A Composite Divergence Approach to Robust Multivariate Estimation under Cellwise and Casewise Contamination

What the math gives to ML

The paper introduces a composite density-power divergence that robustifies low-dimensional likelihood components without requiring a tractable full joint density. Its transferable asset is an explicit bounded-influence weighting mechanism: components assigned low predictive probability contribute less to the gradient, protecting training from whole-example and individual-feature corruption. A direct neural-network adaptation is to replace the usual negative log-likelihood with a weighted composite DPD loss over output heads, tokens, labels, nodes, or edges. The normalization term involving the model density raised to power 1+alpha prevents the method from degenerating into arbitrary loss clipping.