Time-series based Clustering TClustern-SMOTE

Abstract

Handling Imbalance data is one of the most important topic in machine learning. Imbalance data are skewed data towards one majority class rather than other classes because of this trained model by these kind of data sets are biased towards the majority class, which is not considered as effective model and affects the final decision. There are many method to tackle this problem like SMOTE, Borderline-Smote, ASN-Smote, which generate the artificial samples to balance the data. Several algorithms have been developed for this purpose, however most are sophisticated but tend to create excessive noise. In this work we present oversampling method using cluster and SMOTE technique, in which recent samples are used to generate new samples. The suggested technique performs better when training data are oversampled, according to comprehensive trials with 5 datasets. Python programming language implementation is made accessible.