Hypertuned Tree based Devising E-learning usability framework

Abstract

The usability of the mobile e-learning app depends generally on functions such as efficiency, effectiveness, learning, cognition, memory, etc. There is no tool or theory for evaluating and accessing user perceptions and usability features which are categorized according to user perception. In this paper, we have used hybrid research based on quantitative and quality analyses. In quantitative analysis, we have designed surveys that collect student usability and perception demands for mobile applications. From the user's point of view (student), it was analyzed that efficiency is the highest quality which satisfies students, and then we used genetic algorithm to measure usability. Each usability feature has been classified by weight. The GA-based Decision Trees and Random Forests hypertuned model has shown up to 90% accuracy, based on the Machine Learning Model prediction, to forecast the highly suggested usability functions. This methodology can be used in education to assess the key e-learning apps usability criteria.

Full text

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
Hypertuned Tree based Devising E-learning usability
framework
Muhammad Asghar1,*, Muhammad Ijaz2*, Fawwad Hassan Jaskani3
1Department of Computer Science, Islamia University of Bahawalpur
2Department of Computer Systems Engineering, Ontario Tech University
3Department of Computer Systems Engineering, Islamia University of Bahawalpur
Abstract
The usability of the mobile e-learning app depends generally on functions such as efficiency, effectiveness, learning,
cognition, memory, etc. There is no tool or theory for evaluating and accessing user perceptions and usability features which
are categorized according to user perception. In this paper, we have used hybrid research based on quantitative and quality
analyses. In quantitative analysis, we have designed surveys that collect student usability and perception demands for mobile
applications. From the user's point of view (student), it was analyzed that efficiency is the highest quality which satisfies
students, and then we used genetic algorithm to measure usability. Each usability feature has been classified by weight. The
GA-based Decision Trees and Random Forests hypertuned model has shown up to 90% accuracy, based on the Machine
Learning Model prediction, to forecast the highly suggested usability functions. This methodology can be used in education to
assess the key e-learning apps usability criteria.
Keywords: Machine Learning, Creative Technologies, Mobile Learning Applications
1. Introduction
The key problem for mobile learners is the usability of mobile apps to provide consumers with a more or less optimal mobile
learning experience. In the study, learning, efficiency, memorability, and precision and user pleasure are the important five
components [1]. The usability element mostly rests on the concept of literature which is easy to use. It is easy to use if a user
is able to carry out tasks quickly and thus focus on boosting user efficiency and reducing problems from the user interface. It
also provides the pleasure of users to enhance attractive and aesthetically acceptable mobile learning applications for
consumers.
Remote training has unique advantages. It gives a winning technique to meet specific needs such as packed educational
facilities and can help students and instructors live away from schools and colleges every time and everywhere. Your time
"anywhere." If a training document is provided, it can be valuable for certain kinds of students, including students with
disabilities. E-learning is the latest technology for remote learning using online training materials and activities. Detailed data
and services are available to users, including cultural events, technological encounters, facilities and physical and cognitive
talents. E-learning facilities are extremely important for students around the country to reduce the digital divide and gap in
social and cultural applications. For e-learning software producers, maximum user compatibility and access should be an
essential objective and a precondition for users to use these applications financially.
The usability of a mobile app has numerous components that affect the overall efficiency of the user. Three main results
should be reached via an interface:
 Usability of the mobile app enables users to become comfortable with the UI · Users should achieve their application
objectives.
 Error-free programmes are necessary. If your application doesn't work well, the other portions are quiet.
 Usability problems are further exacerbated by user interface flaws that need to solve problems instead of every step or
click. Due to the fact that mobility users suffer compatibility problems while installing third party solutions, the system does
not support formats and data transmission. Because of the lack of reactive interfaces on diverse devices, users have
problems with usability. Even with mobile apps, when a new interface is shown, it is bothersome for users [1][4]. The
problems revealed in the study are only linked to patterns of usability in qualitative M-learning applications. This study also
shows these challenges resulting from approaches of algorithms to handle other usability problems. The proposed solution

