Models of Classification in Educational Data Mining and

Preparing to load PDF file. please wait...

0 of 0
100%
Models of Classification in Educational Data Mining and

Transcript Of Models of Classification in Educational Data Mining and

International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 20 (2018) pp. 14717-14727 © Research India Publications. http://www.ripublication.com
Model of Tuned J48 Classification and Analysis of Performance Prediction in Educational Data Mining
Anoopkumar M 1*, A. M. J. Md. Zubair Rahman 2 1 Research Scholar, Research and Development Centre, Bharathiar University, Coimbatore – 46, Tamilnadu, India.
*Corresponding Author 2 Principal, Al-Ameen Engineering College, Erode, Tamilnadu, India.

Abstract
Information-Technology (IT) and Data Engineering in Education nowadays decides the quality and pace of learning achieved by. Use of Information Engineering and Machine Intelligence generate precise advantages in Optimisation problems of performances and prediction. Many Data Mining algorithms have so far proved to be serving their best in many cases. A tuned J48 is all-around considered in this paper along with the base performance evaluation and analysis of Naïve Bayes, Bayes Net, Multilayer Perceptron, SVM, REPTree and Random Forest with Two Datasets given. Researches in EDM reveal that classification outperforms other mining methodologies. Hence, a comprehensive analysis of Models of Classification is also proposed in this research. Besides these, in this paper, Weka, is used for the base performance calculations. The prerogative of this research reveals that though many models are very competent, few are not; the performance of tuned J48 invites the attention, as its result provide realistic best, compared to others. It is obvious that combinational use of models or an improved and optimised method of J48 classification would be a great choice to enhance the classification accuracy and prediction of Students.
Keywords: Classification, J48 Decision Tree Model, Precision, Accuracy, Prediction Models, MLP, Educational Data Mining
INTRODUCTION
Just as environmental changes witnessed every year, every business and service sector change now and then. Also the Educational Sector. Considering the traditional and modern practices of the education, it is evident that an enormous change has happened to the teaching-learning process, nature of students coming into education, expected outcomes of education, urgency of education, lifestyle of teacher and student, industrial relevance of education and many more. The hidden information unable to identify are now considered as the vital driving forces of this specific sector.
India with 72.1% adult literacy and 95.2% youth literacy between age 15 to 24 stands at fifth according to UNESCO [40]. Even though India is at this position there are a lot more to can be done with the educational sector as being one of the country of leading intelligence worldwide. From the era of ancient civilisation India had laid remarkable contributions to the world education. And now, consider one of the highest literacy state Kerala. Here the prime importance is always

given to primary education and now the trend has changed to further education including professional educations. As is known, very recently, due to various reasons of industry and global phenomena the professional education is going through enhancement stage. Wherein, indeed a more quality ascertainment are getting established with intelligent feedbacking and monitoring and controlling systems. Despite the intended purpose of conventional use, organisations started exploring organisational information for a better operation and opportunities. Storage, efficient retrieval and utilisation of educational information using data mining, data science, and machine learning have become inevitable for this sector to decide effectiveness of expected outcome. Taking these opportunities and concerns into account, in this work, experiments and research works in Educational Data Mining (EDM) [1]and its predictive tasks are formally discussed and analysed and results are again compared to find most distinguish methods of choice. It also considers, the existing methods of modelling and mining based on the contextual requirements of what to be produced. Based on a similar survey and suggestions in [1] , this paper mainly concentrate on various classification techniques of EDM [such as Naïve Bayes, Bayes Net, J48, Multilayer Perceptron [MLP], SMO [SVM], REPTree, Random Forest] and methodologies used by researchers to optimise the work. At the final stages of this work, a real dataset KTU_SNG is used to check actual conformance of the statements with the new dataset, and results using Data mining and Analysis tool called Waikato Environment for Knowledge Analysis (Weka) developed at the University’ of Waikato, New Zealand, are interpreted.
The entire work is organised in different sections and this is indeed to add a clear view of this paper. A hopeful introduction towards the work has been written in Section-I, as preceding, to other sections. The organization of rest this paper is as follows; Section-II handles a context and the problem statements. Next to problem statement in Section-III consists of a narration of this research and brief discussion about the problem, also present, a work proposed. Moving on to next, in Section-IV the related research literatures are taken into discussion proceeded by the implementation of it in Section-V. Section-VI of this work explains the experimental results and observations on results. A relative comparison up to possible extent is also given. Pre-final stage, the SectionVII covers the entire work with a conclusion and future guidelines extending this research. And finally, for the reader’s assistance and deep book of facts complete list of reference is also supplemented in the next.

