Causal Inference Guided Stable Learning: Improving Explainability and Prediction Stability in Machine Learning

Current AI algorithms face two major risks: lack of explainability and instability. These arise because most models are correlation‑driven and do not distinguish true causal links from spurious associations such as confounding and selection bias.

To address these issues, the speaker proposes using causal inference as a powerful modeling tool to recover genuine causal relationships and guide machine learning, thereby achieving interpretable and stable predictions.

Why instability occurs

Instability often stems from data problems (distribution shift, violation of IID assumptions) and model problems (reliance on correlation). Correlation can be caused by causation, confounding, or selection bias. Only causation yields stable, explainable links; the other two produce spurious, unstable correlations.

Proposed solution

The causal‑constrained stable learning framework consists of two parts: causal inference and stable learning. In the causal inference stage, each predictor’s causal effect on the outcome is evaluated, allowing removal of non‑causal variables (e.g., “grass” in a dog‑recognition example). In the stable learning stage, the recovered causal graph is used for Causation‑based Learning.

Data‑driven variable separation algorithm

Instead of treating all observed variables as confounders, the algorithm decomposes them into three groups: confounders X (affect both treatment T and outcome Y), adjustment variables Z (independent of T, related only to Y), and irrelevant variables I (independent of both). By correctly separating variables, propensity scores are computed only on X, while Z is used to regress Y, reducing variance.

Confounder discriminative balancing algorithm

This method learns variable‑specific weights (β) and sample weights (W) to differentiate the impact of each confounder, addressing the limitation of treating all confounders equally. It reduces both confounding bias and variance in ATT estimation.

Experimental validation

Experiments on synthetic data and a real‑world WeChat LONGCHAMP dataset (56 user features predicting ad click) show that the proposed data‑driven algorithm (DVD) achieves higher accuracy and lower variance than traditional propensity‑score methods. Causal regularizer combined with logistic regression or deep models yields stable RMSE across diverse test distributions.

Conclusion

Machine learning is largely correlation‑based, with sources of correlation being causation, confounding, and selection bias. Only causation provides stable, explainable relationships. By integrating causal inference into a stable learning framework, we can obtain models that are both interpretable and robust to distribution shifts, and future work may combine causal reasoning with deep representation learning for even greater stability.