MSc Thesis Defense: Süleyman Onur Doğan, COFINE: HIERARCHY-AWARE SEMI-SUPERVISED LEARNING FOR FINE-GRAINED BEHAVIORAL CLASSIFICATION, Date & Time: 08 July, 2026 – 1:00 PM, Place: FENS L027

COFINE: HIERARCHY-AWARE SEMI-SUPERVISED LEARNING

FOR FINE-GRAINED BEHAVIORAL CLASSIFICATION

SÜLEYMAN ONUR DOĞAN
Computer Science & Engineering, MSc Thesis, 2026

Thesis Jury

     Assoc. Prof. Öznur Taştan (Thesis Advisor)

  Prof. Berrin Yanıkoğlu

  Assoc. Prof. Ramazan Gökberk Cinbiş

Date & Time: July 8th, 2026 – 1:00 PM

Place: FENS L027

Keywords : Hierarchical semi-supervised learning, Fine-grained visual recognition,

Representation learning, Behavioural classification, Computational ethology

Abstract

Sleep research in Drosophila melanogaster has traditionally relied on coarse measurements such as locomotor inactivity. Although scalable, these approaches may miss subtle associated microbehaviours. Automated behavioural classification is therefore important for scalable and reproducible analysis of fine-grained behavioural states in future biological studies. Biological datasets often contain annotations at different levels of granularity, where coarse labels are easier to obtain and may be available for all samples, whereas fine-grained labels require expert annotation and are often limited. To address this problem, we propose CoFINE, a hierarchyaware semi-supervised learning algorithm for fine-grained behavioural classification. CoFINE uses the known coarse-to-fine relationship between behavioural classes as structural supervision and combines supervised fine-grained classification, coarselevel supervision, coarse-guided pseudo-labeling, hierarchical contrastive learning, and entropy regularization. Experiments on a hierarchically labelled Drosophila behavioural dataset with three coarse behaviours and nine fine-grained subclasses show that CoFINE outperforms classical and semi-supervised baselines, achieving a fine-grained F1-macro of 0.787 ± 0.020 and a coarse-grained F1-macro of 0.962 ± 0.012. These results demonstrate that hierarchy-aware semi-supervised learning is effective when fine-grained annotations are scarce but coarse labels are abundant.