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
supports a wide variety of applications through classic GA and generative machine learning techniques. This study also helps
to identify issues of usability for M-learning applications and to provide an integrated approach to problems of usability for
mobile apps
2. Literature Review
Oztekin et al [1] purposed a new usability evaluation method for e learning systems using machine learning . The proposed
approach will provide usability experts with useful feedback about what steps should be changed in order to optimise system
usability for a specific group of E-learning system end-users. The evaluation of E-learning systems' usability is a critical but
difficult task. Usability testing can benefit from machine learning techniques. Severity index values can be determined using a
pareto-like analysis. The most significant usability variables can be ranked using sensitivity analysis.
Georgouli et al [2] purposed a new framework for the variation of E-learning. They present a structure for fusing e-learning
into a conventional course in this paper. The system can be used to create an instructional model that incorporates a relevant
pedagogical setup that integrates learning and "learner-cantered" variables. They use a case study to demonstrate how the
recommended approach was used to make an effective change to a blended learning design that incorporates face-to-face
sessions with distance correspondence. They believe that e-learning approaches and resources can likewise help in efficiently
supporting students and enhancing the nature of learning, based on this experience and the results of assessments they have
conducted over the most recent couple of years (3 surveys with 316 members).
Fetaji et al [3] purposed a new devising M-learning usability framework. The proposed Framework of Usability Guidelines for M-
Learning is the research study's contribution. The developed 'usability paradigm' for m-learning is based on the empirical study
discussed here, which can be of great benefit and assistance to practitioners working in the field of m-learning.
Elkaseh et al [4] purposed a new framework for Critical Success Factors of E-Learning Implementation in Higher Education.
This paper provides a conceptual framework for e-learning implementation success factors in higher education. The aim of
this study was to identify critical success factors that influence the success of e-learning implementation.
Granic et al [5] purposed a new framework for e-Learning. The paper goes into detail on how to build an empirically validated
pedagogical system for e-Learning. The structure is intended to support the subject matter in a classroom environment.
Makina et al [6] purposed a new framework for dealing with the quality use of web recordings in open distance and e-
learning environments. This paper expects to design and establish a system for dealing with the quality use of digital
recordings in ODeL environments for teaching and learning. A developmental qualitative research design was used to create a
methodology utilizing a literature review. The discoveries established a structure for academic developers, learning
technologists, and course designers concerned with online quality.
Kalaivani et al [7] provide an analysis of generic e-learning framework. This thesis is an honest attempt to create an ideal
domain-independent architecture for e-learning environments. The generic system, it is suggested, would suit all e-learning
contexts, from arts to medicine. It is primarily concerned with the possibility of having all environments under one roof,
despite their diversity. This project will result in an analysis model that incorporates current e-learning principles proposed by
IEEE LTSC. The research model is based on a modern e-learning framework that has been proposed.
Sood et al [8] purposed a new enhancing E-learning through educational games. Virtual reality (VR) combined with cloud
computing has paved the way for e-learning to be available to people all over the world. Educational games based on
electroencephalography (EEG) are commonly used to help students develop their cognitive and learning skills. The
framework's efficiency is assessed by calculating average latency, network use, and energy consumption. The findings
demonstrate the optical network's major advantages in playing EEG-based games.
Sigei et al [9] purposed a new E-learning framework based on responsive web design. This research aims to recommend an
ideal responsive web design strategy for an e-learning environment that will enhance e-learning systems' mobile usability.
Because of its popularity, customizability, and broad community support, a case study of the Modular Object-Oriented
Dynamic Learning Environment (MOODLE) Learning Management System (LMS) was considered in this work. We created an
e-learning system based on Google Web-Friendly Test's findings about the MOODLE LMS's mobile usability issues. The
developed responsive e-learning system was compared to the MOODLE LMS, which is not responsive. The RWD-based e-
learning system earned a 95 percent mobile user experience ranking, compared to 63 percent for the non-responsive
MOODLE LMS.
Czerkawski et al [10] purposed a new framework for fostering student engagement in online learning. This structure is the
climax of a careful literature review on student investment, fully intent on presenting the discoveries in a consistent manner
for online teachers. Online designers and teachers need better approaches to increasing student interaction for e-learning
design and development to be effective, and the creators hope that the proposed framework offers such a methodology.
Omar et al [11] purposed a new framework to personalize e-learning. We suggest a structure for personalizing e-learning that
emphasizes the importance of giving close consideration to singular learning styles. We concentrate on understanding learners'
learning propensities and the order in which they choose learning materials in relation to their learning styles. A prototype for