14717

International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 20 (2018) pp. 14717-14727 © Research India Publications. http://www.ripublication.com

PROBLEM STATEMENT
Education is a progressive and a continuing process. Quality education and easy management decisions are a look forward to by every organisation. And hence, now the quality control and assurance in these sectors are rising up drastically. One of the best way to implement all these is the usage of Organisational systems equipped with Data Warehousing and Data Mining (DWDM), and Machine Learning (ML) capabilities for a sustainable decision making. The major portion of EDM activities during 2005 to 2015 were concentrated on classification [1] and related methods. The main objective of this paper to investigate and identify suitable classification techniques for Students Academic Performance Datasets. Two datasets are used in this entire work. xAPI-EduData [Kalboard-360 LMS data of University of Jordan, The dataset consists of 480 student records and 16 features] and KTU_SNG-Data [Kerala Technological University__SNG College of Engineering Data, 496 records and 60 features]). We will get familiarised with these data, later in this paper.
Most of the educational institutions, once were pioneers in results are concerned that KTU students and their results in Kerala had a sudden fall in result percentage when compared to the results of similar Courses of Other Universities taught in same colleges. The findings of this work definitely help them by providing how best a method of classification can be used in Multivariate and Qualitative datasets of students.
 To review concepts related to the above statement  To identify the methods of solutions and their success %
for it  To select best methods of choice and compare its
performance on KTU_SNG dataset  Result analysis of selected algorithms on KTU_SNG
dataset and a slight comparison with already stated proven results of similar researches
The modus operandi of this experimental work is defined including the related works with their success ratios and experimenting performances of Naïve Bayes, Bayes Net, J48, Multilayer Perceptron [MLP], SMO [SVM], REPTree, Random Forest algorithms with the two datasets in order to select a best method of choice and finally to present Result analysis of selected algorithms on KTU_SNG (and the xAPIEdu-Data) dataset and a slight comparison with already stated proven results of similar researches.
PROPOSED RESEARCH WORK
Problem Context
As per the surveys and suggestions [1] it is evident that classifications techniques are highly used in the area of performance prediction and optimisation. Experimenting with problems are always produced good result and progress. At least an improvement from the state of ignorance to gained knowledge. This work is an attempt to address the problems in an analytical way and propose a convenient research outcome towards the eager community.

Classification is a data mining method of predicting the value of a categorical variable (normally called as target or class) by building a predictive model based on one or more numerical and/or categorical variables normally called as predictors or attributes. Those are otherwise called an independent variable of a dataset. As given in figure (Fig. 1), the classification techniques can be generally categorised into 1) frequency based, 2) covariance based, 3) similarity function based and 4) other method based classifications. Base classifier ZeroR, OneR, Bayesian (many), Decision Tree (many), Linear Discriminant Analysis (LDA), Logistic Regression, k-Nearest Neighbour, Artificial Neural Network (ANN), Support Vector Machines (SVM) are few among the four distributed categories.
Figure 1: Sample Classification Methods and Categorisation [39]
Here in this experiment, a few of the identified classification algorithms and methods are going to be considered for the research analysis. The methods such as Naïve Bayes, Bayes Net, J48, Multilayer Perceptron [MLP], SMO [SVM], REPTree, Random Forest are given primary importance in the prediction and performance analysis using two datasets (xAPIEdu-Data [Kalboard 360 LMS data of University of Jordan] and KTU_SNG [Kerala Technological University__SNG College of Engineering Data]). The KTU students and their results in Kerala had faced a sudden fall when it is compared to the results of similar Courses of Other Universities taught in same colleges. It is in this regard, an analysis is being planned and executed to populate data. And the result is the KTU_SNG dataset, though it has around 496 records and 60 attributes, for this analysis it is now with only 232 records 45 features are considered, consisting of two differently performed departments. The data can be extended by collaborating with other organisations. According to various researches in this area, the results of findings are varying, however there are methods repeatedly used and produced a better outcome, when using these multivariate datasets. The main agenda for this research is to explore foresaid algorithms and their performances on two similar data sets (Data-1: xAPI-Edu-Data [3, 4] and Data-2: KTU_SNG-Data)

14718

International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 20 (2018) pp. 14717-14727 © Research India Publications. http://www.ripublication.com
of students and to check its nature of behaviour with each other. Both datasets are collected from Educational Management Systems and Learning Management Systems of respective institutions.

General framework of educational system A general framework of Educational Systems and the relation of Data Mining with it is that it combines various category of data of students into a single frame of dataset. And then, various inevitable processes of Data Mining being applied to excavate hidden knowledge about students. A similar one is illustrated in the figure (Fig. 2). Most of the system have different combination of mining and reporting facilities to come across suitable decisions. The point of uncertainty of indirect information, is cleared while using data mining rather than normal querying. Selection of methods is decided, based on the problem it has to address.
Figure 2: Educational System Framework.
Educational management data mining framework When an educational institution like SNG is considered, the practices and requirements are very precise and most of the critical decisions are filtered through a Mining Module attached to the EDM and LMS. The figure (Fig. 3) Educational Management Data Mining Framework (EMDM) illustrates basic granularity of elements a system used in educational mining has to have. The segment LMS DATA of the similar frameworks deployed at institutions are the main sources of datasets. The EDM system serves as the brain of every intelligent decision and reporting is going to be provided. Optimising and maximising its efficiency decides the success and failure of the organisation. Here in this work, a part of the concern are taken into account and rest of the portions will give information on what methods can be more depended, to do better mining in these areas.

