Keywords

5.1 Introduction and Preliminaries

During the past several years, investigation has been concentrated on diverse groups based on the machine learning algorithms due to the extreme require of accurate prophecies. Machine learning is not about giving tight rules by analyzing the datasets rather it is used to predict the likelihood of future events with some certainty. Classification is a machine learning approach used to fore tell cluster association for documents illustration and is a widely used technique in various fields [1].

Machine learning, pervasive computing, statistical analysis, data analytics, etc., are the applications of artificial intelligence (AI), whereas machine learning allows training and strengthens from practice to estimate the eventualities [2,3,4,5, 5,6,7,8].

A well-known test sample label is correlated with a separate result from the model. The extent of precision of the proportion of instances of the test set is grouped consequently by the framework. If precision is tolerable, then this model is used to separate tuples of data class labels, which are unknown [9, 10].

5.2 Literature Work and Methodologies

Classification has been considered as a seminal issue in the area of machine learning [11]. All the time, there has been absolutely a number of enormous surveys on classification algorithms [12, 13], performance evaluation [14,15,16], collations, and assessment of various classification algorithms [2, 17] beside their uses in figuring out real-life problems in the applications of business [9, 18,19,20,21], engineering, medicine [1, 22, 23], etc.

Amudha and Abdul Rauf [24, 25] applied data mining techniques as an approach for intrusion detection to identify whether the deviation from normal usage patterns can be flagged as intrusions and performed a correlative investigation of various classification algorithms.

Voznika and Viana [26, 27], described different approximation algorithms such as statistical algorithms, genetic programming, neural networks and concluded that the best model can be found by trial and error trying different algorithms in order to obtain the best results possible.

Kesavaraj, Sukumaran [13] performed investigation on multifold categorization methods to furnish an exhaustive analysis of machine learning algorithms.

Chintan Shah and Anjali Jeevani [17] compared decision tree, K-nearest neighbor, Naive Bayesian using parameters like correctly classified illustrations, time taken, relative absolute error, kappa statistic, and root relative absolute error on breast cancer dataset.

Dogan and Tanrikulu [18], performed a study that collate and contrast the precision of the classification algorithms. The application of certain classification models on multiple datasets is done in three stages. The research addressed the reliability of the classifiers, studied by demonstration on various datasets.

Rutvija and Pandya [28] performed the extensive analysis on various categorization techniques.

Keerthana [29] focused on image classification approach in order to identify better algorithm for medical image classification.

Classification is an approach of grouping or allocating class labels to a pattern set under the direction of an instructor. Classification is also termed as supervised learning. The patterns are primarily segregated into training and test sets. Training set is used to prepare the classifier, and the test set is prone to estimate the precision of a classifier. The classifiers are categorized into tree-based, rule-based, Bayes, functions, etc. The algorithms that have been chosen for this predictive data mining task include J48 from trees, multilayer perceptron from functions, ZeroR from rules, and Naive Bayes from Bayes.

The most popular supervised classifier which can work well on noisy data is decision tree classifier. There are various other types of classifiers such as Bayesian classification, neural network-based classifier, and support vector machine. A great deal of research has been done for developing efficient methods in the field of machine learning.

The J48 algorithm uses information gain and gain ratios to construct the decision tree for a given dataset. It works by recursively dividing the data on a single attribute, according to the information gain calculated. Each split in the tree represents a node where a decision must be taken, and you go to the following node and the next till you reach the leaf that expresses you the predicted output.

The steps in the J48 algorithm are as follows:

  1. (i)

    If the requirements are the identical group, the tree illustrates the leaf so that the leaf is substituted by designating in the identical class.

  2. (ii)

    The feasible information is intended for every characteristic, determined by a check on the attribute. Then, the gain in information is premeditated that would outcome from a examination on the characteristic (attribute).

  3. (iii)

    Then the best characteristic (attribute) is identified on the foundation of the current selection criterion and that attribute is adopted for ramification.

5.3 Information Gain and Gain Ratio

The information gain is based on the entropy after a dataset is split on an attribute, where entropy is used to estimate the similarity of a sample. If the instance is completely identical, the entropy is zero, and if the instance is evenly separated, it has entropy of one.

