Feature Interaction Interpretability via Interaction Detection

This is the official code repository for the paper, "Feature Interaction Interpretability: A Case for Explaining Ad-Recommendation Systems via Neural Interaction Detection".

Example Explanations

Global Interpretations

Local Interpretations (of ResNet classifications)

Michael Tsang, Dehua Cheng, Hanpeng Liu, Xue Feng, Eric Zhou, Yan Liu, Feature Interaction Interpretability: A Case for Explaining Ad-Recommendation Systems via Neural Interaction Detection, ICLR 2020.

Neural Interaction Detection:
Michael Tsang, Dehua Cheng, Yan Liu, Detecting Statistical Interactions from Neural Network Weights, ICLR 2018.

Setup

In a linux environment with Python 3.6:

pip install -r requirements.txt

We require CUDA 10 support to use GLIDER.

Usage

1. MADEX

MADEX (Model-Agnostic Dependency EXplainer) is a method for interpreting feature interactions from a black-box prediction model per data instance. It contains two versions of Neural Interaction Detection (NID): the original NID and GradientNID. NID is a fast and accurate method to detect arbitrary-order interactions in polynomial time, whereas GradientNID exactly detects interactions from an explainer MLP. The following domains are showcased: DNA, graph, image, and text modeling.

Show instructions

cd 1.\ madex/

The following notebooks are available to demo MADEX:

madex_example_dna.ipynb
madex_example_graph.ipynb
madex_example_image.ipynb
madex_example_text.ipynb

2. GLIDER

GLIDER (GLobal Interaction Detection and Encoding for Recommendation) takes MADEX beyond model interpretation on recommendation tasks (or tablular data modeling). GLIDER detects feature interactions that reoccur across data instances from a source recommender model, then explicitly encodes the interactions in a target recommender model. This process is a form of automatic feature engineering.

Show instructions

cd 2.\ glider/

A. Data Preparation

Please follow instructions in the AutoInt repo for how to prepare data splits.

The same code is also provided in this repo and follows the same series of commands. The Criteo dataset is found here. Place it in the path data/autoint/criteo.

mkdir data/autoint/criteo
python data/initial_data_prep/criteo/preprocess.py
python data/initial_data_prep/Kfold_split/stratifiedKfold.py
python data/initial_data_prep/criteo/scale.py

B. Global Interaction Detection

First, train a baseline AutoInt model.

python models/autoint/train.py --exp baseline --data data/autoint/criteo --save_path experiments/autoint/criteo/baseline/ --run_times 1 --gpu 0

Then, run global interaction detection on this model.

python detect_global_interactions.py --save_path experiments/autoint/criteo/baseline/1/ --data criteo

using defaul arguments instead

par_batch_size is the number of data instances to process in parallel. Set this based on the number of CPU processes and GPU memory available.
SAVEID shows up again later. Use a descriptive identifier.

C. Cross Feature Generation

This is where I cannot progress past

To generate cross features:

python make_cross_feature_data.py --data_file experiments/detected_interactions_criteo_SAVEID.pickle --exp cross_K40 --K 40 --data criteo --autoint_save_path data/autoint/criteo --deepctr_save_path data/deepctr/criteo --save_base_data true

D. Train DeepCTR models:

Wide&Deep: WDL
DeepFM: DeepFM
Deep&Cross: DCN
xDeepFM: xDeepFM

Baseline:

python train_deepctr.py --model WDL --ds criteo --exp baseline --patience 5 --test_id baseline_experiment --gpu 0

Baseline + GLIDER (distillation):

python train_deepctr.py --model WDL --ds criteo --exp cross --patience 5 --test_id cross_experiment --gpu 0 --d_cross_exp cross_K40 --n_cross 40

E. Train AutoInt models:

Baseline + GLIDER (enhancement):

python models/autoint/train.py --exp cross --data data/autoint/criteo --save_path experiments/autoint/criteo/cross/ --gpu 0 --cross_exp cross_K40

References

@inproceedings{tsang2020feature,
  title={Feature Interaction Interpretability: A Case for Explaining Ad-Recommendation Systems via Neural Interaction Detection},
  author={Michael Tsang and Dehua Cheng and Hanpeng Liu and Xue Feng and Eric Zhou and Yan Liu},
  booktitle={International Conference on Learning Representations},
  year={2020},
  url={https://openreview.net/forum?id=BkgnhTEtDS}
}

Neural Interaction Detection:

@article{tsang2017detecting,
  title={Detecting Statistical Interactions from Neural Network Weights},
  author={Michael Tsang and Dehua Cheng and Yan Liu},
  journal={arXiv preprint arXiv:1705.04977},
  year={2017}
}

haz/interaction_interpretability