Figure 3: EMDM Framework
Proposed framework of classifier evaluation The specific student experimental model is designed for predicting the outcome of students by 9 classifiers including the base classifier ZeroR. The other classifiers of the interest are Naïve Bayes, BayesNet, DT (tuned J48), Multilayer Perceptron (MLP), SMO, SMO (J48), REPTree, and Random Forest. The basic framework of implementation is given below in Figure (Fig. 4). x-API-Edu-Data, KTU_SNG-Data are the two similar datasets used here. Those are really collected from two geographical are and the first one available in UCI repository, later one is a new real dataset of SNG College affiliated to Kerala Technological University, Kerala, India. x-API-Data was a complete data, whereas KTU_SNG-Data have missing values, hence data preparation is being followed required file processing. The framework proposed have a core stage where basic checks were done and the datasets were prepared using 10Fold Cross Validation and Percentage Split before, preprocessing, and selection. Data get transformed to train and test data. Later these data were pre-processed, attributes are selected and given data for classification with various analysis based on a set standard conditions. Base classifications and classifications with probabilistic, frequency table, associated learning, neural network and ensemble models have been considered and evaluated . This will give an efficient analysis on student prediction accuracy and help to compare them individually. The results of Model were taken and compared.
Figure 4: Framework of Classifier Evaluation

14719

International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 20 (2018) pp. 14717-14727 © Research India Publications. http://www.ripublication.com

A specific model with tuned J48, the best performing model to predict and give better accuracy can be found from these analyses. It will help SNG stakeholders to take a better managerial decision on educational service and students can correct themselves using these results.
Measures of evaluation
The performances are evaluated using confusion matrix values are identified as in Table. 1 and the basic formulas of performance measures as follows in the Equations (1) to (4) referring to Table 1.
Table 1. Basic Confusion Matrix

----(5)
Gain, is to calculate the decrements happening to entropy after splits. Entropy of parent and entropy of the child decide its value.
----(6) Gain Ratio, is used to reduce bias operations in multivalued attributes. And the information gain ratio is the ratio of the information gain to the split information:

----(7)

Accuracy= Precision= Recall= F-measure=2

-----(1) -----(2) -----(3) -----(4)

An example confusion matrix:
Table 2. Confusion Matrix Example
=== Confusion Matrix ===
a b <-- classified as
4 5 | a = yes
4 1 | b = no
The number of correctly classified instances a (yes) divided by the number of predicted instances will give the Precision a, is computed as 4/8, while Recall is 4/9, the number of correctly classified instances a (yes) divided by the total number of true instances a. The precision and recall for the other class instances b is the just reverse. Various other measures of calculation for performance evaluation are such as listed below.
Entropy, is normally a split criterion used in ID3, C4.5 (J48) Decision Tree algorithms. The sample is completely homogeneous the entropy is 0 and if the sample is an equally divided it has entropy of 1.

Gini index, is used to measure the impurity (or purity) of CART normally. Target variable t in Binary Classification takes value 0 and 1. If Target Variable t is categorical variable and takes k different values it varies from 0 to k. Gini will be max when target values are equally distributed. For binary classification it is 0.5, for Nominal variable with k level, the maximum value Gini Index is = 1 - 1/k. Gini Index will be minimum and 0 when all observations belong to one label
----(8)
Can also be considered to have an around the flow analysis
RELATED LITERATURES
Data mining and the related technologies are making real changes to Educational sector. Variety of methods are in demand based on the outcome and data it needs are existing. As it is evident in [1] that the methods such as Classification, Clustering, Association and Statistical methods have significant roles in prediction. However the use of classification techniques in EDM is one of the prominent methods, highly used in these kinds of analysis.
Classification is a special method where the nature of dataset plays an important role, which in turn is the basis of methods to be used. Selected Works from 2002 to 2017 have been considered for this familiarisation. The paper [1], in which major role of different mining methods have discussed and it identifies classification as the major contributor in EDM. This paper conducts a comprehensive study on the recent and relevant studies put through in this field to date. The study focuses on methods of analysing educational data to develop models for improving academic performances and improving institutional effectiveness. This paper promote the use of elearning systems in education.
Pre-process and analysis of educational data using various features [2] and performance prediction using ensemble methods have been addressed in the literature [3] by E. A. Amrieh. Method of Pre-processing for performance improvement, has produced a result with an accuracy of 61.3 for DT, 73.8 for ANN and 72.5 for NB. Amrieh in [2], reveals

14720

International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 20 (2018) pp. 14717-14727 © Research India Publications. http://www.ripublication.com

