Chapter 1, Introduction to Data Imbalance in Machine Learning, serves as an exploration of data imbalance within the context of ML. This chapter elucidates the nature of imbalanced data, distinguishing it from other dataset types. It also provides a comprehensive introduction to the essential components of ML and model performance metrics most relevant for cases when there is a data imbalance. The chapter looks into the issues and concerns involved in dealing with imbalanced data, explaining when it can occur and why it can sometimes be a challenge. More importantly, we will go over when not to worry about data imbalance at all or when it may not be worth worrying about. Furthermore, it introduces the imbalanced-learn library, offering invaluable insights and general guidelines to navigate the intricacies of dealing with imbalanced datasets effectively.
Chapter 2, Oversampling Methods, introduces the concept of oversampling, outlining when to employ it and when not to, and various techniques to augment imbalanced datasets. It guides you through the practical application of these techniques using the imbalanced-learn library and compares their performance across classical ML models. Practical advice on the effectiveness of these techniques in real-world scenarios concludes the chapter.
Chapter 3, Undersampling Methods, presents the concept of undersampling as an effective approach for data balancing when standard oversampling isn’t an option. This chapter covers strategies to effectively remove examples from imbalanced data, different ways of addressing noisy observations, and procedures for handling easily categorized instances. We will also discuss when to avoid undersampling of the majority class.
Chapter 4, Ensemble Methods, explores the application of ensemble techniques, including bagging and boosting, to enhance the performance of ML models. Moreover, it tackles the challenge of imbalanced datasets, where traditional ensemble methods may be ineffective, by combining the ensemble methods with the techniques introduced in previous chapters.
Chapter 5, Cost-Sensitive Learning, explores some alternatives to sampling techniques, including oversampling and undersampling. This chapter highlights the significance of cost-sensitive learning as an effective strategy to overcome the problem of imbalanced datasets. We also discuss threshold-tuning techniques, which can be very relevant in the context of data imbalance.
Chapter 6, Data Imbalance in Deep Learning, presents the core concepts of deep learning and walks through the issues posed by imbalanced datasets. You will investigate typical types of imbalanced data challenges in various deep learning applications and develop an understanding of their impact.
Chapter 7, Data-Level Deep Learning Methods, marks a transition from classical ML to deep learning, exploring the adaptation of familiar data-level sampling techniques and unveiling opportunities for enhancing these methods in the context of deep learning models. It dives into combining deep learning with oversampling and undersampling techniques, covering dynamic sampling and data augmentation for images and text. It emphasizes the fundamental differences between deep learning and classical ML, particularly the nature of the data they handle, whereas deep learning deals with unstructured data such as images, text, audio, and video. The chapter also explores techniques to address class imbalance in computer vision and their applicability to Natural Language Processing (NLP) problems.
Chapter 8, Algorithm-Level Deep Learning Techniques, expands on the concepts from Chapter 5, Cost-Sensitive Learning, and applies them to deep learning models. We adapt deep learning models through loss function modifications using the PyTorch deep learning framework, ultimately enhancing model performance and enabling more effective predictions.
Chapter 9, Hybrid Deep Learning Methods, explores innovative techniques that bridge the gap between data-level and algorithm-level methods from the previous two chapters. This chapter introduces the concept of graph ML and employs a real-world Facebook social network dataset to provide valuable insights and practical applications for addressing data imbalance in deep learning. We will also introduce the concept of hard mining loss and build upon it to explore a specialized technique called minority class incremental rectification, which combines hard mining with cross-entropy loss.
Chapter 10, Model Calibration, takes a different angle of addressing data imbalance. Rather than focusing on data preprocessing or model building, this chapter highlights the post-processing of prediction scores obtained from trained models. Such post-processing can be valuable for both real-time predictions and offline model evaluation. The chapter offers insights into measuring the calibration of a model and explains why this aspect can be indispensable when dealing with imbalanced data. This is particularly important since data balancing techniques can often lead to model miscalibration.
Appendix, Machine Learning Pipeline in Production, offers a foundational guide to constructing ML pipelines in production environments that encounter imbalanced data. This appendix provides a brief roadmap, going over the sequence and stage at which techniques for addressing data imbalance should be integrated.
