Causal Explanations for Neural Networks on Tabular Data
Published:Dec 25, 2025 17:47
•1 min read
•ArXiv
Analysis
This paper addresses the crucial problem of explaining the decisions of neural networks, particularly for tabular data, where interpretability is often a challenge. It proposes a novel method, CENNET, that leverages structural causal models (SCMs) to provide causal explanations, aiming to go beyond simple correlations and address issues like pseudo-correlation. The use of SCMs in conjunction with NNs is a key contribution, as SCMs are not typically used for prediction due to accuracy limitations. The paper's focus on tabular data and the development of a new explanation power index are also significant.
Key Takeaways
- •Proposes CENNET, a novel method for providing causal explanations for neural networks on tabular data.
- •Combines neural networks with structural causal models (SCMs) to address issues like pseudo-correlation.
- •Introduces a new explanation power index using entropy.
- •Focuses on improving the interpretability of neural network predictions, particularly for tabular data.
Reference
“CENNET provides causal explanations for predictions by NNs and uses structural causal models (SCMs) effectively combined with the NNs although SCMs are usually not used as predictive models on their own in terms of predictive accuracy.”