犯人像推定における統合的分類アルゴリズムの頑健性の検証

抄録

An integrated classification algorithm is a decision-making method that is not limited to a single classifier. It comprises multiple classifiers to maintain a high classification performance for various datasets. This study investigated the feasibility of an integrated classification algorithm for offender profiling. Offender profiling is the analysis of a crime scene using statistical and psychological methods to estimate information such as the age, job, and criminal record of the offender. In this study, the following 12 machine learning algorisms were used: decision tree (C5.0, CART by entropy or Gini), logistic regression analysis (LR), naïve bayes (NB), random forest (RF), bagging, boosting, support vector machine (SVM by radial basis function or polynomial), k-nearest neighbor (KNN), and neural network (NN). The results of the study showed that the classification performances of each algorithm varied for different objective variables of the dataset (e.g., criminal record, age, or job of offenders of residential burglar). However, the majority decisions made by a combination of three classifier algorithms (e.g., decision tree, LR, and NB) showed high classification performance regarding any dataset.