Abstract:

Isolated bovid teeth dominate the fossil record of southern Africa and offer important clues about past environments and early human behavior, but classifying them is difficult, especially when specimens are only partially preserved. This talk presents a shape-based machine learning framework using elastic shape analysis and the square root velocity framework (SRVF) to classify both fully and partially observed fossil teeth. For complete teeth, SRVF-based features generally outperform elliptical Fourier analysis, and accuracy improves substantially when size information is included. For incomplete teeth, we propose a hot-deck imputation method that draws on similar complete teeth as donors to generate multiple plausible completions. Simulations and an application to fossil teeth from Gladysvale, South Africa, show that imputation can improve species-level classification in some settings, although direct classification of partial shapes often performs better at the tribe level. These results illustrate the potential of shape-based methods to refine taxonomic classification and paleoenvironmental reconstruction, while highlighting the limitations of imputation.