Data Representations and Processing in Machine Learning

This lecture focuses on the importance of data representations and processing techniques in machine learning. It begins with a recap of polynomial feature expansion and kernel functions, emphasizing their role in transforming linear algorithms into nonlinear ones. The instructor discusses the concept of kernels, which serve as a similarity measure between data points, and introduces the Representer theorem, which allows for kernelizing various algorithms. The lecture then shifts to data representations, highlighting the need for effective representations to handle heterogeneous data, such as text and images. The Bag of Words model is introduced for text data, illustrating how to create a common representation for varying lengths of text samples. The concept of visual words is also discussed for image data, where image patches are treated similarly to words. Finally, the lecture addresses the issue of class imbalance in datasets and presents strategies for sampling methods and cost-sensitive approaches to improve model performance.