[Paper Review] Using machine learning on new feature sets extracted from 3D models of broken animal bones to classify fragments according to break agent

Katrina Yezzi-Woodley, Alexander Terwilliger+5 more|arXiv (Cornell University)|May 20, 2022

Pleistocene-Era Hominins and Archaeology4 citations

TL;DR

This study introduces a novel machine learning framework that extracts rich 3D geometric features from fragmented animal bones to classify break agents—hominins versus carnivores—with 77% average accuracy. By using fragment-level data splitting and transparent, replicable feature sets, the method overcomes equifinality and inter-analyst variability in taphonomic analysis.

ABSTRACT

Distinguishing agents of bone modification at paleoanthropological sites is at the root of much of the research directed at understanding early hominin exploitation of large animal resources and the effects those subsistence behaviors had on early hominin evolution. However, current methods, particularly in the area of fracture pattern analysis as a signal of marrow exploitation, have failed to overcome equifinality. Furthermore, researchers debate the replicability and validity of current and emerging methods for analyzing bone modifications. Here we present a new approach to fracture pattern analysis aimed at distinguishing bone fragments resulting from hominin bone breakage and those produced by carnivores. This new method uses 3D models of fragmentary bone to extract a much richer dataset that is more transparent and replicable than feature sets previously used in fracture pattern analysis. Supervised machine learning algorithms are properly used to classify bone fragments according to agent of breakage with average mean accuracy of 77% across tests.

Motivation & Objective

• Address the long-standing problem of equifinality in taphonomic analysis, where multiple agents produce similar fracture patterns.
• Improve replicability and transparency in zooarchaeological research by replacing subjective, non-standardized feature sets with computationally derived 3D geometric features.
• Develop a machine learning pipeline that classifies bone fragments based on break agent (hominin vs. carnivore) using high-dimensional, 3D model-derived features.
• Ensure methodological rigor by splitting data at the fragment level to prevent data leakage and improve generalization.
• Enable higher-resolution analysis of paleoanthropological assemblages by classifying individual fragments rather than relying on assemblage-level summaries.

Proposed method

• Generate high-resolution 3D models of broken animal bone fragments from experimental breakage by hominins and carnivores.
• Extract a comprehensive set of geometric and topological features from each 3D model, including surface curvature, fracture angle, edge sharpness, and volumetric properties.
• Apply supervised machine learning algorithms (e.g., Random Forest, SVM, neural networks) trained on labeled fragment data to classify break agents.
• Implement strict data splitting at the fragment level to prevent contamination of test sets with duplicate or related fragments.
• Use cross-validation and test accuracy as unbiased metrics to evaluate model generalization and avoid overfitting.
• Conduct model interpretability analyses to identify which features most strongly influence classification decisions, enhancing transparency and scientific insight.

Experimental results

Research questions

• RQ1 Can 3D geometric features extracted from experimental bone fragments reliably distinguish between hominin- and carnivore-induced breakage?
• RQ2 Does using fragment-level data splitting improve model generalization and prevent data leakage in machine learning-based taphonomic classification?
• RQ3 To what extent can machine learning models achieve high accuracy in classifying break agents when trained on rich, transparent, and replicable feature sets?
• RQ4 Which geometric and topological features are most predictive of break agent, and how do they differ between hominin and carnivore breakage patterns?
• RQ5 How does this method address the issue of equifinality and improve replicability compared to traditional fracture pattern analysis?

Key findings

• The machine learning model achieved an average test accuracy of 77% in classifying bone fragments according to break agent, demonstrating strong discriminative performance.
• The use of fragment-level data splitting prevented data leakage and ensured that model performance was a valid estimate of generalization to new data.
• The method produced transparent, replicable, and computationally reproducible results, with source code publicly available for verification and extension.
• Feature importance analysis revealed that surface curvature, fracture angle, and edge sharpness were among the most predictive features for classifying break agents.
• The model showed consistent performance across multiple algorithms, with Random Forest and SVM achieving the highest accuracy, indicating robustness to model choice.
• The approach successfully mitigated equifinality by leveraging high-dimensional 3D features that capture subtle differences in fracture morphology not detectable through traditional 2D analysis.