Interventional Multi-Instance Learning with Deconfounded Instance-Level Prediction

When applying multi-instance learning (MIL) to make predictions for bags of instances, the prediction accuracy an instance often depends on not only itself but also its context in corresponding bag. From viewpoint causal inference, such bag contextual prior works as a confounder and may result model robustness interpretability issues. Focusing this problem, we propose novel interventional (IMIL) framework achieve deconfounded instance-level prediction. Unlike traditional likelihood-based strategies, design Expectation-Maximization (EM) algorithm based intervention, providing robust selection training phase suppressing bias caused by prior. Experiments pathological image analysis demonstrate that our IMIL method substantially reduces false positives outperforms state-of-the-art MIL methods.