A Model of Causal Explanation on Neural Networks for Tabular Data

A MODEL OF CAUSAL EXPLANATION ON NEURAL NETWORKS FOR TABULAR DATA Takashi Isozaki Sony Computer Science Laboratories, Inc. Tokyo, Japan Takashi.Isozaki@sony.com Masahiro Yamamoto Sony Computer Science Laboratories, Inc. Tokyo, Japan Masahiro.A.Yamamoto@sony.com Atsushi Noda Sony Corporation of America CA, USA Atsushi.A.Noda@sony.com ABSTRACT The problem of explaining the results produced by machine learning methods continues to attract attention. Neural network (NN) models, along with gradient boosting machines, are expected to be utilized even in tabular data with high prediction accuracy. This study addresses the related issues of pseudo-correlation, causality, and combinatorial reasons for tabular data in NN predictors. We propose a causal explanation method, CENNET, and a new explanation power index using entropy for the method. 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. We show that CENNET provides such explanations through comparative experiments with existing methods on both synthetic and quasi-real data in classification tasks. Keywords Neural networks · Explainable AI · Structural causal models · Causality · Prediction 1 Introduction Research on interpretable machine learning (ML) (e.g., [1, 2]) still attracts much attention. Complex methods, including gradient boosting machines (GBMs) [3] and neural networks (NNs), have performed better on many tasks than classical methods such as regression analysis and decision trees. However, methods generally face a tradeoff between interpretability and predictive performance. Many methods have been proposed to provide explanatory properties to predictive models post-hoc, including research on neural net-specific methods such as Grad-cam [4], and local, model-independent methods such as LIME [5], Anchors [6], and SHAP [7]. These are being used in industry, in part because they can be easily integrated into the pipeline of the forecasting process using existing ML methods. Nevertheless, several challenges remain in this area, some of which are discussed below. As a challenge for ML methods that provide explanatory properties, we note that pseudo-correlation and causality are not well addressed. Information about causality is crucial to human understanding of things and situations. The journey of understanding various phenomena in natural sciences such as epidemiology and in social sciences such as psychology is evidence that humans try to understand things through causal relationships (e.g., [8], [9]). Even in our immediate surroundings, cases in which we wish to understand things through causal relationships, such as the causes of crimes and abnormal weather, are an everyday occurrence. If an ML explanatory system that does not consider causality uses a contributing variable to explain its output, the user may interpret the variable as being causally related without being aware of whether it is a pseudo-correlated variable. As a result, incorrect causal knowledge may accumulate in the general public and in local organizations such as marketing and manufacturing companies. In other words, ML and AI arXiv:2512.21746v1 [cs.LG] 25 Dec 2025 Isozaki et al. risk instilling a false understanding in people. There is also a concern that if the causal relationship is Markovianly not direct, multiple similar explanations will be presented, which will be redundant and difficult to understand. Therefore, it will be effective to present reasons using variables that directly affect the prediction results. This will also increase the likelihood of smoothly leading to intervention actions to change the predicted outcome. In other words, incorporating such structural causality into explainability will not only improve understanding of the reasons for the prediction but also increase the feasibility of a system that seamlessly links prediction and intervention. However, the methods for explaining AI by inferring structural causality among input variables behind the data are still underdeveloped to the best of our knowledge. Another challenge is the following: while many explanatory methods, such as LIME and SHAP, assume additivity of multiple variables, there is concern that the predictive model itself is learning more complex patterns among explanatory variables but may be losing some amount of information [10] due to the assumption [11] when providing explanatory properties. One could say that ML should account for complex reasons that frequently occur in reality, but this does not seem to have been well addressed. For instance, the causes of some diseases that are difficult to detect are often thought to arise only when complex factors are combined, and the contribution of a single variable itself is often not high. For example, gender is rarely the sole reason for the occurrence of ce

This content is AI-processed based on open access ArXiv data.