Research & Projects

Research:

1. Title: An Effective Multilayered Ensemble Classifier (ongoing)

Abstract:

Often when a single learner fails to produce accurate results in prediction of class labels in supervised learning tasks, an ensemble of classifiers is applied. These ensemble classifiers have been quite successful in recent years to solve complex machine learning problems in different domains. In this paper, we present a novel multi-layered ensemble classifier for solving binary classification tasks. Our layered ensemble technique performs the classification in multiple stages or layers. In each layer a single classifier is applied that divides the instances intro two sets according to the predicted labels. In the later stages or layers these subsets undergo further learning by another set of single classifiers and this process goes on until the classification performance is satisfactory or the subsets produced in these way are too small. The over-fitting issues are dealt by the termination criteria. On a selected benchmark data-sets used to test our method, it performs significantly better or similar to other state-of-the-art ensemble classifiers.

2. Title: Big Data with Decision Tree Induction (submitted on an Int. Con.)

Supervised by Dr. Dewan Md. Farid

Abstract:

Big data mining is one of the major challenging research issues in the field of machine learning for data mining applications in this present digital era. Big data consists of 3V’s: (1) volume - massive amount of data/ too many bytes, (2) velocity - high speed streaming data/ too high a rate, and (3) variety - data are coming from different sources/ too many sources. Collecting and managing real-life big data is a difficult task, as big data is so big that we cannot keep all the data together in a single machine. Therefore, we need advanced relational database management systems with parallel computing to deal with big data. Knowledge mining from big data employing traditional machine learning and data mining techniques is a big issue and attract computational intelligent researcher in this area. In this paper, we have used the decision tree (DT) induction method for mining big data. Decision tree induction is one of the most preferable and well-known supervised learning technique, which is a top-down recursive divide and conquer algorithm and require little prior knowledge for constructing a classifier. The traditional DT algorithms like Iterative Dichotomiser 3 (ID3), C4.5 (a successor of ID3 algorithm), Classification and Regression Trees (CART) are generally built for mining relatively small datasets. So, we need a more scalable decision tree learning approach for mining big data. In this paper, we have engendered several trees employing two scalable decision tree algorithms: RainForest Tree and Bootstrapped Optimistic Algorithm for Tree construction (BOAT) using seven benchmark datasets from Keel Repository and UCI Machine Learning repository. We have compared the performance of RainForest and BOAT algorithms. Also, we have proposed a decision tree merging approach, as decision tree merging is a very complex and challenging task .