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
an adaptive web-based course has been developed, in which the learning environment adjusts its behaviour to coordinate
with the learning styles of the students.
Dong et al [12] purposed an E-Learning framework embracing cloud computing . The Blue-sky cloud architecture provides E-
Learning systems with dependable, flexible, and cost-effective services. For E-learning systems, physical computers are
virtualized and allocated on request. It addresses the problems of resource allocation optimization, dynamic concurrency
demands, and rapid storage development.
Joseph [13] purposed a new machine learning framework for feedback and observing E-Learning Impact. This investigation
tries to use machine learning approaches to customize not the content, but rather the presentation of the content based on
practically universal learning objectives - like how a teacher could change the presentation of the content based on student
feedback if certain aspects are unclear to them. The most troublesome piece of creating such an observing framework is
deciding what elements of the interaction ought to be controlled and how these ought to be interpreted as contribution with
actionable experiences - in other words, deciding the learning schema and afterward applying learning calculations to gauge
the learner's interest or disinterest in the content provided.
Abbas et al [14] purposed a semantic lattice based E-Learning framework. We present SELF, a semantic lattice based
ELearning Framework, in this paper. By proposing a well-defined interaction plan among currently available devices and
technologies, SELF expects to identify key enablers in a realistic network based E-learning environment and to minimize
technical reworking. On top, we have E-learning-specific implementation layers, and underneath, we have semantic network
based help layers. In SELF, we likewise map the latest open and freeware developments to different components.
Neji et al [15] purposed a new affective e-Learning framework . This paper investigates the use of affective communication in
simulated environments based on facial expression (FE). We investigate VEs' ability to communicate affective states, as well
as aspects of communication related to emotional response. Affective behaviours, we conclude, will add a new dimension to
interactive e-learning systems.
Senthilnayaki et al [16] purposed a new framework for intelligent web based e-Learning. To teach the subject Database
Management Systems, the proposed cosmology discovery approach was empirically tested in an e-Learning setting. The use of the
Jaccard Similarity measure and the K-Means clustering estimation for clustering learners, similarly as the use of cosmology for term
comprehension and learning style recognition, are the paper's principle commitments. This guides adaptive e-learning by offering
appropriate decision-creation feedback, and it employs decision rules to provide intelligent e-learning.
Shrivastava et al [17] purposed a new framework for web based e-learning. Author present a report on how to extract useful
information from the web, as well as a basic understanding of semantic web and categories. This paper outlines an existing E-
Learning paradigm and suggests a new structure for it.
Chirila et al [18] purposed a new framework for a competencies based E-Learning. Traditional education programmes are
extremely theoretical, almost entirely dependent on memorising student skills. European society's growth today focuses on
functional skills, globalisation, and competition. Existing Romanian E-learning systems aim to spectacularly present and clarify
information using multimedia, but none of them are focused on the principle of competence. In this context, the alternative
is a modern educational framework focused on skills in real-life situations.
Chan [19] purposed a new framework for assessing usage of web-based e-learning. The use of a web usage mining method to
discover students' information from their usage sequences is presented in this paper. An eLearning system's lesson structure
can be used as a website metaphysics to promote usage mining. We likewise present a system for representing evaluation
results and usage patterns using a hierarchy of stream outlines, which can be used as decision estimations to evaluate
relationships between usage patterns and student success. For forecasting student success and delivering timely lessons,
usage patterns can be translated into decision rules or trees.
Shi et al [20] purposed a new framework for e-learning. To begin, we created a multidimensional knowledge outline framework that
stores learning objects in multiple classes separately. Then, in the information outline, we proposed six key semantic relationships
between learning objects. Second, a learning way recommendation model based on the multidimensional knowledge graph
framework is designed to meet different learning needs. It can generate and recommend customized learning ways based on the e-
target learner's learning object. The results of the experiment show that the proposed model can generate and recommend
qualified personalized learning approaches to help e-learners improve their learning experiences.
Kurdi et al [21] purposed a theoretical framework for e-learning. The aim of this study is to figure out what factors affect
students' acceptance of E-learning. On the basis of technology acceptance, a theoretical framework was developed. The best
and most critical predictors of students' intention to use E-learning systems are social impact, perceived pleasure, and self-
efficacy, perceived utility, and perceived ease of use.
Sudhana et al [22] purposed a new framework for e-learning system. This paper describes an ontology-based paradigm for
context-aware adaptive learning systems, including extensive discussions of categorization, modelling, and using ontology to
specifically define learner context in an e-learning setting. Finally, we demonstrate the utility of the proposed ontology by
providing an architectural description of an e-learning environment.
Colace et al [23] purposed a sentiment analysis framework for e-learning. This paper outlines an ontology-based architecture
for a context-aware adaptive learning system, including extensive discussions of categorization, modelling, and the use of
ontology to specifically define the context. In an e-learning environment, the learner background is essential. Finally, we
demonstrate the utility of the proposed ontology by providing an architectural description of an e-learning environment.

