Thyroid Prediction on Imbalanced Datasets Using Machine Learning Classifiers

Abstract

Classification is technique to classify our data into a desired (distinct) classis where we can allot an exclusive label to each class. Applications of classification include biometric identification, speech recognition, handwrit ing recognition, nondiabetic detection, document classification etc. Classifiers can be a Binary classifier where we have only 2 distinct classes or 2 possible outcomes (example: Fraudulent Transaction and Non-Fraudulent Transaction) or Multi-Class classifiers: where there are more than two distinct classes (For Example: Predicting Thyroid Disease, possible class labels are Hyperthyroidism, Hypothyroidism and Normal). Evaluating Machine learning classifier models can be a tough task. In this study we will attempt to explain Performance Metrics for Classification problems in Machine Learning. We will also try to explain how inferences drawn from confusion matrix can be misleading at times and also, we will see what other metrics we can turn to for better evaluation of the Machine learning classifier models. We will mostly focus on Binary classifiers and later would mention a few changes with regards to the Multi-class classifiers.