that DT (C4.5/J48) have got only 6% to 7% performance improved using this method. However, in the other work [3] the same collection of methods with the ensemble is used, in which the boosting outperformed in all algorithms with an average of 74% of accuracy. Almost 26% of improvement is drawn in this paper. While testing and validation with x-APIEdu-Data [2] the three algorithms ANN, NB and DT have produced over 80% results. The performances are evaluated using confusion matrix values and the basic formulas of performance measures.
The main focus of the paper [4] is on clustering, outlier detection and univariate analysis in order to increase the performance of predictions. Still few explanation on classification have been given. Special mention of traditional methods of classification such as kNN, J48, Naive Bayes, Naive Bayes Updatable have also given. B. Baradwaj, in this research [5] conducted a study on performance evaluation of students using classification techniques. Algorithms ID3, ASSISTANT and C4.5 have taken into account and various measures of calculation for performance evaluation such as Entropy, Gini index, Gain, Gain Ratio are also considered. Special category of students who need attention can be easily identified of his methods and work presented.
Classification, sequential pattern mining are being considered as the research experiment in this [6] paper. The concepts such as Resampling, SMOTE, Association, and Classification using J48 have been addressed in this work.
Though there are other researches, A. Cufoglu in this contribution [11] concentrated on NB, IB1, SimpleCART, NBTree, ID3, J48 and SMO. The seven classifiers. NBTree found good among the whole and favoured over others in the personalization applications where classification accuracy performance is most important. SimpleCart and J48 classifiers achieved a classification accuracy of 86%.
Improving educational quality [13], and the role of character identification of student using Moodle data and students data from Facebook have been discussed in this [14] experiment. A personality reconciliation helped students to improve their performance. The model tested and gave best 93.23% for Random Forest and 92.19 for J48.
Few similar papers where analysed, there was a presentation and a brief comparison of ID3 and C4.5 [16] and in addition to this CART is also added to have, three most commonly used decision tree algorithms discussed in this paper [31], stimulating their use and scalability on different types of attributes and feature. Few papers discussed about topics such as student retention and attrition [17] to find new methods to recommend personalised learning. J48 is considered as a very fundamental classifier in Weka, and there were studies towards improving J48, either by updating or in collaboration with other complying environments. An implementation in association with at-Lab is discussed in [20] and it has produced almost accuracy up to 99.87 %.
Decision tree, K-NN, Naïve Bayes have been discussed in these [12], [19], [22], [23] and [24] works. Similarly, author with this specific work [29] came up with an outcome of

Naïve Bayes 77.94%, k-Star (K-NN) 69.42% and Decision Tree (J48) 77.26%.
Similar to the work in [23], R. Sumitha have conducted a research on prediction of outcome of students [32] using a dataset contains 300 student data. Among the dataset around 250 are been used as training dataset and 50 datasets as test data to design student model. Classification of the students based on the attributes selected reveals that the prediction rates of algorithms, the range of prediction varies from (80-98%). Among Naive Bayes, Multilayer Perceptron (MLP), SMO, REP tree and J48, where J48 came up with an accuracy of 97%.
The classification accuracy, sensitivity and specificity of J48 and Naïve Bayes have studied in this investigation research [27] and shown that the efficiency and accuracy of j48 with an average of 59% (31%-Yes, 87%-No) is better than that of Naïve Bayes with an average of 49% (9%-Yes, 89%-No).
Real data about 670 high school students from Zacatecas, Mexico is being used in this work and also Interpretable Classification Rule Mining (ICRM) [7] helped to analyse the performance of each with 10 algorithms. JRip, NNge, OneR, Prism Ridor, ADTree J48, RandomTree, REPTree, SimpleCart,ICRM v1, ICRM v2, ICRM v3 are the few list of those algorithms used in this research experiments. To add more prediction of student failure was the main focus.
Association and clustering techniques used in EDM was very rare [1], say almost 8% of all the methods, based on the review of the season 2002 to 2017. All the same, here in this literature [9], association rule especially Apriori algorithm[38] with the help of the open source tool WEKA (Waikato Environment for Knowledge Analysis) used to predict the student performance. One of the other options of Apriori have used in this [38] paper by the author.
Three different methods namely Clustering, Decision Tree, NNW have used to evaluate Students performance in educational activities [10]. For the same Data collected from Sri Sai University – Palampur. With the usage of these techniques Teachers can easily evaluate the performance of the students.
Use of instance-based learning Classifier, Decision Tree and Naïve Bayes in predicting students at risk is discussed in [12], and also shown that K-Star has better results. However in this [15] the research used a couple of classifiers such as Discriminative or probabilistic classifier, Decision trees, Bayesian classifiers, Neural networks, K-Nearest Neighbour classifiers, Support vector machines and provided approaches for classification and analyse their suitability to the educational domain.
Many researches have used ID3 classification, whereas an implementation [8] of the same and the basic calculations with suitable examples are hard to find in simple terms. Almost detailed information about ID3 had given in this literature.
As similar to [11], [16] and [31] the experiment in [35] and [36] have used ID3, C45 and CART. The experimental results show that CART is the best algorithm for classification of data. The result of prediction accuracy former one is as

