Towards an implementation of instance-based classifiers in pedagogical environment
Monitoring individual student academic performance is indispensable to educational institutes since they are required to provide evidence of their students' academic performance to diverse governmental bureaus. Machine learning classifiers appear productive tools for this purpose; however, inst...
Saved in:
| Main Authors: | , , , |
|---|---|
| Other Authors: | |
| Format: | Article |
| Published: |
Taylor's University
2023
|
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Universiti Tenaga Nasional |
| Summary: | Monitoring individual student academic performance is indispensable to educational institutes since they are required to provide evidence of their students' academic performance to diverse governmental bureaus. Machine learning classifiers appear productive tools for this purpose; however, instance-based machine learning classifiers have acquired the least consideration. This research measures the suitability of instance-based classifiers, exclusively k-Nearest Neighbours (k-NN) and Locally Weighted Learning (LWL), in the pedagogical environment and proposes solutions to issues related to this class of classifiers. The performance of these classifiers depends upon the number of nearest neighbours (k) and the distance metrics. We performed experiments, with varying values of k and different distance metrics, to evaluate the performance of k-NN and LWL. To authenticate the conclusions drawn from these experiments, we carried out experimental evaluation with 3 more datasets taken from another research. This comparison evidences the suitability of instance-based classifiers, in pedagogical environment, especially LWL which is one of the least addressed classifiers. The comparative analysis highlights the fact that varying value of k and changing the distance metric optimistically affect the classifier's performance. Even though Manhattan distance metric dominates in achieving higher accuracy; however, classifiers may act differently for dissimilar datasets. To resolve this shortfall, we propose a novel framework which carries out extensive experiments with varying value of k and changing distance metrics and conclude a prediction model which emerges appropriate for the provided training dataset. The framework takes training dataset from an instructor and recommends suitable instance-based learning prediction model. � 2021 Taylor's University. All rights reserved. |
|---|