Several different algorithms are applied for building the classification model, each of them using different classification techniques. In the first stage of this analysis, learners’ variables were entered in a Microsoft Excel spread sheet file and saved in CSV format, which is suitable for WEKA tools (Plate 1). After changing MS Excel to the CSV format, data was extracted from the CSV into the WEKA tool (Plate 2). The classification panel enables us to apply classification and regression algorithms to the resulting dataset, estimate the accuracy of the resulting predictive model, and visualise erroneous predictions or the model itself. The classification algorithms used for this study are ID3, C4.5, and CART. A 10-fold cross-validation has been performed to verify each classifier. After performing the ten tests, the average test output was used to assess the accuracy of the model. The authors do the data-cleaning process, such as resolving missing values in the dataset and facilitating datasets with the correct attributes.
Next, the authors checked the standardisation or distribution of data to determine whether the distribution of data was normal or balanced and to minimise the effects of prediction errors during the modelling process. In addition, the imputed value technique was implemented to fill in the missing value in the parents' income variable with its mean values to mitigate bias in the dataset. Lastly, the authors have 150 data points with 14 variables as parameter inputs from the data-cleaning process.
3.1. Socioeconomic and Demographic Features
Based on the data in Plate 2, 64% of the learners were male, while women made up only 33% of the sample respondents. A percentage of 44.67 per cent of learners worked below 8 hours a day, while 47.33% of them worked more than 8 hours a day. Regarding the age of the learners, 72% of adult learners are mostly above 46 years of age, followed by those between 30 and 45 years of age, and 9.33% of the learners are less than 30 years of age. The dataset also showed that 70% of learners had previously completed primary school, while 30% could not complete primary school. The majority of the spouse’s educational attainment was above the primary school certificate. The learner's income is mostly below N30, 000 a month. These data metrics, in terms of demographic features, identify that the dataset has a relatively good variation to be used in the learner dropout indicator prediction.
3.2. Results for Classification Algorithms
Table 2 and Fig. 2 show the accuracy of ID3, C4.5, and CART algorithms for the classification using 10-fold cross-validation. Since there is no separate evaluation data set, this is necessary to get a reasonable idea of the accuracy of the generated model. The model is generated in the form of a decision tree.
Table 2: Performance of the classifiers
|
Algorithms
|
Accuracy
|
Precision
|
Recall
|
Incorrectly classified instance
|
Error rate
|
Time taken
|
C4.5
|
80
|
74
|
95.9
|
2.81
|
0.12
|
0.032sec
|
ID3
|
79.50
|
73
|
92.3
|
1.92
|
0.059
|
0.05sec
|
CART
|
78.80
|
71
|
92.3
|
1.91
|
0.0522
|
0.58sec
|
The C4.5 decision tree algorithm used gave a prediction accuracy of 80%, an incorrect classification instance of 2.81%, a recall of 95.9%, a precision of 74.4%, and a time taken of 0.032 seconds sequentially. The ID3 algorithm used gave a prediction accuracy of 79.50%, an incorrect classification instance of 1.92%, a recall of 92.3%, a precision of 74%, and a time taken of 0.5 sec sequentially. The Simple CART tree algorithm used gave a prediction accuracy of 78.8%, an incorrect classification instance of 1.92%, a recall of 92.3%, a precision of 71%, and a time taken of 0.58 seconds sequentially. This result is similar to the results obtained by [14], which compared CART, C4.5, and ID3 in predicting students’ academic performance. Their experimental results show that C4.5 performed better than ID3 and CART. [15] also compared the classification accuracy of four algorithms, namely ID3, C4.5, Simple CART, and an alternating decision tree, on the student’s dataset. According to their simulation results, the C4.5 classifier outperforms the ID3, CART, and AD Tree in terms of classification accuracy. Moreover, [16] applied the same set of algorithms (CART, C4.5, and ID3) to predict students’ academic performance. Their results show that the CART has the best classification accuracy when compared to ID3 and C4.5, which is not consistent with our results. Similarly, [17] study on student dropout predictors in Kavumu TVET School indicated that ID3 was the best algorithm compared to C4.5 and CART.
3.3. Results for Rank Attributes and Simulation
The simulation studies involve ranking attributes using the gain ratio presented in Table 3.
Table 3: Simulation Result for Rank Attributes using Gain Ratio
|
No.
|
Variable
|
Value
|
Rank
|
1
|
Age
|
1.22
|
1st
|
2
|
Program relevance
|
1.10
|
2nd
|
3
|
Number of Working Hours
|
1.08
|
3rd
|
4
|
Previous Education
|
1.08
|
3rd
|
5
|
Monthly Income
|
0.94
|
5th
|
6
|
Employment Status
|
0.94
|
5th
|
7
|
Mobility
|
0.89
|
7th
|
8
|
Home location
|
0.89
|
7th
|
9
|
Occupation
|
0.85
|
9th
|
10
|
Spouse Education
|
0.79
|
10th
|
11
|
no of time you visit the hospital for treatment
|
0.75
|
11th
|
12
|
Gender
|
0.74
|
12th
|
13
|
Marital status
|
0.66
|
13th
|
14
|
Number of children
|
0.57
|
14th
|
Through a series of tests, the authors identified that the key attributes that affect learner dropouts are mostly age, the relevance of the programme, Number of Working Hours, income, employment status, and mobility. The three least important items stated were gender, marital status, and number of children. The result of this study is contrary to that obtained by [4] who have found that the location of a student's residence, whether rural or urban, has been related to the attributes of dropouts [4]. [18] and [7] acknowledged that financial aid contributes to academic persistence. Similarly, [17] in their study on student dropout predictors in Kavumu TVET School, found that the key attributes that profoundly affect student dropouts are academic performance.