14721

International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 20 (2018) pp. 14717-14727 © Research India Publications. http://www.ripublication.com

follows ID3-52% C45-45% and CART- 56%, however the observation from the later had shown that J48 outperformed all. It says that depending on the data the results may vary, and one optimum solution with a single scenario is difficult. The next one [37] is a different implementation of C4.5 (formally J48 in Weka), TIAGA Algorithm. Recursive, Iterative and Hybrid forms of TIAGA were used to optimise C4.5. It handles continuous attribute values with the same naïve algorithm as used by RainForest, that is, when calculating the AVC-sets, one has to count the occurrences of all pairs of distinct attribute values and class labels. It can be improved by following methods proposed in [28].
Various data mining processes including Cluster Analysis, Classification, Regression Model were considered [18], and an elaborated study on EDM for the benefit of predicting students’ performance, raised by the interest of stakeholders have discussed.
Applying selected data mining algorithms for classification, aimed at revealing the high potential of data mining applications to predict student performance, was the keen topics over the work [19], and the project was implemented at a Bulgarian university. The prediction rates are not remarkable, it varied between 52% - 67%.
Many of the researches in the area of EMD are focussed on implementing a suitable solution for the dataset they use. Classification model generation comprehensive analysis of datasets using Bayesian, Support Vector Machine have been addressed in [21, 22]. Besides these MLP, J48, REPTree were also analysed [23] and found that MLP have produces almost
75% results and with experimenter option using F-Measure was about 82%. Therefore, performance of MLP is relatively higher than other classifiers. K-Nearest Neighbour, ID3 were also there, however there is no single classification which always produce best result, and did vary on datasets.
Apart from [21] and [22], a slight improvisation to data mining procedures in EDM can be increased, utilising Apriori, and AdaBoost. R. Kumar in this paper [23] provide an inclusive survey of different classification algorithms.
S. A. Kumar in the paper [24] implemented the J48 DT on student data sets to predict students’ performance based in their internal examination Marks, and found useful in many scenarios. While paper [24] dealt with multivariate data Lakshmi in [25] was trying to use qualitative data of students, to predict performances. The work focused on ID3, C4.5 and CART later their results are compared. The findings shown that CART has the best classification accuracy when compared to ID3 and C4.5. Also there was one another research [30] on similar combinations combination of methods like ID3 and C4.5.
It was interesting to see a lengthy list of algorithms namely Naïve Bayes, Random Tree, Multi-Class Classifier, which are Naïve Bayes, Random Tree, Multi-Class Classifier, Conjunctive Rule and Nearest Neighbour-Lazy IB in a recent work of H. Nawang [26]. The study was to classifying students’ performance in SijilPelajaran Malaysia Pelajaran Malaysia. This helped teachers plan suitable teaching

activities for their students based on the students’ performance.
The best among three strategies (calculation of local threshold, calculation of Gain and calculation of Gain in main memory method of RainForest algorithm) by this author in [28] for computing the information gain of continuous attributes and all adopt a binary search of the threshold in the whole training set starting from the local threshold computed at a node. It improved the performance of C4.5 decision tree.
The research work [33] used a dataset, most similar to KTUSNG-data, was an extracted useful knowledge of graduate students data collected from the College of Science and Technology – Khanyounis. Though data was similar, it concluded that association, classification, clustering and the outlier detection can help to improve students’ performance in one way or the other. A proposal to suitable data mining techniques are required to measure, monitor and infer these non- cognitive factors (set of behaviours, skills, attitudes), there are large numbers of factors that play significant role in performance, for prediction have made in this literature[34] by analysing a decade of research works from 2002 to 2014.

IMPLEMENTATION
Dataset, xAPI-Edu-Data is taken from a Learning Management System Kalboard-360 of University of Jordan, and the dataset consists of 480 student records and 16 features. At the same time, the KTU_SNG-Data is collected from one of the colleges of Kerala Technological University renowned as SNG College of Engineering, and the sample data contain 496 records and 60 features. A sample description of the attributes of data is given in table (Table 3)

Table 3. Sample Attributes of the acquired dataset with descriptions and domain values

Attributes Gender DOB
Mother Tongue Economically
Backward Handicapped Admission Type
Branch
Internal
Attendance
Total Credits Result

Description

Values

Student’s sex

{M,F}

Date of Birth

Varying for samples

For language fluency {Malayalam, Tamil}

Financial status of the family

{True, False}

Health-based analysis

{True, False}

Type of Admission {Regular, Lateral}

Department of the {CS, CE, ME, ECE, EEE,

student

NASB}

Class performance

Based on internal class performance

Regularity

Based on the attendance count

Marks obtained

Given in percentage

Calculated from marks

{PASS, FAIL}

As is known it contain personal data, qualification data, physical data, activity data, and academic data with class value

14722

International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 20 (2018) pp. 14717-14727 © Research India Publications. http://www.ripublication.com

as the final one. Since the actual dataset contain 496 records and 60 features most of them are not really informative, of which dataset which contains 13920 records is considered (includes 232 samples with 60 attributes). Later it is being preprocessed to one containing 45 attributes in this research. xAPI-Edu-Data, KTU_SNG-Data are the two similar datasets used here. Those are indeed collected from two geographical locations and the first one available in UCI repository, later one is a new dataset of SNG College affiliated to Kerala Technological University, Kerala, India. x-API-Data was a complete data, whereas KTU_SNG-Data have missing values, hence data preparation is being followed required file processing. The framework proposed have a core stage where basic checks were done and the datasets were prepared using 10Fold Cross Validation and Percentage Split before, preprocessing, and selection. Data get transformed to train and test data. Later these data were pre-processed, attributes are selected and given data for classification with various analysis based on a set standard conditions. This will give an efficient analysis on student prediction accuracy and help to compare them individually. The results of Model were taken and compared. Results are Obtained and tabulated in table (Table 4). It is clear from the result that few were underperforming compared to other algorithms those perform comparatively better to others.
Table4. Result of J48 and others with Datasets of default features

