PERJ Vol. 5 No.

2 (2023), 97-106

Predicting Physics Students’ Achievement Using In-Class Assessment Data: A

Comparison of Two Machine Learning Models
Purwoko Haryadi Santoso1,2*, Hayang Sugeng Santosa3, Edi Istiyono1,4, Haryanto1, Heri Retnawati1
1Graduate School of Educational Research and Evaluation, Universitas Negeri Yogyakarta, Indonesia
2Department of Physics Education, Universitas Sulawesi Barat, Indonesia
3Graduate School of Educational Research and Evaluation, Universitas Muhammadiyah Prof. Dr. HAMKA, Indonesia
4Department of Physics Education, Universitas Negeri Yogyakarta, Indonesia

ARTICLE INFO ABSTRACT

Article history: Data is the primary source to scaffold physics teaching and learning for teachers and
Submitted : February 27th, 2023 students, mainly reported through in-class assessment. Machine learning (ML) is an
Revised : May 19th, 2023 axis of artificial intelligence (AI) study that immensely attracts the development of
Accepted : July 11th, 2023 physics education research (PER). ML is built to predict students’ learning that can
support students’ success in an effective physics achievement. In this paper, two ML
algorithms, logistic regression and random forest, were trained and compared to
Keywords: predict students’ achievement in high school physics (N = 197). Data on students’
Achievement Prediction; Assessment; Machine achievement was harvested from in-class assessments administered by a physics
Learning; Physics Education; Students teacher regarding knowledge (cognitive) and psychomotor during the 2020/2021
academic year. Three assessment points of knowledge and psychomotor were
employed to predict students’ achievement on a dichotomous scale on the final term
examination. Combining in-class assessment of knowledge and psychomotor, we
could discover the plausible performance of students’ achievement prediction using
the two algorithms. Knowledge assessment was a determinant in predicting high
school physics students’ achievement. Findings reported by this paper recommended
open room for the implementation of ML for educational practice and its potential
contribution to supporting physics teaching and learning.

COPYRIGHT (C) 2023 PHYSICS EDUCATION RESEARCH JOURNAL

Introduction in the educational process. EDM and LA were born

when data was a potential resource to be addressed in
Since the emergence of the integration of information the current era.
and communication technology in physics learning,
educational institutions and academic processes have Regarding the functions of EDM and LA, artificial
been sources of complex and large digital data; not intelligence (AI) technology is the forerunner to the
only data collected through the class assessment birth of these two research groups (Fynn et al., 2022).
process but learning channeled through the learning AI studies have created predictive technology
management system is one channel for collecting processed through machine learning (ML) for several
learning data that can be produced at any time and is purposes, including education (Leitner et al., 2017).
large (Santoso et al., 2022). The emergence of data of One of the ML algorithms' analytical tasks is analyzing
this type and large size can be used to support learning, state predictions beyond model training ((Shafiq et al.,
as attempted by the Educational Data Mining (EDM) 2022; Albreiki et al., 2021). ML algorithms can learn
and Learning Analytics (LA) research groups (Santoso patterns from extensive training data to provide
& Munawanto, 2020). These two groups were created predictions to support learning decisions (Romero &
so that we can be more aware of the existence of data Ventura, 2020). Although the application of ML in

*Correspondence email: [email protected]

doi: 10.21580/perj.2023.5.2.14127 97
 Phy. Educ. Res. J. Vol. 5 No. 2 (2023), 97-106

physics education is still relatively new, several studies Two divisions of ML learning types are most widely
predicting physics learning achievement have been cited in the literature, namely unsupervised and
carried out since 2019 (Aikenhead, 2023; Lin et al., supervised (Susilawati et al., 2021)—the difference
2023). The results of the ML study report some between the two lies in the labels or targets available in
success in supporting physics learning, but further the training dataset. Unsupervised algorithms are more
studies still need to be carried out. exploratory because interpretation of the results still
requires the role of the model user. An example of an
Most ML models built by physics education unsupervised ML algorithm is clustering. In contrast,
researchers are random forests and logistic regression supervised ML methods have predetermined targets
as tested. Previous researchers could predict student before model training. ML is trained to predict one of
learning achievement with an accuracy of up to 80% the labels contained in the target variable. Several tasks
using these two algorithms. Teachers can use this are included in the supervised learning type (Bloor &
information as feedback that can be given to students Santini, 2023). One of them that is often applied in
to support student learning (Atmam & Mufit, 2023; education is classification. The topic of classification,
Ndoa & Anastasia, 2022). According to the namely the prediction of student learning, is the basis
constructivist approach, effective physics learning for the birth of the field of ML studies in the field of
provides students with independent learning, which is education, namely educational data mining (EDM) and
supported by the teacher's activeness in monitoring learning analytics (LA). Even though they have slightly
learning, which the results of ML predictions can different terminology, there is a common goal of
generate. supporting student learning.
The research presented in this article aims to explore Several ML algorithms are often used to predict
assessment data collected by teachers to predict student learning achievement. A systematic review of
student physics learning achievement. The data is the history of algorithms that have been used has been
processed to train an ML model based on logistic synthesized, such as decision trees, random forests, k-
regression and random forest algorithms, as has been nearest neighbors, support vector machines, naive
done by many physics education researchers above. Bayes, logistic regression, and artificial neural
The prediction performance produced in the two ML networks (Sekeroglu et al., 2021). Due to the
models was then evaluated and compared to the most limitations of the study focus addressed in this article,
optimal one. only two ML algorithms as the basis for developing
To guide the investigation, three research questions ML models were used in predicting learning
were posed as follows. achievement reported in this article.
Problem 1. How is the prediction performance of
physics learning outcomes produced by the logistic Logistic Regression
regression model? Although its name is mentioned in the term
Problem 2. How is the prediction performance of ‘regression,’ which is another task in the supervised
physics learning outcomes produced by the random learning type; logistic regression is mainly used for
forest model? tasks. This algorithm has the same properties as linear
Problem 3. What does data contribute most to regression. The difference lies in using the dependent
predicting physics learning outcomes? variable in categorical form, where the sigmoid
One of the implications of this article aims to open a function is used to classify the class of the dependent
discussion space in physics education to support variable.
physics learning through ML predictive technology.
Random Forest
Machine Learning This model is one of a family of decision tree
Machine learning (ML) artificial intelligence (AI) is a algorithms for carrying out classification tasks on
branch that focuses on utilizing data and algorithms to categorical data. Random forest is a model from a
imitate how a knowledge machine constantly improves group of tree-based algorithms that grows several trees
its capabilities. ML differs from statistical approaches (decision trees) to determine classification decisions
that rely heavily on probability theory for hypothesis for a class of targets as an ensemble. This method is
testing. The terms often used in ML studies are also known as optimization of decision trees by
training and testing that do not aim to generalize to a considering the regression average of trees that grow
population. However, ML is trained to make as an ensemble.
predictions from input data not used in model training
(Aikenhead, 2023).

