A Survey on Confidence Calibration of Deep Learning-Based Classification Models Under Class Imbalance Data

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Confidence calibration is crucial for accurate probability estimation in classification models.
• Class imbalance data negatively impacts model learning and probability predictions, particularly for underrepresented classes.
• Existing research often addresses confidence calibration and class imbalance separately.

Purpose of the Study:

• To comprehensively investigate confidence calibration methods for deep learning models under class imbalance.
• To theoretically explore the impact of class imbalance on confidence calibration.
• To provide practical guidelines and insights for selecting appropriate methods.

Main Methods:

• Review of 60 state-of-the-art confidence calibration methods for imbalanced data.
• Categorization of methods into six groups based on differences.
• Systematic comparison of seven properties and evaluation methodologies (datasets, metrics).
• Conducting comparative experiments to validate findings.

Main Results:

• Class imbalance significantly challenges confidence calibration, especially for minority classes.
• Theoretical explanations provided for empirical observations in existing studies.
• Comparative analysis identifies strengths and weaknesses of different calibration methods.

Conclusions:

• The study offers a systematic overview and comparative analysis of confidence calibration techniques for imbalanced datasets.
• Experimental results provide practical guidance for researchers and practitioners.
• Identifies promising future research directions in this critical area.
• Meta_Agnosticism is a philosophical stance that questions the possibility of knowing ultimate reality or the existence of God. It suggests that certain truths are unknowable or beyond human comprehension. This abstract explores the concept of agnosticism, its historical context, and its implications for various fields, including philosophy, theology, and ethics. The study examines different interpretations of agnosticism, from weak agnosticism (lack of evidence) to strong agnosticism (inherent unknowability). It also discusses the relationship between agnosticism and atheism, highlighting their distinct approaches to belief and disbelief. The research delves into the ethical and societal consequences of agnosticism, considering its impact on moral frameworks and personal decision-making. Furthermore, the study explores the role of agnosticism in scientific inquiry and the limitations of empirical knowledge. The findings suggest that agnosticism, while challenging traditional belief systems, offers a nuanced perspective on the nature of knowledge and the human condition. The abstract concludes by emphasizing the importance of intellectual humility and open-mindedness in navigating complex questions about existence and reality.Meta Description: Explore agnosticism: the philosophical view that ultimate reality is unknowable. This study examines its history, interpretations, and impact on ethics and science.
• Meta_Description