RESULTS ANALYSIS AND DISCUSSIONS
One of the interesting observation made in this research found to be the contribution or utilisation of C4.5 (or Equivalent J48 in Weka) throughout the selected, analysed and listed literatures. The results of these works have, but, a huge variation in their outcome. J48 have produced result accuracy ranging from 31% to 99.87%.
More than 50% of all related works discussed in Section IV have considered J48 as one of the classifier among others. Huge number of works have been considered this due to its capabilities to handle, nominal, numeric and missing values in data. Interestingly, found that J48 have produced a wide range of accuracy in prediction at various researches. The variation is plotted in figure (Fig. 6). So it is considered as a next stage of research to experiment with J48 in one another situation and come up with a new model of results. Hope that, it shall also add knowledge into the existing treasure achieve.

Figure 6: Min and Max Accuracy obtained by J48

During this study of selected classification algorithms corresponding to the parameters given in the (Section III. (V)), and available measures of evaluation in Weka are highly used and kept for analysis one of the sample decision tree. Precision (2), Recall (3) and F-Measure (4) are considered more along with the traditional solutions. Classifier is given along with figure (Fig. 5).
Figure 5: Tuned J48 Decision Tree Classification Sample of KTU_SNG-Data

The findings says that the result are varying for the same configuration setup of algorithms applied to two similar datasets. However, the result of basic set-up had shown surprisingly best result with J48 Decision Tree algorithms (100%, 77.83%), algorithms Multi-Layer Perceptron (99.04%, 76.38%), REPTree (100%, 65.38%) have also shown a good result. As it is found in table (Table 4) that the KTU-SNGData contain some misleading attributes which filters other attributes to contribute. However it is strongly believed that, and few independent variable have a good correlation misleading with the result. The table (Table 6) depicts these based on the comparative analysis on the results. And the figure (Fig. 7) portraits the result of algorithms in a more meaningful way. The J48 algorithm with both datasets have good results, however the initial experiments have produced quite surprising and unrealistic results. It is found in the analysis that quite a few features in the data sets are misleading the results by eliminating others to contribute or the one which found misleading are not a realistic, and hence is not a decisive feature of the dataset. Having eliminated this feature, it has a fall of almost 10% to15% in the results of Tuned J48 (90.09%, 83.34%). The REPTree and MLP with 89.47%, 85.71% and 88.02%, 86.77%. The fall show that the feature values are critical but the values have to be retained more realistically so that it can also contribute the similar way as other features of the dataset. A good method is to be

14723

International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 20 (2018) pp. 14717-14727 © Research India Publications. http://www.ripublication.com
devised to tune these values for NatureOfStudents in KTU_SNG-Data.
Table 6. Result of Tuned-J48 with eliminated features
Figure 8: Cost curve of PASS J48 with 10 Fold

Figure 7: Tuned-J48 Classifier Evaluation
The following table (Table 7) precisely explains with TP, FP, PRECISION, RECALL and F-MEASURE of the basic analysis of classification algorithms, by providing along with the results of experiments with the refined data and tuned J48. In this process, dataset in which misleading feature consciously eliminated and experimented. Setting a threshold of 4 for the FN and 2 for the FP for various algorithms, it is found that the cost-sensitive calculation is increased 20% more than base cost of 48% for PASS. And the average cost estimated is 0.2931%. The Cost Curve of DT (J48) at two scenarios are given in figures (Fig. 8 and Fig. 9).
Table 7. Result Obtained During the Implementation of J48

Figure 9: Cost curve of PASS J48 with 70% split.
The result of Tuned J48 is given in figure (Fig.10). During the experiments, with a percentage population of 78.45%, target percentage of 94.02% and score threshold of 0.146, the Tuned J48 has produced up to 91.3791% of Accuracy. This was the one of the best accuracy produce by J48% for PASS, the picture (Fig. 11) conveys the values effectively.

Figure 10: Result received by Implementing J48 14724

International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 20 (2018) pp. 14717-14727 © Research India Publications. http://www.ripublication.com
models, ensemble models or even a Hybrid implementation of classification techniques are good choices and an improved framework of these complying with KTU_SNG-Data may power good realistic results. A most suitable model of J48 to predict better accuracy for KTU_SNG-Data can be found from further researches. A 5% improvement would be a great achievement. Whatsoever, the final outcome of the continuing research will help stakeholders and miners to take a better managerial decision on their own sector and students can have an early awareness and correct themselves using these results.
Also these studies can be extended as future work to other areas of academics such as pedagogy evaluation, involvement, and attrition of students towards higher classes for multi-class classification problems with high dimensional values. The work may also have extended towards sports, healthcare, logistics for which it has to deal with huge data.