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
 [24] purposed a new framework for e-Learning. A pedagogical framework (PF) for e- and m-Learning in secondary
schools was designed, designed, and validated by European partners. The PF is an important part of the conceptualization
and development of an e-Learning system. It provides sound concepts for the creation of learning scenarios that will improve
students' learning experiences. Teachers' implementation of the framework grew in mastery in the school context as their
experience with the system grew. The goal of the project was to create this support structure in order to maximise the
system's benefits for teaching and learning.
Scalise et al [25] purposed a new framework for e-learning. They present a scientific arrangement or categorization of 28
novel item types that could be useful in computer-based testing. The proposed scientific characterization defines a collection
of famous item categories known as "intermediate requirement" objects, which are organized by the degree of constraint on
the respondent's possibilities for answering or engaging with the evaluation item or errand. These item types have responses
that fall somewhere between completely constrained responses (i.e., the conventional multiple-choice question), which can
be excessively restrictive to completely exploit the potential of new information technologies, and completely constructed
responses (i.e., the standard essay), which can be difficult for computers to meaningfully analyse even with labelling. The 28
example types discussed in this paper are organized into seven gatherings, with response imperatives decreasing from
completely selected to completely develop. There are four exemplary examples of each requirement bundle. The proposed
scientific categorization's point is to provide a useful context for the conversation of new assessment arrangements and uses
in computer-based environments, similarly as a realistic resource for assessment developers.
3. Methodology
This section describes the main methodology of this research. First we have collected information form students and on the
basis of their recommendation we have scored the usability features by using Genetic Algorithm, details are as follows:
As we have used supervised and unsupervised classifiers for our data validation, so in this step we train our unsupervised learning
classifier with unlabelled data. And send the results to the normalized database for clustering. First we have used Genetic algorithm
in this approach and then applied support vector machine classification model for predicting the best features.
a. Inclusion of Multi-Objective Optimization
An optimization approach should be used for the right and optimum solution or benefit. The dilemma of optimization
comprises of one or more goals or of seeking minimal or optimum benefit. MOO is called; the problems that have more than
one goal. Such questions can be seen in everyday life as in the fields of architecture, mathematics, economy, social sciences,
agriculture, automobile, aviation and more. In different daily challenges, the priorities under discussion are confounding. And
refining a given approach for a single purpose will contribute to unintended outcomes for the other goals.
A logical and practical approach to a multi-objective dilemma is to pursue any solution that achieves its goals at the required
level without any additional solution. According to our work, two GA and NSGA-II algorithms are built for problems with
multiple objectives.
GA and NSGA-II are metaheuristic and ideally suited to various kinds of issues addressed in the next segment. In the other
side, conventional GA and NSGA-II are updated to include numerous objective challenges by providing answers to the
diversity and the usage of exercise functions.
b. Multi-Objective Genetic Algorithm
Multi-target GA is a population-based solution. It is useful for addressing multi-target challenges and interface difficulties. A
GA may be adapted to find a variety of solutions that are not affected by each other in one round. The capacity of GA to scan
simultaneously multiple areas of solution space and can identify numerous solutions for difficult multimodal, discontinuous
and non-convex solutions.
GA crossover operator may accomplish decent solutions systems with diverse goals in order to create new solutions that are
not affected by unknown areas of the Pareto front.
In comparison, the operator would not use any of the multi-objective GA to scale and prioritize goals. GA was thus the most
popular heuristic approach used in our work for optimizing features.
c. GA Based Random Forests Classification
RF is a random choice forest a grouping and other activities ensemble learning strategy that operates by creating a learning
management in decision-making trees by training and the average assessment of the individual woodlands. Section 5
discusses the more comprehensive and description of the random forest classifier.
On the qualified dataset, DT was applied. It partitions the data into separated layers these regression layers are measured,
apart from that, break points are more and more reliable and the training of less DT samples is easier. Often known as
random decision woods are random forests. The diagram below illustrates the fundamental vocabulary of random forests
and decision trees.

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
Figure 1 Ensemble learning model
d. Training and Validation of Models
To ensure that the concept is well prepared, it must be checked and evaluated. Three data sets, a preparation, validation and
test collection, are used for this. The training sets are used to train the model, while the validation collection is used to
measure the error on the trained model and pick the better model, if several are computed. The test collection is then used
to calculate the precision of the model chosen. If only one model with the training data is educated, a training and test set
would be appropriate. Education, validation and test failures may be calculated by the residual square number (RSS).
e. Hyper parameters for Decision Trees:
Decision tree nodes split on the basis of impurity. Whereas impurity is the measure of homogeneity of the labels on a node.
There are many ways to measure impurity two of which Scikit-learn has implemented is the information gain and Gini
impurity or Gini Index. Entropy measure as the impurity measure and splits a node such that it gives the most amount of
information gain described as:
 =1 + + ( )
Information gain (eq. described below) uses the entropy measure as the impurity measure and splits a node such that it gives
the most amount of information gain. Whereas Gini Impurity measures the divergences between the probability distributions
of the target attribute that it gives the least amount of impurity.
=

log2

log2

+
+
+
+
The total gain (criterion {
=  
Many of the researchers point out that in most of the cases, the choice of splitting criteria will not make much difference in the tree
performance. Each criterion is superior in some cases and inferior in others, as the 
f. Maximum Depth of a Tree
Max_Depth: int or none, optional (default=none)
The maximum depth of the tree, if none, then nodes are expanded until all leaves are pure or until all leaves contain less than
min_samples_split samples.
The theoretical maximum depth a decision tree can achieve is one less than the number of training samples, but no algorithm
will let us reach this point for obvious reasons, one big reason being overfitting. Note here that it is the number of training
samples and not the number of features because the data can be split on the same feature multiple times.
g. Minimum Weight Fraction Leaf
min_weight_fraction_leaf: float, optional (default=0.)

