OptInter

This repository contains PyTorch Implementation of ICDE 2022 paper:

Memorize, factorize, or be naive: Learning optimal feature interaction methods for CTR Prediction.

You can find our paper here.

Usage

Following the steps below to run our codes:

Install

pip install -r requirements.txt

Download Dataset

Please Download the Criteo and Avazu datasets.

• For Criteo dataset, copy the train.txt file under datasets/Criteo and rename it to full.txt.
• For Avazu dataset, copy the train.csv file under datasets/Avazu and rename it to full.csv.

Preprocess Dataset

• For Criteo dataset, run python preprocess/criteo.py
• For Avazu dataset, run python preprocess/avazu.py

Search

• For Criteo dataset, run python learn/CriteoSearch.py
• For Avazu dataset, run python learn/AvazuSearch.py

Re-train

• For Criteo dataset, run python learn/CriteoTrain.py --load XXX --model DNN_cart --alpha_mode 0
• For Avazu dataset, run python learn/AvazuTrain.py --load XXX --model DNN_cart --alpha_mode 0

Here XXX indicates the logs dictionary generated during the search stage.

Baseline Running

• For Criteo dataset, run python learn/CriteoTrain.py --model YYY
• For Avazu dataset, run python learn/AvazuTrain.py --model YYY

Here YYY could be {LR, FM, FNN, IPNN, DeepFM, PIN, Poly2, DNN_cart}

Citation

@inproceedings{OptInter,
  author       = {Fuyuan Lyu and
                  Xing Tang and
                  Huifeng Guo and
                  Ruiming Tang and
                  Xiuqiang He and
                  Rui Zhang and
                  Xue Liu},
  title        = {Memorize, Factorize, or be Naive: Learning Optimal Feature Interaction
                  Methods for {CTR} Prediction},
  booktitle    = {38th {IEEE} International Conference on Data Engineering, {ICDE} 2022},
  pages        = {1450--1462},
  address      = {Kuala Lumpur, Malaysia},
  publisher    = {{IEEE}},
  year         = {2022},
  url          = {https://doi.org/10.1109/ICDE53745.2022.00113},
  doi          = {10.1109/ICDE53745.2022.00113},
  timestamp    = {Sun, 30 Jul 2023 12:39:38 +0200},
  biburl       = {https://dblp.org/rec/conf/icde/LyuTGTHZL22.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}