A Classification Model for Drug Addicts Based on Improved Random Forests Algorithm

Abstract

To accomplish the rapid screening of drug addicts and to meet the requirements of modern police work, this research employs a data-mining technology that utilizes real samples of drug addicts as well as non-drug addicts. The aim is to construct a classification model based on pulse wave data. After the pre-processing of pulse wave data, the original random forest classification model is initially established with high accuracy, but with a relatively low recall rate and F1 score. To resolve this issue, an improved classification model is henceforth proposed. The improved model mainly involves three improvement strategies: firstly, perform cross-validation by dividing multiple training sets and test sets to obtain generalization errors; secondly, balance the sample distribution using down-sampling techniques; and finally, select model parameters based on multi-criteria analysis. According to the evaluation results of accuracy, precision, recall rates, and F1 scores, the performance of the improved random forest classification model has demonstrated its superiority and robustness via experiments using different datasets.