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

Abstract

It is 50 years since Sieveking et al. published their pioneering research in Nature on the geochemical analysis of artefacts from Neolithic flint mines in southern Britain. In the decades since, geochemical techniques to source stone artefacts have flourished globally, with a renaissance in recent years from new instrumentation, data analysis, and machine learning techniques. Despite the interest over these latter approaches, there has been variation in the quality with which these methods have been applied. Using the case study of flint artefacts and geological samples from England, we present a robust and objective evaluation of three popular techniques, Random Forest, K-Nearest-Neighbour, and Support Vector Machines, and present a pipeline for their appropriate use. When evaluated correctly, the results establish high model classification performance, with Random Forest leading with an average accuracy of 85% (measured through F1 Scores), and with Support Vector Machines following closely. The methodology developed in this paper demonstrates the potential to significantly improve on previous approaches, particularly in removing bias, and providing greater means of evaluation than previously utilised.

Perdiz arrow points from Caddo burial contexts aid in defining discrete behavioral regions

“…A support vector machine is a supervised machine learning model regularly used in classifying archaeological materials [14,49,29,37,80,28], which has utility in comparing and classifying datasets aggregated from digital repositories, comparative collections, open access reports, as well as other digital assets. For this effort, linear data were imported and modeled using the scikit-learn Fig.…”

Use and Misuse of Machine Learning in Anthropology

“…To date, ML has been applied in a number of lithic studies addressing a wide variety of anthropological questions: identifying heat-treated raw material nodules, a practice employed to improve the ease of working raw nodules into stone artifacts [15]; identifying the materials worked by a stone tool according to the classification of the use-wear created on its edge [16], [17]; predicting the original flake mass from variables on the striking platform in order to quantify the degree of resharpening (and thus the length of its use-life as a tool) [18]; predicting site formation conditions from the surface alteration of the site's lithic artifacts [19]; creating more quantitatively rigorous approaches to the creation of typologies for studying artifact shape through time and space [20], [21]; predicting the raw material of the stone tool from the cut marks produced by the edge [22]; identifying the geochemical signatures of geological sources of lithic raw materials as a means of studying prehistoric mobility and material selection criteria [23], [24]; distinguishing the flake products from different reduction strategies for exploiting the volume of a core [25]; distinguishing chronological manifestations of lithic behavior between the Middle and Late Stone Age in Africa through the presence vs. absence of types within assemblages [26]; developing virtual knapping software [27]; and quantifying lithic knapping skill acquisition for studying the evolution of human cognition [28].…”

“…The studies [3], [4] apply standardization of the data before a train-test split, while the studies [8], [11] (and other works by these authors) apply PCA for dimension reduction before the train-test split. In [24] the authors use the t-SNE embedding to visualize the dataset in two dimensions and manually remove "outliers" before the train-test split. Due to the difficulty interpreting the t-SNE embedding, the "outliers" removed could in fact be valid datapoints that are simply difficult to classify, thereby artificially increasing accuracy scores.…”

Perdiz arrow points from Caddo burial contexts aid in defining discrete behavioral regions

“…A support vector machine is a supervised machine learning model regularly used in classifying archaeological materials [36,37,38,39,40,41], which has utility in comparing and classifying datasets aggregated from digital repositories, comparative collections, open access reports, as well as other digital assets. For this effort, linear data were imported and modeled using the scikit-learn package in Python [42,43] (supplementary materials), and subsequently split into training (75 percent) and testing (25 percent) subsets.…”