MSc Thesis Defense: Muhammet Nurullah Cebeci, CONSTRAINED CONTEXTUAL BANDITS FOR RESOURCE-LIMITED EEG DECODING: A LEAKAGE-CONTROLLED, MULTI-PARADIGM EVALUATION, Date & Time: 16 July, 2026 – 1:00 PM, Place: L029

CONSTRAINED CONTEXTUAL BANDITS FOR RESOURCE-LIMITED EEG DECODING: A LEAKAGE-CONTROLLED, MULTI-PARADIGM EVALUATION

Muhammet Nurullah Cebeci
Computer Science and Engineering, MSc Thesis, 2026

Thesis Jury

     Prof. Selim Balcısoy (Thesis Advisor)

  Prof. Ayşe Berrin Yanıkoğlu

  Asst. Prof. Günet Eroğlu

Date & Time: 16th July, 2026 – 1:00 PM

Place: FENS L029

Keywords : constrained contextual bandits, online pipeline selection, cognitive load, near-ear wearable EEG, data leakage

Abstract

Electroencephalography (EEG)-based Brain-Computer Interfaces (BCIs) traditionally decode mental states from dense electrode montages that are impractical for wearable, low-resource use, where channel count, computation, and time are tightly constrained. This thesis investigates whether a Constrained Contextual Bandit (CCB) --- which selects, online and per trial, among EEG feature-extraction pipelines under a fixed compute budget --- can decode cognitive load from a minimal, near-ear montage, and how it compares with fixed classical pipelines, a compact convolutional network (EEGNet), and each task's published benchmark, including under the cross-session drift of all-day wear. The CCB is instantiated over a low-channel signal with paradigm-specific Action, Constraint, and Context definitions and evaluated under a strict no-leakage discipline: a higher-channel recording never informs a lower-channel pipeline, and the near-ear (T7/T8) montage is obtained by electrode position rather than any dense-montage statistic. The evaluation panel spans cognitive-load and motor-imagery datasets from two to sixty-four channels. The central result is a deployment-regime ceiling. Within-subject cross-validation on block-homogeneous cognitive-load labels is leakage-confounded, inflating Cohen's kappa; under leakage-clean cross-session and leave-one-subject-out protocols, two-channel near-ear cross-session decoding is hard (kappa approximately 0.06--0.22), and neither the CCB nor EEGNet rescues it. A best-arm diagnostic localises the residual gap to the arm bank rather than the bandit's online selection. The contribution is methodological: a reproducible, multi-paradigm, leakage-controlled characterisation of where such bandits help for low-channel BCI, where they fail, and why.