MinShap: A Modified Shapley Value Approach for Feature Selection

The critical distinction MinShap draws is between a feature being correlated with the outcome versus being directly causal of it. Standard Shapley values, by design, distribute credit across correlated features, which is useful for attribution but actively misleading when the goal is to prune a model's input space.

This sits within a cluster of interpretability work appearing on Modelwire this week. The ORCA paper on structural interpretability in SVMs addresses a similar frustration: post-hoc explanation methods often describe model behavior without isolating which inputs are doing real predictive work. Both papers are essentially arguing that the dominant tooling for explainability was built for auditing, not for feature engineering, and those are different problems. The connection to other stories from this period is limited, but the pairing with ORCA is genuine and worth noting for readers tracking the interpretability beat.

The real test is whether MinShap's feature rankings diverge meaningfully from standard Shapley rankings on established tabular benchmarks with known ground-truth feature relevance. If published follow-up experiments show consistent divergence on datasets like those from the UCI repository, the method has practical traction; if the differences are marginal, the theoretical fix may not matter in practice.