), Yongfeng Zhang (yongfeng.zhang@rutgers.edu). We welcome any issues and requests for model implementation and bug fix.

Citation

If you use the code, please cite our paper:

@article{li2022autolossgen,
  title={AutoLossGen: Automatic Loss Function Generation for Recommender Systems},
  author={Li, Zelong and Ji, Jianchao and Ge, Yingqiang and Zhang, Yongfeng},
  journal={SIGIR},
  year={2022}
}

Environments

Python 3.9.7

Necessary packages:

PyTorch==1.9.0
Numpy==1.20.3
scikit-learn==1.0.2
scipy==1.7.3
pandas==1.4.1
tqdm==4.62.3

Example to run the codes

AutoLossGen has three phases.

An example to run the first phase: Loss Search

python main.py --epoch 10000 --child_num_branches 9 --child_num_layers 10 --search_loss --gpu 0 --sample_branch_id --sample_skip_id --controller_num_aggregate 10 --controller_train_steps 10 --log_file ../log/log_0.txt --formula_path ../model/Formula_0.txt --train_with_optim --dataset ml100k01-1-5 --model_name BiasedMF --random_seed 42

Note: If the number of epochs is set too small, the performance may not good enough as the search is not throughout. We recommend a large number of epochs, and you can stop it manually.

The second phase: Validation Check is run by

python loss_valid_check.py

The file path and decode method are currently hard coded in this python file.

The last phase: Effectiveness Test. An example is:

python main.py --epoch 10000 --log_file ../log/log_1.txt --model_name BiasedMF --loss_func MaxR --metric AUC --dataset ml100k01-1-5 --smooth_coef 1e-6

Reference

• We leveraged the dataset of NCR projects to implement our experiment.
• We implemented our Controller part referring to this project, which is a PyTorch implementation of paper Efficient Neural Architecture Search via Parameters Sharing.