Murty M N - Representation in Machine Learning

• A timely report of state-of-the art analytical techniques

• A bridge between new research results, as published in journal articles, and a contextual literature review

• A snapshot of a hot or emerging topic

• An in-depth case study or clinical example

• A presentation of core concepts that students must understand in order to make independent contributions

This book deals with the most important issue of representation in machine learning (ML). While learning class/cluster abstractions from the data using a machine, it is important to represent the data in a form suitable for effective and efficient machine learning. In this book, we propose to cover a wide variety of representation techniques that are important in both theory and practice.

In practical applications of current interest, the data typically is high dimensional. These applications include image classification, information retrieval, problem solving in AI, biological and chemical structure analysis, and social network analysis. A major problem with such high-dimensional data analysis is that most of the popular tools like the k-nearest neighbor classifier, decision tree classifier, and several clustering algorithms that depend on interpattern distance computations fail to work well. So, representing the data in a lower-dimensional space is inevitable.

Some of the popular linear feature extractors are based on principal components, random projections, and nonnegative matrix factorization. We cover all these techniques in the book. There are some misconceptions in the literature on representing the data using a subset of principal components. It is typically believed that the first few principal components make the right choice for classifying the data. We argue and show practically, in the book, how such a practice may not be correct.

It is argued in the literature that deep learning tools are the ideal choices for nonlinear feature selection; also they can learn the representations automatically. These tools include autoencoders and convolutional neural networks. We discuss these tools in the book. Further, we argue that it is difficult even for the deep learners to automatically learn the representations.

We present experimental results on some benchmark data sets to illustrate various ideas.

The coverage is meant for both students and teachers and helps practitioners in implementing ML algorithms. It is intended for senior undergraduate and graduate students and researchers working in machine learning, data mining, and pattern recognition. We present material in this book so that it is accessible to a wide variety of readers with some basic exposure to undergraduate-level mathematics. The presentation is intentionally made simpler to make the reader feel comfortable.