Paw-sitive Identification: Machine Learning with Biometrics Improves Canid Detection

Summary

This is an abstract from the "Machine-Learning Approaches to Studying Ancient Human-Environmental Interactions" session, at the 90th annual meeting of the Society for American Archaeology.

Zooarchaeological canid identifications are made using an array of techniques, many of which were only ever designed to separate dogs from wolves and have never been tested against large samples. Skeletal measurements (termed biometrics) coupled with statistical analyses can improve identification replicability, but require large sample sizes to capture morphological variability and precise modeling to gauge prediction error. Machine learning techniques offer a way to overcome several related issues through processes of data splitting and the calculation of model performance metrics. Here, we apply several modeling algorithms (Linear Discriminant Analysis, Random Forest, and K-Nearest Neighbors) on a dataset of 957 canid mandibles from ten North American canid taxa. We then apply these models to archaeological specimens from the U.S. Southwest to reassess previous identifications. We argue that improved canid detection must fully incorporate measures of uncertainty, machine learning helps us estimate uncertainty, and that machine learning improves data quality in zooarchaeology.

This Resource is Part of the Following Collections

• Society for American Archaeology 90th Annual Meeting, Denver, CO (2025)   •
• Machine-Learning Approaches to Studying Ancient Human-Environmental Interactions

Keywords

Record Identifiers

Abstract Id(s): 54005