An enhanced feed-forward neural networks and a rule-based algorithm for predictive modelling of students' academic performance
Feed-forward Neural Networks, is a multilayer perceptron and a network structure capable of modelling the class prediction as a nonlinear combination of the inputs. The network has proven its suitability in solving several complex tasks. But sometimes, it has challenges of over-fitting, especially w...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2016
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/15251/1/An%20enhanced%20feed-forward%20neural%20networks%20and%20a%20rule-based%20algorithm%20for%20predictive%20modelling%20of%20students%27%20academic%20performance.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Feed-forward Neural Networks, is a multilayer perceptron and a network structure capable of modelling the class prediction as a nonlinear combination of the inputs. The network has proven its suitability in solving several complex tasks. But sometimes, it has challenges of over-fitting, especially when fitting models from massive data of varied data points. This necessitates its enhancement in order to strengthen its performance. Such enhancement would ensure a predictive network model that can generalize well with a set of untrained data. In this research, in order to alleviate the possibility of over-fitting in a network predictive model, a dynamic partitioning of the dataset is proposed. Also, for a more efficient exploration of students‟ data collected for this research, a Rule-Based Algorithm is proposed and implemented. The predictive models emanated from the two approaches were evaluated in order to validate their effectiveness. The enhancement done to the Feed-forward Neural Networks (FNN) in the first approach, ensure partitioning of the dataset that is based on the size of the data available for creating the model. The evaluation carried out on the Enhanced Feed-forward Neural Network (EFNN) models show that, there is a decrease in error from 0.261 to 0.029. Similarly, another set of 2000 students‟ data is trained, the error recorded when the network model is simulated with untrained 500 data show that, there is a reduction in error from 0.0095 to 0.00033. Most of the training performance generated from the network models created also shows that, the EFNN has lower errors and converge faster. The implementation of the rule-based algorithm proposed in the second approach, shows outputs that are consistently accurate. Its efficiency is compared to some existing techniques reported in the literature for the predictive modelling of students‟ academic performance. Findings from the comparison show that, the proposed RBA explores students‟ data much better. It can also serve as an alternative algorithm to the use of machine learning techniques in the exploration of students‟ data for prediction purposes. |
|---|