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
The minimum weighted fraction of the sum total of weights (of all the input samples) required to be at a leaf node. Samples
have equal weight when sample weight is not provided.
Min_weight_fraction_leaf is the fraction of the input samples required to be at a leaf node where weights are determined by
sample weight, this is a way to deal with class imbalance. Class balancing can be done by sampling an equal number of
samples from each class, or preferably by normalizing the sum of the sample weights for each class to the same value. Also
note that min_weight_fraction_leaf will then be less biased toward dominant classes than criteria that are not aware of the
sample weights, like min_samples_leaf.
If the samples are weighted, it will be easier to optimize the tree structure using weight-based pre-pruning criterion such as
min_weight_fraction_leaf, which ensure that leaf nodes contain at least a fraction of the overall sum of the sample weights.
h. Minimum Purity Decrease
min_impurity_decrease: float, optional (default=0.)
A node will be split if this split induces a decrease of the impurity greater than or equal to this value.
Min_impurity_decrease helps us control how deep our tree grows based on the impurity. But, what is this impurity and how
does this affect our decision tree? Remember in the criterion section we quickly looked at Gini Index and Entropy, well, these
are a measure of impurity. The impurity measure defines how well a number of classes are separated.
i. Maximum Weights
class_weight: dict, list of dicts, none, default=none
Weights associated with classes in the form {class_label: weight}. If not given, all classes are supposed to have weight one.
For multi-output problems, a list of dicts can be provided in the same order as the columns of y.
Class_weight is used to provide a weight or bias for each output class. But what does this actually mean, see when the
algorithm calculates the entropy or gini impurity to make the split at a node, the resulting child nodes are weighted by the
class_weight giving the child samples weights based on the class proportion you specify.
This can be highly useful when you have an imbalanced dataset. Usually, you can just start with the distribution of your
classes as the class weights and then depending on where your decision tree lean, you can try to increase or decrease the
other class weights so that the algorithm penalizes samples of one class relative to the other. The simplest way is to specify

Note that this isn                 
change, it        
node, and it will change to
Weight * (the number of samples from a class in the node) / (size of class)
j. Random Forest
It is a supervised learning method which is ensemble from decision trees, use bagging algorithm to train and use data to
improve the results of accuracy
k. C4.5
Itof decision tree algorithm that uses the concept of information entropy in order to train itself.
l. ID3
ID3 uses a top-down greedy approach to build a decision tree i.e. to train itself in order to evaluate the data and get better
results m. Extra Trees
Extra Trees algorithm works by creating a large number of un-pruned decision trees from the training dataset. Predictions are
made by averaging the prediction of the decision trees
n. Custom Accuracy
The Custom accuracy showed that random forest showed better result as compared to others. As Custom accuracy is the
difference between actual value and predicted value and R- mean value is the proportion of the variance for a dependent
variable that is explained by an independent variable or variables in a the model. For the best customer accuracy, random
forest has been selected for the predictive model of T20 World Cup.
o. Performance Parameters
Performance parameters such as Custom accuracy and R- mean Square values are calculated for authentication of proposed
techniques. As the custom accuracy is the difference between Actual and predicted values whereas R- mean square is a
statistical measure that represents the proportion of the variance for a dependent variable that's explained by an
independent variable or variables in a regression model given below.
Custom Accuracy= Actual Value  Predicted value
(3.1)
R2=1-
(3.2)
Table 1 Performance Parameters
F1-Score 2* (Recall * Precision) / (Recall +
Precision)
Recall TP/TP+FN
Precision TP/TP+FP

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
Loss
1
Function

n
=0
Loss function used for finding MSE
4. Results
Displays the comparative performance of Machine learning algorithms. Two simulation parameters (Custom Accuracy and R-
Mean score) are used to evaluate the effectiveness of each algorithm. Random forest algorithm proved to be the best among
four selected machine learning algorithm.
Comparison
100
90
80.86
79.67
74.69
79.73
80
64.68
62.46
70
60.8
60
44.2
50
40
30
20
10
0
Random
Extra Trees ID3 (DT)
C4.5 (DT)
Forests
Custom Accuracy
R-Mean Score
Linear (Custom Accuracy)
Figure 2 Comparison between ExtraTrees, ID3, C4.5 and Random Forests. Results of all trees without any hyper-tuning
Random forest algorithm proves better with both criteria and has achieved 8086% and 64.68% respectively.
a. Results with Hyper-tuned Parameters
For improving our results we have also applied hyper-tuning. We have hyper-tuned the parameters in order to get optimized
results. Following parameters are hyper-tuned to optimize the results. For improving our results we have also applied
Hypertuning. We have hyper tuned the parameters in order to get optimized results. Following parameters are hyper tuned
to optimize the results. As we already know our model is not over fitted so we will not imply those features, which are used
during over fitting. We will change the criterion of all decision trees and then check whether the improved or not. We have
used three criterion for checking the model           tropy Index and
Random Criteria for checking whether the impurity exist in the nodes or not. Gini Index calculates the probabilities of having
an impurity in a node by showing the value between 0 and 1, whether the node has multiple classes or in a complexity or not.
Every time when we try to find impurity and improve the results the accuracy of each of the tree changes with better results.
Entropy index also checks the change and entropy of each of the node vertically. And if both of the criterion are not selected
then impurity does not improves if it exists.
1. Criteria
2. Minimum Impurity Decrease
3. Class Weights
As we already know our model is not over-fitted so we will not imply those features, which are used during over fitting.
i. Criterion
We will change the criterion of all decision trees and then check whether the improved or not. There are two hyper -tuned
criterion which we called as "Gini" and "Entropy". Following equations calculated the Gini Index Gain and ii. Entropy Index
Gain:
We have used three criterion for checking the model  
Random Criteria (shown in Figure 7) for checking whether the impurity exist in the nodes or not. Gini Index calculates the
probabilities of having an impurity in a node by showing the value between 0 and 1, whether the node has multiple classes or in a
complexity or not. Every time when we try to find impurity and improve the results the accuracy of each of the tree changeswith