Figure 11: Cost/Benefit Analysis of J48
The variation happened in this analysis is found to be proceeded to other experimental researches. A better classification technique or a further improved implementation of decision tree J48 may bring a far better results with less variance. Combinational models, ensemble models or even a Hybrid implementation of classification techniques may be good choice for the next stages of analysis.
CONCLUSION AND FUTURE PROSPECTS
The research explains performances of classification methods against custom tuned J48. The evaluation of ten classifiers and their results generated, and results in classification, depending on the datasets used. Performances of the Naïve Bayes, Bayes Net, Multilayer Perceptron, SVM, REPTree, and Random Forest were analysed against tuned J48, and found that later one performed to generate better realistic outcomes. It is found that the cost-sensitive calculation of J48 produced with 90.8% preciseness whereas with custom tuned J48 it has generated 91.38% accuracy. The next stage of this research is to generate a better classifier model for KTU_SND-Data trying out different possibilities. The work proposed, in this literature is found effective in laying a stepping stone towards the continuing process of performance optimisation of classifier J48 with the aforesaid dataset.
It is also apparent that J48 had higher variation in results. Hence a better implementation of the more refined J48 algorithms is indeed a respectable idea of future research. Tuned J48 can be comparatively considered because of the competitive 91.38% rate and its nature of predictions and based on the characteristics analysed through this research. 71.57% of TP, 16.81% TN 3.88% FP and, 4.74% FN are the Confusion Matrix percentage rate on a 2x2 cost matrix for values (2, 4) of Tuned J48. Altogether 8.62% incorrect classification has been produced in this. Combinational

ACKNOWLEDGEMENT
One of the dataset used in this work was provided by the Principal, SNG College of Engineering, Ernakulam, KeralaIndia. The author, therefore, acknowledge with thanks Principal, SNG for dataset and service support.
REFERENCES
[1] Anoopkumar M and A. M. J. Md Zubair Rahman, “A Review on Data Mining techniques and factors used in Educational Data Mining to predict student amelioration,” Proc. 2016 IEEE Int. Conf. on Data Min. Adv. Comput. SAPIENCE 2016, pp. 122–133, 2016 doi: 10.1109/SAPIENCE.2016.7684113.
[2] P. Ajith, M. Sai, and B. Tejaswi, “Evaluation of Student Performance: An Outlier Detection Perspective,” Int. J. Innov. Technol. Explore. Eng., vol. 2, no. 2, pp. 40–44, 2013.
[3] E. A. Amrieh and I. Aljarah, “Preprocessing and Analyzing Educational Data Set Using X- API for Improving Student ’ s Performance,” no. November, 2015.
[4] E. A. Amrieh, T. Hamtini, and I. Aljarah, “Mining Educational Data to Predict Student’s academic Performance using Ensemble Methods X-API Data,” no. September, 2016.
[5] B. Baradwaj and S. Pal, “Mining educational data to analyze student’s performance,” Int. J. od Advanced Comput. Sci. Appl., vol. 2, no. 6, pp. 63–69, 2012.
[6] J. Barracosa and C. Antunes, “Anticipating teachers’ performance,” KDD 2011 Work., 2011.
[7] S. M. Bhandari and K. Patel, “--good--A Review on Using Clustering and Classification Techniques to Predict Student Failure with High Dimensional and Imbalanced Data,” pp. 50–54, 2015.
[8] R. Bhardwaj and S. Vatta, “Implementation of ID3 Algorithm,” Int. J. Adv. Res. Comput. Sci. Softw. Eng., vol. 3, no. 6, pp. 2128–2277, 2013.

14725

International Journal of Applied Engineering Research ISSN 0973-4562 Volume 13, Number 20 (2018) pp. 14717-14727 © Research India Publications. http://www.ripublication.com

[9] S. Borkar and K. Rajeswari, “Predicting students academic performance using education data mining,” Int. J. Comput. Sci. Mob. Comput., vol. 2, no. 7, pp. 273–279, 2013.
[10] E. R. Chuchra, E. Rimmy, and C. M. Tech, “Use of Data Mining Techniques for the Evaluation of Student Performance : A Case Study,” Int. J. Comput. Sci. Manag. Res., vol. 1, no. 3, pp. 425–433, 2012.
[11] A. Cufoglu, M. Lohi, and K. Madani, “A Comparative Study of Selected Classifiers with Classification Accuracy in User Profiling,” in 2009 WRI World Congress on Computer Science and Information Engineering, 2009.
[12] E. Er, “Identifying At-Risk Students Using Machine Learning Techniques,” Int. J. Mach. Learn. Comput., vol. 2, no. 4, pp. 476–480, 2012.
[13] P. Gulati and A. Sharma, “Educational Data Mining for Improving Educational Quality,” Int. J. Comput. Sci. Inf. Technol. Secur., vol. 2, no. 3, pp. 648–650, 2012.
[14] M. S. Halawa, M. Soliman Halawa, M. E. Shehab, and E. M. Ramzy Hamed, “Predicting Student Personality Based on a Data-Driven Model from Student Behavior on LMS and Social Networks,” in IEEE, 2015, pp. 294–299.
[15] W. Hämäläinen and M. Vinni, “Classifiers for Educational Data Mining,” Handb. Educ. Data Mining, Data Min. Knowl. Discov. Ser., p. 57--71., 2010.
[16] B. HSSINA, A. MERBOUHA, H. EZZIKOURI, and M. ERRITALI, “A comparative study of decision tree ID3 and C4.5,” Int. J. Adv. Comput. Sci. Appl., vol. 4, no. 2, pp. 13–19, 2014.
[17] R. A. Huebner, “A survey of educational data- mining research,” Nature, vol. 194, no. 4833, pp. 1006–1006, 2012.
[18] J. Jacob, K. Jha, P. Kotak, and S. Puthran, “Educational data mining techniques and their applications,” 2015 Int. Conf. Green Comput. Internet Things, ICGCIoT 2015, pp. 1344–1348, 2015.
[19] D. Kabakchieva and Dorina, “Predicting student performance by using data mining methods for classification,” Cybern. Inf. Technol., vol. 13, no. 1, pp. 61–72, Jan. 2013.
[20] G. Kaur and A. Chhabra, “Improved J48 Classification Algorithm for the Prediction of Diabetes,” Int. J. Comput. Appl., vol. 98, no. 22, pp. 13–17, 2014.
[21] P. Kaur, M. Singh, and G. S. Josan, “Classification and Prediction Based Data Mining Algorithms to Predict Slow Learners in Education Sector,” in Procedia Computer Science, 2015, vol. 57, pp. 500–508.
[22] G. Kesavaraj and S. Sukumaran, “A study on classification techniques in data mining,” 2013 Fourth Int. Conf. Comput. Commun. Netw. Technol., pp. 1–7, 2013.
[23] R. Kumar and D. Verma, “Classification Algorithms for Data Mining: A Survey,” Int. J. Innov. Eng. …, vol.

