SHAP analysis identifies Hapax Ratio as the most influential stylometric feature in Random Forest predictions, followed by NSR, Noun Ratio, Adverb Ratio, and TTR¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Global SHAP importance across held-out windows ranked Hapax Ratio first, followed by NSR, Noun Ratio, Adverb Ratio, and TTR; Adjective Ratio contributed least. → Rajesh Kumar 2026

Evidence¶

Rajesh Kumar 2026¶

Rajesh Kumar, Nabeel Siddiqui, Alexander Fuchsberger. (2026). Detecting GPT-Assisted Writing Using Interpretable Stylometric Features. https://arxiv.org/abs/2609.26687 doi:10.48550/arxiv.2609.26687

q2 · i? · design · r2

SHAP analysis of the Random Forest on held-out windows (Figures 3–4). Local examples showed a GPT-assisted median window at probability 0.947 with NSR and Hapax Ratio making the largest positive contributions, and a human window at 0.077.

"Hapax Ratio is the most influential feature, followed by NSR, Noun Ratio, Adverb Ratio, and TTR. Verb Ratio, Sentence-Length Variability, and Word Entropy have smaller contributions, while Adjective Ratio contributes least."

Discussion¶

Related Claims¶

• GPT-assisted writing shows higher lexical diversity, richness, and noun/non-stopword ratios, while independently authored writing shows higher adverb ratio and sentence-length variability — related
• Document length alone carries some signal (word-count classifier ROC-AUC 0.68), but the fixed-word sliding-window design largely prevents stylometric features from capturing length differences — related
• Random Forest using nine interpretable stylometric features detects GPT-assisted student writing with held-out ROC-AUC of 0.870 and F1 of 0.842 — related
• Academic word list count, word count, and Flesch-Kincaid grade level are the most important features for predicting cognitive engagement in discussion posts — related
• In the random forest, word count was the most important readability feature while traditional readability formulas ranked near the bottom — related
• Engineered practice-balance features outperform raw activity counts: the Intensity Ratio ranks third among Codeforces predictors, and upsolving habit dominates the survey model — a broader claim this one bears on
• Feature-based model predictive power derives from a large and varied set of features rather than a few dominant variables — related
• The stylometric detector misclassified 4 of 18 independently authored documents as GPT-assisted (FPR 22.2%) and 2 of 18 GPT-assisted documents (FNR 11.1%), with wide confidence intervals — related
• Humanization works by systematically suppressing detector-salient surface cues — long-token ratio, AWL density, and lexical diversity — rather than fragmenting syntax — related