Awesome

Infinite Recommendation Networks: A Data-Centric Approach (Distill-CF)

This repository contains the implementation of Distill-CF (using ∞-AE) from the paper "Infinite Recommendation Networks: A Data-Centric Approach" [arXiv] where we leverage the NTK of an infinitely-wide autoencoder to perform data distillation i.e. instead of training models on large datasets, train them on terse, high-fidelity, and synthetic data summaries generated by Distill-CF. Notably, Distill-CF:

• Synthesizes fake users which are later used for model training
• Is easy to implement (<100 lines of relevant code)
• Converges in O(100) iterations
• Models trained only on 500 users generated by Distill-CF are enough to outperform SoTA models trained on the full dataset

The paper also proposes ∞-AE: a SoTA implicit feedback recommendation model that has a closed-form solution, and only a single hyper-parameter. We provide ∞-AE's code in a separate GitHub repository.

If you find any module of this repository helpful for your own research, please consider citing the below paper. Thanks!

@article{inf_ae_distill_cf,
  title={Infinite Recommendation Networks: A Data-Centric Approach},
  author={Sachdeva, Noveen and Dhaliwal, Mehak Preet and Wu, Carole-Jean and McAuley, Julian},
  booktitle={Advances in Neural Information Processing Systems},
  series={NeurIPS '22},
  year={2022}
}

Code Author: Noveen Sachdeva (nosachde@ucsd.edu)

Setup

Environment Setup

pip install -r requirements.txt

Data Setup

This repository already includes the pre-processed data for ML-1M, Amazon Magazine, and Douban datasets as described in the paper. The code for pre-processing is in preprocess.py.

How to distill data?

• Edit the hyper_params.py file which lists all config parameters of Distill-CF.
• Finally, type the following command to train and evaluate ∞-AE on data synthesized by Distill-CF:

CUDA_VISIBLE_DEVICES=0 python distill.py

• Alternatively, to run a grid search on various hyper-parameters of Distill-CF, please edit the configuration (including GPU-ID) in grid_search_distill.py and then run:

python grid_search_distill.py

Results sneak-peak

<br><br> Below are the nDCG@10 results for the datasets used in the paper:

Note: The user synthesis budget for Distill-CF is only 500 for this table. ∞-AE's results on the Netflix dataset (marked with a *) are obtained by training only on 5% of the total users. Note however, all other methods are trained on the full dataset.

MIT License