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
better results. Entropy index also checks the change and entropy of each of the node vertically. And if both of the criterion are not
selected then impurity does not improves if it exists. Figure 7 shows the results without hyper-tuning of the criterion parameter.
While Figure 8 and Figure 9 shows the hyper-tuning of the parameters by replacing the parameter of criteria from 
and Entropy respectively, which has improved the impurities of the nodes and improves the accuracy as well.
Gini Index
100
86.5
90
79.9
81.55
76.3
80
62.3
70
61.7
55.76
61.44
60
50
40
30
20
10
0
Random
Extra Trees ID3 (DT)
C4.5 (DT)
Forests
Custom Accuracy
R-Mean Score
Linear (Custom Accuracy)
Figure 35.2. Calculation of Gini Index after Hypertuning
Entropy Index
100
89
90
77
79
80
73
70
63
59
65
55
60
50
40
30
20
10
0
Random Extra Trees ID3 (DT) C4.5 (DT)
Forests
Custom Accuracy
R-Mean Score
Linear (Custom Accuracy)
Figure 4 Calculation of Entropy Index after Hypertuning
iii. Minimum Impurity Decrease
This feature helps us to decrease the impurity of a node by splitting the node of a value greater or equal than a threshold which
induces a decrease in the impurity of a node. Following is the equation to calculate the minimum impurity decrease in a node.
Nt
NtL


N
Nt
Where N is the total number of samples, Nt is the number of samples at the current node, NtL is the number of samples in
the left child, and NtR is the number of samples in the right child.
N, Nt, NtR and NtL all refer to the weighted sum, if sample-weight is passed.
Now the question is why to use this hyper-tuned parameter. We have already used Gini and Entropy, if the index of Gini is 0 it
means the impurity is almost 0, while in our case it is not zero. So we have used this parameter to tend this impurity to zero by
ize of class

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
increasing the integer values in the argument index of minimum impurity decrease. Following are the results Figure 10 and
Figure 11 shows results after tuning minimum impurity decrease:
Gini Index with Minimum Impurity
Decrease
100
92
80
75
81
80
62
64
54
60
45
40
20
0
Random
Extra Trees ID3 (DT)
C4.5 (DT)
Forests
Custom Accuracy
R-Mean Score
Linear (Custom Accuracy)
Figure 5 Calculation of Gini Index after Hypertuning with Minimum impurity Decrease
Entropy Index with Minimum
Impurity Decrease
100
91.44
86
90
78
78
80
64
64
70
60
56
60
50
40
30
20
10
0
Random
Extra Trees
ID3 (DT)C4.5 (DT)
Forests
Custom Accuracy
R-Mean Score
Linear (Custom Accuracy)
Figure 6 Calculation of Entropy Index after Hypertuning with Minimum impurity Decrease
iv. Class Weights
Distribution of Output class depends upon the class weights, so it is an important feature for hyper -tuning. The purpose of
using this parameter is to adjust the class weights after tuning the output classes so that the output clearly shows which node
and class of a decision tree leans. Figure 12 and Figure 13 shows the improvement in the results by hyper-tuning the
parameter of class weights. Following equation is used for calculating class weights:
Class weights = Weight x (the number of samples from a class in the node)

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
Class Weights and Minimum
Impurity Decrease for Gini Index
100
94
82.3
87
90
81
80
65
62
65
70
57
60
50
40
30
20
10
0
Random
Extra Trees
ID3 (DT) C4.5 (DT)
Forests
Custom Accuracy
R-Mean Score
Linear (Custom Accuracy)
Figure 7 Calculation of Gini Index after Hypertuning with Minimum impurity Decrease adjusting Class Weights
Class Weights and Minimum
Impurity Decrease for Entropy Index
100
93
87
90
80
80
80
65
65
70
61
60
56
50
40
30
20
10
0
Random Extra Trees ID3 (DT) C4.5 (DT)
Forests
Custom Accuracy
R-Mean Score
Linear (Custom Accuracy)
Figure 8 Calculation of Entropy Index after Hypertuning with Minimum impurity Decrease adjusting Class Weights
After selecting these hyper tuned parameters for random forests and decision trees we have evaluated that random forests
are the best to predict the match winning team so we have applied random forest on datasets to predict usability features,
random forests gives the best accuracy of 92% with hypertuned parameters.
5. Conclusions
The purpose of this article is to highlight usability difficulties in education apps in Pakistani educational institutions. It aims to
make it easier to adapt digital technologies in education. This work is focused on quantitative and quality analysis and we will
take interviews and questionnaires in this study to determine the mental perception of users within the framework of the
designer. This paper identifies the main aspects for the usability of these applications and will also offer the design heuristics

