Feature Selection for Aiding Glass Forensic Evidence Analysis

The evaluation of glass evidence in forensic science is an important issue. Traditionally, this has depended on the comparison of the physical and chemical attributes of an unknown fragment with a control fragment. A high degree of discrimination between glass fragments is now achievable due to advances in analytical capabilities. A random effects model using two levels of hierarchical nesting is applied to the calculation of a likelihood ratio (LR) as a solution to the problem of comparison between two sets of replicated continuous observations where it is unknown whether the sets of measurements shared a common origin. Replicate measurements from a population of such measurements allow the calculation of both within-group and between-group variances. Univariate normal kernel estimation procedures have been used for this, where the between-group distribution is considered to be non-normal. However, the choice of variable for use in LR estimation is critical to the quality of LR produced. This paper investigates the use of feature selection for the purpose of selecting the variable for estimation without the need for expert knowledge. Results are recorded for several selectors using normal, exponential, adaptive and biweight kernel estimation techniques. Misclassification rates for the LR estimators are used to measure performance. The experiments performed reveal the capability of the proposed approach for this task.