The entropy is calculated using the following formula.

$${\text{entropy}}\left( {p1,p2, \ldots ,pn} \right) = - p1\log p1 - p2\log p2 - \cdots - pn\log pn$$
(5.1)
$${\text{entropy}}\left( {p1,p2, \ldots ,pn} \right) = - \sum {pi\log pi}$$
(5.2)

Entropy on the other hand is an estimate of impurity. It is characterized for a binary class with estimates a and b as:

$${\text{entropy}} = - \, p\left( a \right)*\log \left( {P\left( a \right)} \right){-}p\left( b \right)*\log \left( {p\left( b \right)} \right)$$
(5.3)

Using the above formula, we calculate two entropy values, namely entropy before and entropy after. The entropy before value is calculated before splitting, and entropy after is computed after considering the split. Now by assimilating the entropy before and after the split, we derive an estimate of information gain as denoted below:

$${\text{Information gain}} = {\text{entropy before}}{-}{\text{entropy after}}$$
(5.4)

At each node of the tree, this computation is carried out for every feature, and the feature with the largest information gain is chosen for the split in a greedy manner. This process is applied recursively from the root-node down and stops when a leaf node contains instances all having the same class, i.e., it stops when the node cannot be divided further. Constructing a decision tree is all about finding attribute that has the highest information gain.

Gain ratio is a modification of the information gain that reduces its bias. It takes into account the number and size of branches while choosing an attribute. There are chances of getting negative values in the existing information gain and gain ratio algorithms.

The idea of the proposed algorithm is to eliminate the negative values. The accuracy of the algorithm can be improved by eliminating the negative values. The proposed algorithm checks if the entropy before value is less than entropy after value and return 0; otherwise, it returns the unknown rate calculated.

Because of the outliers pruning is a significant step to the result. Some instances are present in all datasets which are not well defined and differ from the other instances on its neighborhood.

The classification is performed on the instances of the training set, and tree is formed. There exist various algorithms for performing classification, extracting salient features, opinion mining, processing of scalable web log data using map reduce framework etc. [30,31,32,33,34,35,36,37]. The pruning is performed for decreasing classification errors which are being produced by specialization in the training set. Pruning is performed for the generalization of the tree.

5.4 Results and Discussion

Evaluation of the datasets has been done by using the proposed classification algorithm. Various classification algorithms analyzed are evaluated by the evaluation criteria such as true positive rate, false positive rate, mean absolute error, and root mean square error. Heart disease, mushrooms, and birds are the datasets used for the analysis. These two datasets are taken from UCI machine learning repository. The classification algorithms are applied on the data using tenfold cross validation technique, and the results are then recorded. A sample description of datasets has been represented in Table 5.1.

Table 5.1 Sample description of datasets
Full size table

The considered sample heart disease dataset has 14 attributes and 303 instances that are categorized into 5 classes. Mushrooms dataset has 23 attributes and 8124 instances that are categorized into 2 classes. Experimentation has been done on each dataset.

5.4.1 Data Set 1 (Heart Disease Dataset)

The considered sample heart disease dataset has 14 attributes and 303 instances that have only 5 classes. This dataset has taken from UCI machine learning repository.

True positive rate, false positive rate, root mean square error, and mean absolute error are calculated for J48, Naive Bayesian, multilayer perceptron, ZeroR, and the proposed algorithm, which are represented in Table 5.2.

Table 5.2 Results of the classification algorithms on heart disease dataset
Full size table

True positive rate of J48, Naive Bayesian, multilayer perceptron, ZeroR, and the proposed algorithm are 0.558, 0.559, 0.574, 0.541, and 0.558, respectively. It shows that the proposed modified algorithm is able to show the same performance as J48 algorithm.

False positive rate of J48, Naive Bayesian, Multilayer Perceptron, ZeroR, and the proposed algorithm are 0.238, 0.273, 0.189, 0.541, and 0.238, respectively. It shows that the proposed modified algorithm is able to show the same performance as J48 algorithm.

