Evaluating machine learning techniques for archaeological lithic sourcing: a case study of flint in Britain

Area of Science:

• Archaeological science
• Geochemistry
• Computational archaeology

Background:

• Geochemical sourcing of stone artefacts has evolved significantly since 1970s research.
• Recent advances in instrumentation, data analysis, and machine learning have spurred global interest.
• Variability in the application quality of these advanced techniques necessitates robust evaluation.

Purpose of the Study:

• To provide an objective evaluation of popular machine learning techniques for geochemical sourcing of artefacts.
• To present a standardized pipeline for the appropriate application of these methods.
• To improve upon previous approaches by reducing bias and enhancing evaluation metrics.

Main Methods:

• Utilized a case study of flint artefacts and geological samples from England.
• Evaluated three machine learning techniques: Random Forest, K-Nearest-Neighbour, and Support Vector Machines.
• Developed and applied a robust pipeline for objective model evaluation, including F1 Score measurement.

Main Results:

• Achieved high model classification performance when techniques were evaluated correctly.
• Random Forest demonstrated the highest average accuracy at 85% (F1 Score).
• Support Vector Machines also showed strong performance, closely following Random Forest.

Conclusions:

• The developed methodology offers significant improvements over previous artefact sourcing approaches.
• The pipeline effectively mitigates bias and provides enhanced means for evaluating geochemical sourcing models.
• This approach enhances the reliability and objectivity of sourcing stone artefacts using geochemical data and machine learning.