Mining Convict Characteristics from a Quantitative History Perspective: A Study on Data Association and Algorithm Adaptation in the Norfolk Island Penal Colony

Abstract

To overcome the limitations of traditional historical research, which is predominantly qualitative and lacks quantitative depth, this study analyzes a dataset of 6,473 convict records from the Norfolk Island Penal Colony (1788–1855), focusing on two key research directions. First, it explores the association patterns between prisoners’ marital status and factors such as literacy, religion, and offense type. Second, it compares the performance of different classification algorithms in predicting key convict attributes—literacy and death in custody. The findings reveal that: (1) association rule mining identifies a weak yet meaningful link between single status and the profile of “illiterate, Roman Catholic, minor offense” (lift 1.05–1.06), a pattern that aligns with the social stratification of the colonial era; (2) in classification tasks, logistic regression performs best for literacy prediction (79% accuracy), while prediction of death in custody remains limited overall (ROC AUC 0.60–0.64) due to inherent data characteristics, though decision trees effectively identify age and sentence length as key risk factors. Furthermore, this study proposes a three-dimensional framework—“differentiated preprocessing, algorithm adaptation, and historical contextualization”—to provide a reusable paradigm for interdisciplinary research using historical data.