Feature Set Formation and Comparative Analysis of Classification Algorithms for AI Generated Code Detection

Abstract: The paper presents a comprehensive approach to to construct-ing a feature space for detecting artificially generated Python source code. We developed the Algorithmic_Analyzer class to extract 27 features categorized into four groups: basic code met-rics, structural patterns, keywords, and libraries. Additionally, lexical patterns are captured using word n-grams. Experiments using classical machine learning algorithms demonstrate that structural characteristics exhibit significantly higher signifi-cance than lexical features. The study identifies the most in-formative features for artificial code detection and establishes that the XGBClassifier model achieves the best performance, with an average F1_macro score of 0.90.

Keywords: code classification, feature analysis, language models, source code, machine learning