A Revisit to Classification Algorithms

Author Name(s): Chandaka Bhavani Sai Sivani, Dantuluri Thanusha, Akella Surya Rohit, Chukka Kundana Siri
Author Email: bsaisivani@gmail.com

Abstract

Artificial Intelligence is the term often heard in the field of technology over the past few years. The vision of this paper is to provide a keen understanding about three Classification techniques for Machine Learning. They are Decision trees, Naive Bayes algorithm and Support Vector Machines. The foundation for this paper is based on several researches which were done previously on these classification techniques. The paper is presented in such a way that it illustrates the techniques used by the researchers in their papers briefly. It also re-views the previous research done on each of the classification techniques and shows a clear view of the optimal techniques. Every research was analyzed thoroughly and has been shown in contrast with respect to other researches. These researches are an extended work to the already existing techniques with newly proposed methodologies. The review says how each technique is better than the other. To have a brief understanding of comparisons each evaluation measures were depicted in tabular format.

Introduction

Artificial Intelligence is a field in computer science that involves the study of intelligent agents. Artificial Intelligence is the stimulation of human like intelligence and behavior by the machines .The programs can mimic human intelligence like reasoning, decision making, perception, visualization, voice recognition, social intelligence etc. With its vast range of applications AI began to maneuver the researchers. As an intelligent machine it must have the ability to learn from its past its experience. The program must train itself. In decision trees the paper reviews the studies in the construction of accurate and incremental decision tree learning algorithm. Very fast decision tree (VFDT) algorithm has been proposed by the researchers along with the perks of extremely fast decision algorithm (EFDT). The paper compares their methodologies against C4.5 standard with UCI benchmark datasets, C4.5 is a decision tree generating algorithm widely. VFDTcNB has been introduced which have the ability to apply naïve Bayes classifiers in tree leaves to classify test examples without any overhead in the training phase . Also HAAT is proposed which imparts anytime property to the Hoeffding tree (VFDT), it is an incremental decision tree learner for large data sets. Another approach for classification is by using naïve Bayes classification. It is usually employed for text classification. One of the papers deal with the clarification for the confusion encountered in the two different models in the text classification, the multi-variate Bernoulli model and multinomial model, both are probabilistic models. The multi-variate Bernoulli model does not capture the number of times the words have occurred, here probability of all the attribute values are multiplied for the calculation of probability. In multinomial model the number of occurrences of each word is captured, here the probability of the words occurred is multiplied for the calculation of probability. The research clarifies this doubt by comparing the result obtained when these two models are implemented on five datasets using “naïve Bayes assumption”. The assumption is that the probability of each word occurring in a document is independent of the occurrence of other words in a document.

Conclusion

The conclusion of is that VFDTcNB and VFDTcMC are same and VFDTcNB generates a stable predictive decision tree. CVFDT and HATT were built based on the split revaluation, but their objectives and methods were different. As NT inherits the superiority of C4.5, it performs better than Naïve Bayes and C4.5 in accuracy, running time and tree size. Decision trees don’t work well with continuous target variable, so Naïve Bayes classifier was introduced. Multinomial Naïve Bayes model performs much efficient than multi-variate Naïve Bayes model. But MNB produces optimistic bias results with SMO. So, SVM is preferable for better accuracy. SVM is good as compared to other classifiers as the computational complexity is reduced and classification efficiency is increased when compared to any other nonlinear classifier. SVM identifies the malware sent through various IP addresses through the Random Sampling with repetition of training technique. In order to improve the efficiency and to overcome the limitations of SVM, CART technique was introduced. Through this SVM and other few processing techniques, the imbalanced classed data-sets can be converted to balance data-set and can be evaluated to increase the accuracy and precisions.

 

1,401 total views, no views today

Download PDF File

About the author: admin