Identifying non-anthropogenic accumulation in zooarchaeological assemblages using naive Bayesian classifier: A trace-oriented actualistic taphonomic approach in the hyperarid coasts of the Atacama desert

Abstract

The taphonomic evaluation of zooarchaeological assemblages is paramount for compelling interpretations of past subsistence strategies, palaeoecological interactions and overall biasing of the anthropogenic bone deposition and modification by non-anthropogenic agents. A recent trend in this field, has focused on bone surface modifications (BSM) to successfully generate reference frameworks for critical evaluation of the interpretive potential of selected zooarchaeological samples by examining BSM patterns produced by known agents in experimental settings. These approaches compare anthropogenic and non-human generated patterns using different machine learning (ML) algorithms, classifying and assigning sets of traces to specific processes such as gnawing, scoring or intentional fracture. We propose a simplification in two key aspects to make this new arising area of ML assisted taphonomic interpretation more accessible for zooarchaeologists. First, we encourage a trace oriented multivariate data recording protocol for taphonomic analysis including the most commonly recorded BSM plus others such as shape modification, penetrative modification and element or tissue loss modifications thus expanding comparative potential. Second, we offer a ML approach based on Naïve Bayes Classifier (NBC) that discriminates each given observation by comparing to actualistic taphonomic patterns to determine the average percentage of non-anthropogenic modification within zooarchaeological samples. To illustrate the potential value of these proposals for taphonomical interpretation, we analyzed seven different Mid to Late Holocene assemblages containing avian and mammal remains from the coastal platform in the hyperarid core of the Atacama Desert. We compared their alteration with a local sample of actualistic data recovered to characterize regional taphonomic patterns. A NBC was trained with a 75% random selection of the total archaeological and actualistic samples. Afterwards we tested the remaining sample, achieving over 90% accuracy in determining whether a random specimen from any sample corresponded to an actualistic or an archaeological context implying key differences in taphonomic trace patterns.