Shadman Islam

Shadman Islam, Agustinus Kristiadi, Mostafa Milani

Synthetic instance-dependent label noise (IDN) benchmarks are widely used to evaluate noisy-label learning methods, yet existing approaches typically generate noise through imperfect annotators or classifier raters, leaving the source of ambiguity implicit. We introduce CILN, a benchmark generation framework that creates IDN through controlled input corruptions. A diverse voter pool labels corrupted instances, producing benchmark datasets in which both the source and severity of ambiguity are explicit and controllable. Using CIFAR10, MNIST, and Adult, we construct 90 benchmark settings spanning multiple corruption families and severity levels. Our experiments show that the resulting benchmarks exhibit genuine instance-dependent noise, provide diverse confusion structures, and, on CIFAR-10, can produce label distributions that are closer to human uncertainty than an existing synthetic IDN benchmark. We further demonstrate that corruption-mediated IDN can expose failure modes of popular noisy-label learning methods, including Co-Teaching and DivideMix, that are not observed under comparable levels of rater-fallibility noise. These findings suggest that noise structure, not only noise rate, plays an important role in benchmark difficulty and algorithm behavior. By making ambiguity generation explicit and controllable, CILN provides a complementary benchmarking framework for studying noisy-label learning under diverse sources of instance difficulty.

Hazim AbdElazim, Shadman Islam, Mostafa Milani

Data cleaning is often framed as a technical preprocessing step, yet in practice it relies heavily on human judgment. We report results from a controlled survey study in which participants performed error detection, data repair and imputation, and entity matching tasks on census-inspired scenarios with known semantic validity. We find systematic evidence for several cognitive bias mechanisms in data cleaning. Framing effects arise when surface-level formatting differences (e.g., capitalization or numeric presentation) increase false-positive error flags despite unchanged semantics. Anchoring and adjustment bias appears when expert cues shift participant decisions beyond parity, consistent with salience and availability effects. We also observe the representativeness heuristic: atypical but valid attribute combinations are frequently flagged as erroneous, and in entity matching tasks, surface similarity produces a substantial false-positive rate with high confidence. In data repair, participants show a robust preference for leaving values missing rather than imputing plausible values, consistent with omission bias. In contrast, automation-aligned switching under strong contradiction does not exceed a conservative rare-error tolerance threshold at the population level, indicating that deference to automated recommendations is limited in this setting. Across scenarios, bias patterns persist among technically experienced participants and across diverse workflow practices, suggesting that bias in data cleaning reflects general cognitive tendencies rather than lack of expertise. These findings motivate human-in-the-loop cleaning systems that clearly separate representation from semantics, present expert or algorithmic recommendations non-prescriptively, and support reflective evaluation of atypical but valid cases.