ClassShift: Improving Fixation-to-Class Assignment in UML Class Diagrams

Gaze offset in eye tracking studies biases AOI-based metrics. In UML studies, classes are often operationalized as AOIs and enlarged to mitigate offset, threatening reproducibility. We propose ClassShift, an ensemble-based post-hoc offset-compensation algorithm that improves fixation-to-class assignment in UML class diagrams without enlarging AOIs. We evaluate ClassShift on synthetic data with ground-truth fixation-to-class labels (9332 scenarios) and real eye tracking data derived from 32 participants. Evaluation on synthetic data shows a 19% Macro-F1 improvement relative to baseline containment-based labeling. In the 9.6% of scenarios where ClassShift underperforms, losses are small (median ΔMacro-F1 = -0.033). On real data, visual inspection suggests plausible fixation-to-class assignments. Participant self-reports further indicate which classes were reported as having been inspected particularly long. After applying ClassShift, relative dwell time for these classes increases significantly over the uncorrected data (𝑝 = .028, 𝑟 = 0.50). Together, these results suggest ClassShift improves eye tracking analyses under systematic offset.