Mean absolute error of J48, Naive Bayesian, multilayer perceptron, ZeroR, and the proposed algorithm are 0.2, 0.1838, 0.2768, 0.2591, and 0.2, respectively. It shows that J48 and the proposed modified algorithm are better in terms of mean absolute error, compared with ZeroR and multilayer perceptron. But for this dataset Naive Bayesian performs better by exhibiting low mean absolute error, i.e., 0.1838.

Root mean squared error of J48, Naive Bayesian, multilayer perceptron, ZeroR, and the proposed algorithm are 0.3867, 0.3368, 0.378, 0.3592, and 0.3367, respectively. It shows that the proposed modified algorithm is better in terms of root mean squared error, compared with the other algorithms, i.e., 0.3367.

For the considered dataset, the proposed modified algorithm shows better performance than other algorithms in terms of the root mean square error. The root mean square error values of various algorithms are pictorially represented in Fig. 5.1 on heart disease dataset.

Fig. 5.1
A bar graph compares the root mean squared error versus different algorithms. J 48 and the modified algorithm have maximum values of 0.3867, while N B has a minimum value of 0.3368.

Comparison of root mean squared error of classifiers on heart disease dataset

Full size image

5.4.2 Data Set 2 (Birds Dataset)

Birds dataset used here is an images dataset. It contains 600 images of 34 samples of 6 types of birds. We applied some filters before classifying the data. This dataset has taken from a Ponce research group repository. True positive rate, false positive rate, root mean square error, and mean absolute error are calculated for J48, Naive Bayesian, multilayer perceptron, ZeroR, and the proposed algorithm, which are represented in Table 5.3.

Table 5.3 Results of the classification algorithms on a birds dataset
Full size table

True positive rate of J48, Naive Bayesian, multilayer perceptron, ZeroR, and the proposed algorithm are 0.372, 0.325, 0.390, 0.165, and 0.372, respectively. It shows that the proposed modified algorithm is able to show the same performance as J48 algorithm.

False positive rate of J48, Naive Bayesian, multilayer perceptron, ZeroR, and the proposed algorithm are 0.126, 0.095, 0.102, 0.169, and 0.026, respectively. It shows that the proposed modified algorithm is able to show better performance compared with all the considered algorithms.

Mean absolute error of J48, Naive Bayesian, multilayer perceptron, ZeroR, and the proposed algorithm are 0.213, 0.2726, 0.2774, 0.2778, and 0.213, respectively. It shows that J48 and the proposed modified algorithm is better for the considered dataset in terms of mean absolute error, compared with Naive Bayesian, ZeroR, and multilayer perceptron.

Root mean squared error of J48, Naive Bayesian, multilayer perceptron, ZeroR, and the proposed algorithm are 0.2536, 0.3421, 0.3784, 0.3727, and 0.2436, respectively. It shows that the proposed modified algorithm is better in terms of root mean squared error, compared with the other algorithms, i.e., 0.2436.

For the considered dataset, the proposed modified algorithm shows better performance than other algorithms in terms of the root mean square error. The root mean square error values of various algorithms are pictorially represented in Fig. 5.2 on birds dataset.

Fig. 5.2
A bar graph compares the root mean squared error of different algorithms. M L P and the modified algorithm have maximum and minimum values of 0.3784 and 0.2436, respectively.

Comparison of root mean squared error of classifiers on a bird’s dataset

Full size image

The experimentation has shown that the proposed algorithm outperforms other existing algorithms on various considered datasets.

5.5 Conclusion

This study focuses on finding the right algorithm for classification of diverse datasets. The datasets that used are heart disease and mushrooms. For mushrooms dataset, the decision tree algorithm J48 and the proposed modified algorithm gave better results in terms of root mean squared error (RMSE) and for heart diseases dataset; the proposed modified algorithm gave reduced error rates than the traditional J48 algorithm, Naïve Bayes, multilayer perceptron, and ZeroR. However, it is noticed that the performance of a classifier depends on the dataset used. Although there are many algorithms available, the best one is often found by trial and error. For better results, one must compare or even combine the available algorithms. The performance of the existing algorithms can even be improved with minor modifications.