Research on the Application of Feature Engineering-Driven Machine Learning in STR Profile Recognition

Abstract

In forensic evidence identification, STR analysis technology is the core, and the analysis of mixed STR profiles in multi-person criminal cases is crucial for case judgment. The key issues are to identify the number of contributors in mixed samples, infer proportions, match genotypes, and reduce noise. The research content involves constructing models for the above issues, solving them, and conducting evaluations. The methods include using machine learning models such as random forests, weighted inversion models combined with differential evolution algorithms, chi-square test plus random forests, and wavelet transform combined with CNN for noise reduction. The core innovations lie in the accurate identification by fusing multiple models, the proposal of the weighted inversion model, efficient genotype matching, and the excellent effect of combined noise reduction. The conclusions show that the random forest performs the best, the proportion inference is effective, and the noise reduction model has a good goodness of fit.