Preprints

• Sequential Large Language Model-Based Hyper-parameter Optimization (SLLMBO)
  Kanan Mahammadli, Seyda Ertekin
  arXiv preprint, 2024 · View on arXiv
  • Developed SLLMBO, a sequential hyperparameter-optimization framework combining LLM-guided configuration proposals with Tree-structured Parzen Estimation.
  • Evaluated four LLMs across 14 classification and regression tasks; the proposed LLM-TPE sampler achieved better results than the evaluated Bayesian-optimization methods on 9 tasks.
• Class-Specific Data Augmentation: Bridging the Imbalance in Multiclass Breast Cancer Classification
  Kanan Mahammadli, Abdullah Burkan Bereketoglu, Ayse Gul Kabakci
  arXiv preprint, 2023 · View on arXiv
  • Studied how applying a uniform augmentation strategy across classes, together with relying on aggregate evaluation metrics, can obscure degraded performance on minority classes.
  • Developed a class-specific augmentation pipeline using stain normalization and a transfer-learning Vision Transformer, improving minority-class performance on the BreaKHis dataset relative to the evaluated baseline models.