Experimental Results on Data Analysis Algorithms for Extracting and Interpreting Edge Feature Data for Duct Tape and Textile Physical Fit Examinations

• NCJ Number: 308591
• Journal: Journal of Forensic Sciences Volume: Online Dated: 2023
• Author(s): Meghan Prusinowski; Pedram Tavadze; Zachary Andrews; Logan Lang; Divyanjali Pulivendhan; Cedric Neumann; Aldo H. Romero; Tatiana Trejos
• Date Published: December 2023
• Length: 17 pages
• Grant Number(s): 2020-DQ-BX-0012
• Agencies: NIJ-Sponsored
• Publication Type: Research (Applied/Empirical)
• Award: Assessing the Strength of Trace Evidence Fracture Fits Through a Comprehensive, Systematic, and Quantifiable Approach

Annotation

This article describes the creation of a data analysis algorithm that uses mutual information and a decision tree, to support forensic practitioners with the interpretation of trace evidence; the paper provides details on the research study and algorithm development, research results, and implications for practitioners. 

Abstract

A physical fit is an important observation that can result from the forensic analysis of trace evidence as it conveys a high degree of association between two items. However, physical fit examinations can be time-consuming, and potential bias from analysts may affect judgment. To overcome these shortcomings, a data analysis algorithm using mutual information and a decision tree has been developed to support practitioners in interpreting the evidence. The authors created these tools using data obtained from physical fit examinations of duct tape and textiles analyzed in previous studies, along with the reasoning behind the analysts' decisions. The relative feature importance is described by material type, enhancing the knowledge base in this field. Compared with the human analysis, the algorithms provided accuracies above 90 percent, with an improved rate of true positives for most duct tape subsets. Conversely, false positives were observed in high-quality scissor cut (HQ-HT-S) duct tape and textiles. As such, it is advised to use these algorithms in tandem with human analysis. Furthermore, the study evaluated the accuracy of physical fits when only partial sample lengths are available. The results of this investigation indicated that acceptable accuracies for correctly identifying true fits and non-fits occurred when at least 35 percent of a sample length was present. However, lower accuracies were observed for samples prone to stretching or distortion. Therefore, the models described here can provide a valuable supplementary tool but should not be the sole means of evaluating samples. (Published Abstract Provided)