Comparing the Performance of Machine Learning and Traditional Approaches to Archaeological Site Modeling and Prediction

Summary

This is an abstract from the "Novel Statistical Techniques in Archaeology II (QUANTARCH II)" session, at the 84th annual meeting of the Society for American Archaeology.

Site prediction models have helped archaeological resource management in site prospecting, impact mitigation, and information recovery. Beginning in 2009, we developed probability models for the Shoshone National Forest (SNF). These models helped to prioritize inventory of areas burned in wildfires, to rapidly appraise archaeologically sensitive areas. The methodological tool most used in model-building has been logistic regression. However, developments in Data Science via traditional and Machine Learning algorithms provide opportunities to improve predictive ability. Here, we use SNF archaeological presence/absence location data with several modeling techniques to create and compare predictions. Predictive methods include Generalized Linear Modeling, Generalized Additive Modeling, MaxEnt, Multilayer Perceptron, Conditional Inference Trees, Recursive Partitioning and Regression Trees, Random Forest, Flexible and Mixture Discriminant Analysis, Radial Basis Function, Support Vector Machine, Gradient Boosting Machine Learning, Partial Least Squares Regression, k-Nearest Neighbors, and Neural Networks. We compared predictions using 17 diagnostics including gain, true/false positive rates, true/false negative rates, likelihood ratios, diagnostic odds ratio, accuracy, positive/negative predictive values, false discovery/omission rates, the F1 score, Youden’s J statistic, markedness, and Matthews coefficient. Techniques offer prediction tradeoffs depending on the diagnostics used. We recommend users become familiar with the diagnostic tools that best reflect modeling goals.

This Resource is Part of the Following Collections

• Novel Statistical Techniques in Archaeology II (QUANTARCH II)   •
• Society for American Archaeology 84th Annual Meeting, Albuquerque, NM (2019)

Keywords

Record Identifiers

Abstract Id(s): 25262