1, no. 2, pp. 7–14, 2012.
[24] S. A. Kumar and D. V. M.N, “Efficiency of Decision Trees in Predicting Student ’ S Academic Performance,” ©Cs It-Cscp 2011, pp. 335–343, 2011.
[25] T. M. Lakshmi, A. Martin, R. M. Begum, and V. P. Venkatesan, “An Analysis on Performance of Decision Tree Algorithms using Student’s Qualitative Data,” Int. J. Mod. Educ. Comput. Sci., vol. 5, no. 5, pp. 18–27, 2013.
[26] H. Nawang, “CLASSIFICATION MODEL AND ANALYSIS ON STUDENTS’ PERFORMANCE,” J. Fundam. Appl. Sci., vol. 9, no. 6S, pp. 869–885, 2017.
[27] T. R. Patil and M. S. S. Sherekar, “Performance Analysis of Naive Bayes and J48 Classification Algorithm for Data Classification,” Int. J. Comput. Sci. Appl., vol. 6, no. 2, 2013.
[28] S. Ruggieri, “Efficient C4.5,” IEEE Trans. Knowl. Data Eng., vol. 14, no. 2, pp. 438–444, 2002.
[29] T. Sharma, A. Sharma, and V. Mansotra, “Performance Analysis of Data Mining Classification Techniques on Public Health Care Data,” Int. J. Innov. Res. Comput. Commun. Eng., vol. 4, no. 6, pp. 11381–11386, 2016.
[30] S. Singh and V. Kumar, “Classification of Student’s data Using Data Data Mining Techniques for Training & Placement Department,” Int. J. Comput. Sci. Netw., vol. 1, no. 4, pp. 2277–5420, 2012.
[31] S. Singh and P. Gupta, “Comparative study ID3, cart and C4 . 5 Decision tree algorithm: a survey,” Int. J. Adv. Inf. Sci. Technol., vol. 27, no. 27, pp. 97–103, 2014.
[32] R. Sumitha, E. S. Vinothkumar, and P. Scholar, “Prediction of Students Outcome Using Data Mining Techniques,” Int. J. Sci. Eng. Appl. Sci., vol. 2, no. 6, pp. 132–139, 2016.
[33] M. M. A. Tair, A. M. El-Halees, M. M. Abu Tair, and A. M. El-Halees, “Mining Educational Data to Improve Students’ Performance: A Case Study,” Int. J. Inf. Commun. Technol. Res., vol. 2, no. 2, pp. 140–146, 2012.
[34] P. Thakar et al., “Performance analysis and prediction in educational data mining: A research travelogue,” Int. J. Comput. Appl., vol. 110, no. 15, pp. 60–68, 2015.
[35] S. K. S. K. Yadav, B. Bharadwaj, and S. Pal, “Data Mining Applications: A Comparative Study for Predicting Student’s performance,” Int. J. Innov. Technol. Creat. Eng., vol. 1, no. 12, pp. 13–19, 2012.
[36] S. K. Yadav and S. Pal, “Data Mining : A Prediction for Performance Improvement of Engineering Students using Classification,” World Comput. Sci. Inf. Technol. J. WCSIT, vol. 2, no. 2, pp. 51–56, 2012.
[37] Y. Yang and W. Chen, “Taiga: Performance optimization of the C4.5 decision tree construction algorithm,” Tsinghua Sci. Technol., vol. 21, no. 4, pp. 415–425, 2016.
[38] Anoopkumar M and A. M. J. Md Zubair Rahman,

14726
DataClassificationMethodsDatasetStudents