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
for the design of such applications. Considering the impression of the user (student), Efficiency has been analysed as the
highest usability attribute to satisfy the student. After this we employed Genetic Algorithm (GA) to evaluate the usability
features. Each usability feature is measured by the weight of each usability feature. We have determined the rank of main
functions utilised in the Genetic Algorithm questionnaire and benchmark. After that, GA has been used to choose features, GA
have obtained the best features and we have also projected the usability characteristics of students based on selected features
based on the best features. GA based RF model with hypertuned parameters has considered the best algorithmic approach in
predicting the best characteristics with 92% accuracy.
References
[1] A. Oztekin, D. Delen, A. Turkyilmaz, and S. Zaim, -based usability evaluation method for eLearning systems,
Decis. Support Syst., vol. 56, no. 1, pp. 6373, 2013, doi: 10.1016/j.dss.2013.05.003.
[2]  10th
Int. Conf. eLearning (eLearning-2019), 26-27 Sept., no. September, pp. 3944, 2019.
[3] N. Ghatasheh, nt in e-Learning Systems Using Machine Learning and User Activity Analysis, Int. J. Adv.
Comput. Sci. Appl., vol. 6, no. 4, 2015, doi: 10.14569/ijacsa.2015.060415.
[4] A. Moubayed, M. Injadat, A. B. Nassif, H. Lutfiyya, and A. Shami, -Learning: Challenges and Research Opportunities Using
Machine Learning Data Analytics, IEEE Access, vol. 6, pp. 3911739138, 2018, doi: 10.1109/ACCESS.2018.2851790.
[5] S. S. Khanal, P. W. C. Prasad, A. Alsadoon, and A. Maag, commendation systems
for e-learning, Educ. Inf. Technol., vol. 25, no. 4, pp. 26352664, 2020, doi: 10.1007/s10639-019-10063-9.
[6] S. B. Aher and L. M. R. J. Lobo, algorithms for recommendation of courses in E-Learning System
based on historical data, Knowledge-Based Syst., vol. 51, pp. 114, 2013, doi: 10.1016/j.knosys.2013.04.015.
[7] Q. Deng and Z. Wu, arning, IOP Conf. Ser. Earth Environ. Sci., vol.
199, no. 3, 2018, doi: 10.1088/1755-1315/199/3/032042.
[8] D. N. Lu, H. Q. Le, and T. H. Vu, -learning: A machine learning algorithm approach, Educ.
Sci., vol. 10, no. 10, pp. 113, 2020, doi: 10.3390/educsci10100270.
[9] C. J. Huang, S. S. Chu, and C. T. Guan, 
intelligent e-learning systems, Comput. Educ., vol. 49, no. 3, pp. 597614, 2007, doi: 10.1016/j.compedu.2005.11.008.
[10] I. Lykourentzou, I. Giannoukos, V. Nikolopoulos, G. Mpardis, and V. Loumos, -learning courses through the
combination of machine learning techniques, Comput. Educ., vol. 53, no. 3, pp. 950965, 2009, doi:
10.1016/j.compedu.2009.05.010.
[11] T. Mahboob, S. Irfan, and A. Karamat, -learning using Quinlan
Bayes and Random Forest algorithms, Proc. 2016 19th Int. Multi-Topic Conf. INMIC 2016, 2017, doi: 10.1109/INMIC.2016.7840094.
[12] T. Purwoningsih, H. B. Santoso, and Z. A. Hasibuan,         
Machine Learning Approach, Proc. 2019 4th Int. Conf. Informatics Comput. ICIC 2019, no. Icic, pp. 18, 2019, doi:
10.1109/ICIC47613.2019.8985918.
[13] A. Borakati, -learning course with natural language processing and machine learning,
BMC Med. Educ., vol. 21, no. 1, pp. 110, 2021, doi: 10.1186/s12909-021-02609-8.
[14] S. Singh and S. P. Lal,  2013 IEEE Conf. e-Learning, e-
Management e-Services, IC3e 2013, pp. 7378, 2013, doi: 10.1109/IC3e.2013.6735969.
[15] M. Tan and P. Shao, on of student dropout in E-learning program through the use of machine learning method, Int. J.
Emerg. Technol. Learn., vol. 10, no. 1, pp. 1117, 2015, doi: 10.3991/ijet.v10i1.4189.
[16] A. Hamdi-cherif,  Learning Agents and UML for Elearning Settings, vol. 2, no. 1, 2008.
[17] M. M. Iqbal, M. Farhan, Y. Saleem, and M. Aslam, -Bot Implementation using Machine Learning Techniques in
eLearning Paradigm, 11th Mex. Int. Conf. Artif. Intell., no. January, 2012.
[18] M. Krendzelak, ne learning and its applications in e-learning systems, ICETA 2014 - 12th IEEE Int. Conf. Emerg. eLearning
Technol. Appl. Proc., pp. 267269, 2015, doi: 10.1109/ICETA.2014.7107596.
[19] T. Hall and H. Kamper, c learning using machine learning, 2020 Int.
SAUPEC/RobMech/PRASA Conf. SAUPEC/RobMech/PRASA 2020, 2020, doi: 10.1109/SAUPEC/RobMech/PRASA48453.2020.9041140.
[20] N. Yan and O. T.-S. Au, g, Asian Assoc. Open Univ. J., vol. 14, no. 2, pp.
97106, 2019, doi: 10.1108/aaouj-08-2019-0029.
[21] M. Hussain, W. Zhu, W. Zhang, J. Ni, Z. U. Khan, and S. Hussain, -learning System Using
Machine Learning Techniques, 2018 IEEE Conf. Big Data Anal. ICBDA 2018, pp. 123128, 2019, doi: 10.1109/ICBDAA.2018.8629697.
[22] S. Siddiq,   -Learning Computer Science 4490Z Final Report Machine Learning in E-Learning Department of
Computer Science London , Ontari
May, pp. 039, 2017, doi: 10.13140/RG.2.2.36213.37605.
[23] W. Villegas-Ch, M. Román-Cañizares, and X. Palacios-Pacheco,  with the integration of
machine learning and data analysis in an LMS, Appl. Sci., vol. 10, no. 15, 2020, doi: 10.3390/APP10155371.
[24] R. Muralidharan, T. Kanagasabapathy, and R. P. Vijai Ganesh, l
Evaluation with SVM Classifier, J. Phys. Conf. Ser., vol. 1767, no. 1, 2021, doi: 10.1088/1742-6596/1767/1/012012.
[25] M. Krendzelak, -learning systems, ICETA 2014 - 12th IEEE Int. Conf. Emerg. eLearning
Technol. Appl. Proc., no. August, pp. 267269, 2015, doi: 10.1109/ICETA.2014.7107596.

