Fast Generic Interaction Detection for Model Interpretability and Compression

Tianjian Zhang, Feng Yin, Zhi-Quan LuoICLR 2022 · OpenReview

Deep learning models are notoriously "black boxes"—they make highly accurate predictions but offer little insight into why they make them. A critical step toward model interpretability is understanding feature interactions: how variables (like "age" and "income") work together to influence an outcome, rather than just looking at them in isolation.

Existing methods for finding these interactions face a difficult "trilemma":

  1. Efficiency: Calculating all possible interactions in high-dimensional data is computationally expensive.
  2. Generality: Many current tools only work for specific architectures (like ReLU networks) and fail on others (like tree-based models).
  3. Stability: Some methods rely on local derivatives that can be misleading if the model's "landscape" is flat or jagged.

To solve this, researchers have developed a fast, model-agnostic method that not only explains a black box but also uses those explanations to build a better, smaller model called ParaACE.

The Global Advantage over Local Derivatives

In models like ReLU networks, the local derivative (Hessian) is zero almost everywhere because the function is piecewise linear.

Piecewise linear surface of a ReLU network

By looking at the Expected Hessian over a distribution instead of a single point, this new approach captures the true global relationship that local methods miss, ensuring stable and reliable interaction detection.

A Two-Stage Framework

The paper introduces a novel two-stage framework to detect interactions and compress the model:

1. Interaction Detection via Bandits

The authors define interaction strength using the Expected Hessian—a global statistical metric. To avoid the "brute force" cost of calculating this for every feature pair across the entire dataset, they ingeniously frame it as a Multi-Armed Bandit problem.

Mapping feature pairs to bandit arms for UCB sampling

Each feature pair is treated as an "arm." Using the Upper Confidence Bound (UCB) algorithm, the system adaptively samples the most promising feature pairs, significantly cutting down on redundant calculations and speeding up detection by an order of magnitude.

2. ParaACE (Parametric Alternating Conditional Expectation)

Once the top $k$ interactions are found, the researchers build a "Student" model. Unlike a standard dense network, ParaACE is structured to explicitly model "main effects" (individual features) and "interacting effects" (the detected pairs) using small, modular neural network blocks. It also includes an optional "fix-up" layer to handle any residual complexity.

Key Results and Insights

The authors validated their approach on a diverse suite of benchmarks, including synthetic functions, real-world regression tasks (like F16 aircraft control and biomedical voice measurements), and classification datasets.

  • The Power of Inductive Bias: The most surprising finding is that the tiny Student model (ParaACE) often outperforms the massive Teacher model. By explicitly "telling" the model which features interact, we provide a superior inductive bias that prevents the model from wasting parameters on noise.
  • Performance Boost: On average, ParaACE improved prediction performance (NRMSE) by 26% compared to a standard over-parameterized fully-connected network.
  • Massive Compression: The resulting models were frequently 100x to 300x smaller than the baseline, making them ideal for deployment on edge devices.
  • Scientific Discovery: In a drug combination task involving 50 FDA-approved drugs, the model successfully identified 15 pairwise interactions already verified in the DrugBank database, proving its utility for genuine scientific discovery.

Limitations and Future Work

While highly effective for pairwise interactions, detecting higher-order (3-way or 4-way) interactions remains challenging. Additionally, the quality of the detected interactions inherently depends on the quality of the initial "Teacher" model; a poorly trained black box will yield incorrect interaction insights.

Future work aims to extend this framework to even higher-dimensional "big data" scenarios, where the number of features reaches the thousands, requiring even more aggressive screening techniques.

Summary updated

Abstract

The ability of discovering feature interactions in a black-box model is vital to explainable deep learning. We propose a principled, global interaction detection method by casting our target as a multi-arm bandits problem and solving it swiftly with the UCB algorithm. This adaptive method is free of ad-hoc assumptions and among the cutting-edge methods with outstanding detection accuracy and stability. Based on the detection outcome, a lightweight and interpretable deep learning model (called ParaACE) is further built using the alternating conditional expectation (ACE) method. Our proposed ParaACE improves the prediction performance by 26 % and reduces the model size by 100+ times as compared to its Teacher model over various datasets. Furthermore, we show the great potential of our method for scientific discovery through interpreting various real datasets in the economics and smart medicine sectors. The code is available at https://github.com/zhangtj1996/ParaACE.
BibTeX
@inproceedings{zhang2022fast,
  author = {Tianjian Zhang and Feng Yin and Zhi-Quan Luo},
  title = {{Fast Generic Interaction Detection for Model Interpretability and Compression}},
  booktitle = {International Conference on Learning Representations},
  year = {2022},
  url = {https://openreview.net/forum?id=9MqubVN_bg},
}
Figure