Development of an Improvised Technique for Classification in Data Mining

Abstract

Data mining is the process of finding useful hidden patterns from huge amount of data that are produced by information systems and using those patterns in smart decision making. The predefined methods and algorithms that are used to extract these useful patterns are called data mining techniques.Among these, classification is a supervised learning approach. Classification is a technique in which a given unknown instance is categorized into a particular class or label. An algorithm that implements classification, known as a classifier. A classifier analyses the training instances with labels and builds a model based on them to predict labels of test instances that are unknown. Classification algorithms are based on many learning methods such as instance-based, rule-based, tree-based etc.The k nearest neighbor (k NN) algorithm is one of the most widely used classification methods, which is a type of instance-based, non-parametric learning method or lazy learning method. The basis of k nearest neighbor classifiers is learning by finding similarities between instances. The similarities are measured by a distance formula.The motivation behind selecting k NN classification algorithm for this research work is that it has widely been used in applications of data mining and machine learning because of its inherent simplicity in implementation and note-worthy performance.In this research work, several conceptual variations of k NN have been proposed which lead to remarkable improvements in classification performance in various ways. In the first phase of this research, in order to enhance classification performance, an ensemble model has been constructed using stacking approach.Further enhancements in later phases led to certain limitations being eradicated. Rare class problems are one of the limitations of k NN as it is an instance-based non parametric learning method, which means there is no prior model assumption so it is much more sensitive to imbalanced data.