International Journal of Software Engineering and Knowledge Engineering
(0218-1940)
[26] N. Bhaliya, J. Gandhi, and D. K. Singh, Int. J. Innov. Technol.
Explor. Eng., vol. 9, no. 7, pp. 1317, 2020, doi: 10.35940/ijitee.f4078.059720.
[27] M. M. RAMADHAN, I. S. SITANGGANG, F. R. NASUTION, and A. GHIFARI, 
Method for Gender Classification Based on Voice Frequency, DEStech Trans. Comput. Sci. Eng., no. cece, 2017, doi:
10.12783/dtcse/cece2017/14611.
[28] V. M. Menon and H. A. Rahulnath,        
emotional intelligence through social media data, 2016 Int. Conf. Next Gener. Intell. Syst. ICNGIS 2016, 2017, doi:
10.1109/ICNGIS.2016.7854061.
[29] H. Shaikh, F. H. Jaskani, F. Khuhawar, K. Khan, M. A. Azam, and M. Aqeel, 
using Machine Learning Techniques,i: 10.13140/RG.2.2.21875.35366.
[30] Y. V. Mane and A. R. Surve, Proc. 2016 Int. Conf.
Adv. Commun. Control Comput. Technol. ICACCCT 2016, no. 978, pp. 661665, 2017, doi: 10.1109/ICACCCT.2016.7831722.
[31] A. P. J. Joseph, Int. J. Res. Appl. Sci. Eng. Technol., vol. 8, no. 7, pp.
896899, 2020, doi: 10.22214/ijraset.2020.30378.
[32] M. C. Sorkun, Algilayicilarin Ve Pencere BoyunuEtk s 
and Window Size,2018 26th Signal Process. Commun. Appl. Conf. (SIU), Izmir, pp. 14, 2018, doi: 10.1109/SIU.2018.8404569.
[33] C. Science and M. Studies, ion Methods and Classifiers,115, 2015.
[34] C. S. Abhishek, K. V Patil, P. Yuktha, K. S. Meghana, and M. V Sudhamani, 
Learning Techniques,435, 2019, doi: 10.35940/ijitee.B1043.1292S19.
[35] R. G. Mantovani, A. L. D. Rossi, J. Vanschoren, B. Bischl, and A. C. P. L. F. De Carvalho, 
SVM hyper-parameter tuning, Proc. Int. Jt. Conf. Neural Networks, vol. 2015-Septe, 2015, doi: 10.1109/IJCNN.2015.7280664.
[36] I. Syarif, A. Prugel-Bennett, and G. Wills,           
classification performance,Telkomnika (Telecommunication Comput. Electron. Control., vol. 14, no. 4, pp. 15021509, 2016, doi:
10.12928/TELKOMNIKA.v14i4.3956.
[37] R. Schaer, H. Müller, and A. Depeursinge, 
classification in CT using Hadoop,J. Imaging, vol. 2, no. 2, 2016, doi: 10.3390/jimaging2020019.
[38] A. Unnikrishnan, F. Ajesh, and Reshma S. Nair, 
31963210, 2015.
[39] M. Jethanandani, T. Perumal, and A. Sharma, -Labelsets method for human activity recognition with multi-sensor data
in smart home, 2019 IEEE 16th India Counc. Int. Conf. INDICON 2019 - Symp. Proc., 2019, doi:
10.1109/INDICON47234.2019.9030296.
[40] N. Golestani and M. Moghaddam, -based Human Activity Recognition ( MI-HAR ),o. Mi, pp. 1718, 2019.
12