Ananya Uppal
Due to the pervasive deployment of Machine Learning (ML) and Artificial Intelligence (AI) in high-stakes decision-making, it has become essential to understand how data quality issues originate and propagate through end-to-end pipelines that have traditionally been treated as black boxes. This work evaluates how corrupted training points cascade through an ML pipeline and degrade predictive accuracy, and whether data valuation methods can identify those points reliably enough to guide targeted repair. Three corruption mechanisms are injected under controlled conditions: label flips applied uniformly at random, label flips targeted at a protected demographic subgroup, and feature missingness targeted at that same subgroup. Two detection methods are then compared against identical, known corruption on the same pipelines: Shapley-based valuation computed over the full pipeline (DataScope) and entropy-based predictive uncertainty, each measured against a random-ordering baseline that cleans exactly the same rows. Across three datasets spanning two orders of magnitude in scale, two pipeline architectures, and two corruption rates, no single method dominates. The most effective method depends on whether the corruption lies in label space or feature space, and on whether it is spread uniformly or concentrated within a protected subgroup. DataScope's advantage proves to be a property of its ranking rather than of its final outcome: it is concentrated at partial cleaning budgets and disappears at full cleaning, where DataScope converges with random ordering. Entropy-based ranking is competitive on label noise but fails consistently on feature-space corruption. The study further establishes that the choice of correction action, restoring a label as against removing a row, is as consequential to recovered accuracy as the choice of detection method.