Feature Selection for Character N-Grams in Romanian Authorship Attribution
Character N-gram representations are widely used in authorship attribution but often generate extremely high-dimensional feature spaces, making feature selection an important design decision. This study investigates the impact of eight feature selection methods on Romanian authorship attribution using four classifiers and character 3-, 4-, and 5-gram representations. The best overall result was achieved by a Random Forest classifier using F-test filtering with only 500 3-gram features. Despite using substantially fewer features, the model achieved performance comparable to the best results obtained by models trained on the full feature set and exceeded it by a small margin. These findings highlight the potential of feature selection to develop efficient systems without sacrificing accuracy.
Authors
- Dana Lupsa
- Radu Lupșa
Institutions
- Babeș-Bolyai University (RO)
Publication Details
- Journal
- Studia Universitatis Babeș-Bolyai Informatica
- Published
- 2026-11-18
- DOI
- https://doi.org/10.24193/subbi.2026.07
- Primary Topic
- Authorship Attribution and Profiling
- Type
- article
- Field-Weighted Citation Impact
- 0.00