Towards Learning and Explaining Indirect Causal Effects in Neural Networks

Abbavaram Gowtham Reddy, Saketh Bachu, Harsharaj Pathak, Benin Godfrey L, Varshaneya V, Vineeth N. Balasubramanian, Satyanarayan Kar

Abstract

Standard neural networks (NNs) typically assume input features are independent, which prevents them from capturing or explaining indirect causal effects—where one feature influences the output through another. This paper introduces Ante-Hoc Causal Explanations (AHCE), a framework that embeds known causal relationships directly into the network's architecture. By introducing lateral connections in the input layer and employing a two-phase training strategy, AHCE enables models to accurately learn and quantify both direct and indirect causal effects. Experimental results demonstrate that AHCE significantly outperforms existing post-hoc explanation methods in accuracy and computational efficiency.

Introduction

As neural networks are increasingly deployed in high-stakes domains, the demand for interpretable and causally grounded explanations has grown. While methods like Shapley values or Integrated Gradients provide feature importance, they often rely on the assumption of feature independence. Even recent causal explanation methods primarily focus on direct causal effects, ignoring the complex web of interactions where features influence each other.

In real-world scenarios, indirect pathways are common. For instance, socioeconomic status may influence job performance indirectly through its effect on education level. A standard feedforward NN, with its independent input neurons, cannot model these dependencies. Consequently, post-hoc explanations applied to these models fail to reflect true data dynamics, potentially masking biases or providing incomplete insights.

Methodology: Ante-Hoc Causal Explanations (AHCE)

The AHCE framework shifts the focus from post-hoc interpretation to ante-hoc design, ensuring the model is "interpretable by design."

Architectural Augmentation

The core of AHCE is the modification of the NN architecture to include a "Layer 0." This layer introduces feedforward connections among input neurons based on a known causal graph $\mathcal{G}$. This allows the network ($\mathcal{N}^{Ind}$) to explicitly represent how parent features influence child features before the information propagates through the rest of the network.

Comparison of the true causal graph $\mathcal{G}$, a standard neural network $\mathcal{N}$, and the proposed augmented architecture $\mathcal{N}^{Ind}$.

Two-Phase Training and Quantification

AHCE utilizes a two-phase training process:

  1. Partial Training: Freeze Layer 0 and train the subsequent layers.
  2. Topological Training: Train the entire network while forcing input features to be derived topologically (parents predicting children). A regularization loss ensures these derived features remain consistent with the actual training data.

To quantify effects, the authors provide algorithms for Average Direct Causal Effect (ADCE) and Average Indirect Causal Effect (AICE). To handle the computational complexity of marginalization, they employ a second-order Taylor approximation and a "binning strategy" for high-dimensional data, which precomputes interventional statistics to speed up inference.

Evaluation and Results

The authors evaluated AHCE against several benchmarks, including synthetic data with ground-truth causal equations, tabular datasets (Auto-MPG, Lung Cancer, Sachs), and industry-grade flight simulation time-series.

Comparison of explanation methods across various causal effect types.

Key findings include:

  • Accuracy: AHCE consistently outperformed baselines like Integrated Gradients (IG) and Causal Shapley (CSHAP) in capturing ground-truth effects.
  • Distribution Matching: AHCE achieved the lowest Frechet distance scores, indicating its learned causal effects closely mirror the true data-generating process.
  • Efficiency: The binning and Taylor approximation strategies provided a 3 to 10-fold speedup on complex RNN datasets with minimal loss in precision.

Discussion and Conclusion

AHCE demonstrates that architecture is a prerequisite for capturing indirect causal effects. By building structural inductive biases into the network, practitioners can better audit models for fairness and hidden biases that might be obscured in indirect pathways.

While AHCE requires a prior causal graph and assumes no unobserved confounders, it represents a significant step toward practical, scalable, and causally-faithful AI explanations. Future work may explore methods to relax the requirement for a fully known causal graph.

Summary updated

Abstract

Recently, there has been a growing interest in learning and explaining causal effects within Neural Network (NN) models. By virtue of NN architectures, previous approaches consider only direct and total causal effects assuming independence among input variables. We view an NN as a structural causal model (SCM) and extend our focus to include indirect causal effects by introducing feedforward connections among input neurons. We propose an ante-hoc method that captures and maintains direct, indirect, and total causal effects during NN model training. We also propose an algorithm for quantifying learned causal effects in an NN model and efficient approximation strategies for quantifying causal effects in high-dimensional data. Extensive experiments conducted on synthetic and real-world datasets demonstrate that the causal effects learned by our ante-hoc method better approximate the ground truth effects compared to existing methods.
BibTeX
@inproceedings{reddy2024towards,
  author = {Abbavaram Gowtham Reddy and Saketh Bachu and Harsharaj Pathak and Benin Godfrey L and Varshaneya V and Vineeth N. Balasubramanian and Satyanarayan Kar},
  title = {{Towards Learning and Explaining Indirect Causal Effects in Neural Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i13.29399},
}
Figure