Improving Classifier Generalization

ISBN-13: 9789811950759 / Angielski / Miękka / 2023

Rahul Kumar Sevakula; Nishchal K. Verma

Opis

This book elaborately discusses techniques commonly used to improve generalization performance in classification approaches. The contents highlight methods to improve classification performance in numerous case studies: ranging from datasets of UCI repository to predictive maintenance problems and cancer classification problems. The book specifically provides a detailed tutorial on how to approach time-series classification problems and discusses two real time case studies on condition monitoring. In addition to describing the various aspects a data scientist must consider before finalizing their approach to a classification problem and reviewing the state of the art for improving classification generalization performance, it also discusses in detail the authors own contributions to the field, including MVPC - a classifier with very low VC dimension, a graphical indices based framework for reliable predictive maintenance and a novel general-purpose membership functions for Fuzzy Support Vector Machine which provides state of the art performance with noisy datasets, and a novel scheme to introduce deep learning in Fuzzy Rule based classifiers (FRCs). This volume will serve as a useful reference for researchers and students working on machine learning, health monitoring, predictive maintenance, time-series analysis, gene-expression data classification. 

Spis treści

Chapter 1. Introduction to classification algorithms

a. Basics

b. Bias-variance tradeoff

c. Generalization and sampling error

References

Chapter 2. Methods to improve generalization performance

a. Statistical Learning Theory

i. Vapnik-Chervonenkis Dimension

ii. Growth function

b. Maximum margin classifiers

c. Having fewer parameters - Occam's razor

d. Regularization

e. Boosting

i. Gradient Boosting

f. Transfer learning

g. Unsupervised greedy layerwise learning of deep networks

h. Dropout

i. Conclusions

References

Chapter 3. MVPC – a classifier with very low VC dimension

a. Majority Vote Point classifier

b. Implementation of MVPC

c. Evaluating and comparing VC dimension

i. Upper bound on VC dimension

ii. Empirical estimation of VC dimension with search space reduction

iii. Comparing with linear classifiers

d. Case study on Time-series classification

e. Case study on Gene-expression data classification

f. Conclusions

References

Chapter 4. Framework for reliable fault detection with sensor data

a. Data acquisition framework to simulate real time environment

b. Data Pre-processing

i. Normalization robust to outliers

c. Feature Extraction

d. Feature Selection algorithm using Graphical Indices

i. Feature Ranking and Graphical Indices

ii. Dataset Rejection

iii. Dataset Retrieval

iv. Feature Selection Architecture

e. Classification

f. Sensitive Position Analysis (SPA)

i. Parameter range identification (PRI)

g. Case Study on Air Compressor Fault Detection

i. Leakage Inlet Valve (LIV) fault detection

ii. Leakage Outlet Valve (LOV) fault detection

iii. Online Testing

iv. Real time testing

h. Tutorial for time-series classification

i. Conclusions

References

Chapter 5. Membership functions for Fuzzy Support Vector Machine in noisy environment

a. Fuzzy Support Vector Machine (FSVM)

i. Available Membership Functions for FSVM

ii. Limitations of earlier Membership Functions

iii. Convex Hulls analysis on FSVM

b. Set Measures - Distance Measure between Points and Non-Empty Sets

i. Distance between a Point and a Non Empty Set

ii. Hausdorf Distance (HD)

c. Proposed General Purpose Membership Functions

i. GPMFs with density based clustering

ii. Proposed GPMFs with Fuzzy C-Means clustering

d. Case Study on datasets from UCI repository

i. Class Imbalance Learning in FSVM

ii. Performance Evaluation

iii. Experimentation

e. Results and Analysis

i. Statistical Analysis

f. Conclusions

References

Chapter 6. Stacked Denoising Sparse Autoencoder based Fuzzy rule classifiers

a. Stacked Autoencoder (SAE) based FRC

i. Autoencoders and denoising autoencoders

ii. Weight initialization with unsupervised greedy layerwise learning

iii. Naive stacked autoencoder based FRC

b. Data pre-processing Strategies for SDSAE based FRC

i. Pre-processing Strategy I (PP-I)

ii. Gaussian Mixture Model (GMM) for Pre-Processing

iii. Pre-processing Strategy II

iv. Pre-processing Strategy III (PP-III)

v. Preprocessing of Nominal Features

c. Fine Tuning of weights for FRC Modeling

i. Process Definitions

ii. Fine Tuning Strategy I (FT-I)

iii. Fine Tuning Strategy II (FT-II)

iv. Fine Tuning Strategy III (FT-III)

d. Integration with Expert Knowledge

e. Case study on datasets from UCI repository

i. Dataset wise Observations

f. Conclusions

References

Chapter 7. Epilogue

a. Transfer learning for Molecular Cancer Classification

b. Transfer learning for time-series classification

c. Directions for future work

d. Conclusions

References

O autorze

Dr Sevakula Rahul Kumar has over 10 years of research experience in machine learning (ML) and deep learning (DL). He received his Bachelor’s degree from the National Institute of Technology (NIT) Warangal, India in 2009 and later his Ph.D. degree from the Indian Institute of Technology (IIT) Kanpur, India in 2017. He is currently a Sr. Research Scientist at Whoop, and his research interests lie at the intersection of ML, physiological signals, cardiovascular health monitoring (medicine) and wearables. Prior to joining Whoop, he was an Instructor (junior research faculty) at Harvard Medical School and Massachusetts General Hospital, USA, and a Data Scientist at IBM India. He has filed multiple patent disclosures and has published over 45 research papers in international peer-reviewed journals and conferences. He is also a reviewer for several journals of national and international repute.

Dr. Nishchal K. Verma is a Professor in the Department of Electrical Engineering at Indian Institute of Technology (IIT) Kanpur, India.  Dr. Verma's research interest falls in Artificial Intelligence (AI) related theories and its practical applications to inter-disciplinary domains like machine learning, deep learning, computer vision, prognosis and health management, bioinformatics, cyber-physical systems, complex and highly non-linear systems modeling, clustering, and classifications, etc. He has published more than 250 research papers in peer-reviewed reputed conferences and journals along with 4 books (edited/ co-authored) in the field of AI. He has 20+ years of experience in the field of AI. He is currently serving as Associate Editor/ Editorial Board Member of various reputed journals and conferences. He has also developed several AI-related key technologies for The BOEING Company, USA.