Duration: 0:00

38 views

0

Danai Kafetzaki, Jeroen Poblome

Of all archaeological material, pottery provides the most complete dataset for assessing chronological sequences of deposits. Count, weight, functional types, and stratigraphic information can assist as variables to quantitatively infer relative chronology. In recent years, machine learning algorithms have increasingly supplemented archaeological approaches to seriation. In this case study, we present the results of a seriation of Sagalassos Red Slip Ware (25 BCE – 700 CE) datasets, using the Principal Component Analysis algorithm to regenerate the relative time sequence of the deposits. The 9 previously defined chronological phases provide labels for our observations used in the validation of the results. Initially, the unsupervised learning application provides the first component that is used in seriation, as well as additional components that can be used in a multi-dimensional time approach to deal with the complex nature of temporality. We then select components based on factor loadings, so that the new axes convey the diversification of phases. For this lower-dimensional dataset, clustering algorithms are implemented to decide on group membership of the contexts. Finally, we predict and evaluate the estimated relative position of contexts in other SRSW datasets, based on a combination of this data-driven approach and specialist knowledge application. We illustrate every aspect of the developed method using Rstudio and JavaScript, allowing archaeologists to observe and interact with every step of the analysis and incorporate prior information where needed. This paper will thus show how machine learning algorithms and visual analytics can be used to refine and complement traditional archaeological methods.