Λογότυπο αποθετηρίου

Πλοήγηση ανά Συγγραφέα "Chatzipapadopoulou, Anna"

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω
Φιλτράρετε τα αποτελέσματα πληκτρολογώντας τα πρώτα γράμματα
Τώρα δείχνει 1 - 1 από 1
  • Μικρογραφία εικόνας
    Τεκμήριο
    Ensemble learning in uncertainty quantification for multi-label prediction: from theoretical foundations to medical image understanding
    (2025-10-01) Chatzipapadopoulou, Anna ; Χατζηπαπαδοπούλου, Άννα; Koutsopoulos, Iordanis; Pavlopoulos, John; Androutsopoulos, Ion
    Reliable prediction in high-stakes domains requires methods that can handle complex outputs and quantify their own uncertainty. Multi-label classification (MLC) provides the framework to address tasks where multiple labels may be simultaneously relevant, while conformal prediction (CP) provides principled guarantees on the reliability of its uncertainty estimates. This thesis investigates ensemble strategies that combine these two approaches, and in a second part applies multi-label classification methods, including ensemble-based techniques, to the task of medical concept detection. In the first part, we present a theoretical and empirical study of conformal ensembles, combining the formal coverage guarantees of CP with the robustness and diversity benefits of homogeneous and heterogeneous ensembles. We propose an ensemble conformal prediction (ECP) framework for multilabel classification, in which individually conformalized models are aggregated using standard strategies such as majority voting, probability averaging, and F1-weighted fusion. We adapt existing theoretical results to analyze coverage properties under these ensembles, and evaluate their performance across benchmark datasets. Results demonstrate that conformal ensembles consistently improve macro-F1 while maintaining valid coverage, and at the same time produce more compact and informative prediction sets compared to single-model or post-hoc conformal baselines. In the second part, we address the task of multi-label medical image concept detection, examining a range of architectures and strategies, including ensemble-based methods, as part of our participation in the ImageCLEFmedical Caption 2025 challenge. Our approach employs CNN–FFNN architectures with various backbone encoders, per-label threshold optimization to address extreme label imbalance, and diverse ensemble aggregation strategies, including union, intersection, and consensus-driven methods. Experiments on the ImageCLEFmedical dataset show that these ensembles achieved highly competitive performance in concept detection, ranking first in the 2025 competition.