98 https://ejournal.walisongo.ac.id/index.php/perj/index
 Phy. Educ. Res. J. Vol. 5 No. 2 (2023), 97-106

Methods 3. Data preparation.

Based on data from physics teachers, 197
Research Design students are enrolled in physics learning. Students
are divided into five classes X from the science
This research explores training two types of ML department. However, there are cases where
models using the framework suggested by The Cross students do not take part in the assessments
Interprofessional Standard Data Mining Process carried out by the teacher even though the
(CRISP-DM) in data science research (Chapman, teacher has carried out additional assessment
2000). Six stages structure this framework, illustrated sessions for the students. However, there are still
in Figure 1. students who cannot join for several reasons. The
Figure 1 teacher agreed that a grade of zero should be
Work Cycle of Data Science Research According To CRISP-DM given in that case. Data that is ready to be saved
in CSV format.
4. Modeling.
The next stage is training the ML model through
logistic regression and random forest algorithms.
The proportion of sharing 75% of training data
and 25% of test data was chosen in this research.
Modeling was carried out in R using several
packages such as 'caret' (Hochberg et al., 2018),
'randomForest,' 'lme4' (Bates et al., 2015),
'caTools', 'pROC' (Starr et al., 2020), 'varSelRF',
'pscl', 'MASS', as well as 'ggplot2', 'ggthemes',
'RColorBrewer', 'Rmisc', and 'car' for data
visualization purposes.
5. Evaluation.
The ML model that has been trained is then used
to make predictions on the test data at this stage.
The prediction performance of the ML model on
1. Business understanding. the test data is then tabulated using the confusion
The purpose of applying ML in this research is to matrix shown in Table 1. Several evaluation
perform classification tasks. Classification tasks metrics are used in evaluating ML classification
are one of the most researched topics by EDM models based on the number calculated in the
and LA groups (Sekeroglu et al., 2021). Prediction confusion matrix. Several evaluation metrics that
of students' physics learning achievement is an are often used in ML studies predicting learning
example of a classification task in ML research. outcomes include accuracy, precision, sensitivity,
specificity, F1-measure, and area under the curve
2. Data understanding. (AUC) (Darling-Hammond et al., 2020), which
Two types of data obtained through classroom are calculated based on the formula described in
assessments are used to predict physics learning Equation 1 to Equation 6.
outcomes—namely, teacher assessment data on
aspects of knowledge and skills by the Indonesian Table 1
Confusion Matrix
physics curriculum syllabus. The class assessment
data in this research comes from recorded Actual+ Actual -
academic data obtained by a teacher from one of Prediction + True positive (TP) False positive (FP)
the state schools in Indonesia in the 2020/2021 Prediction - False negative True negative
academic year. Student assessments are measured (FN) (TN)
using assessment instruments developed and
validated by teachers based on the class X high TP+TN
Accuracy = (1)
TP+TN+FP+FN
school physics curriculum reference by the
current COVID-19 pandemic conditions. There Precision =
TP
(2)
are three essential physics competencies that TP+FP
teachers focus on during this period, including Sensitivity =
TP
(3)
the nature of physics, measurement, and vectors. TP+FN

https://ejournal.walisongo.ac.id/index.php/perj/index 99
 Phy. Educ. Res. J. Vol. 5 No. 2 (2023), 97-106

Specificity =
TN
(4) Then, another fifth model was created to follow the
TN+FP development of student learning in each basic physics
F1-measure = 2 ∙
𝑝𝑟𝑒𝑠𝑖𝑠𝑖∙𝑟𝑒𝑐𝑎𝑙𝑙
(5) competency. The last two models were created to
𝑝𝑟𝑒𝑠𝑖𝑠𝑖+𝑟𝑒𝑐𝑎𝑙𝑙 measure the contribution of knowledge and skills
aspects in predicting physics learning achievement
Then, the 'stepAIC' function measures how with two ML algorithms.
important a variable (knowledge or skills) is in the First, model accuracy increases as basic competency
logistic regression model and 'varImpPlot' for the (BC) variables increase, reaching the highest peak at
random forest model. basic competency 3. However, model performance
6. Deployment. based on just one measure of accuracy can cause
In this research, this stage was not carried out several problems (Zabriskie et al., 2019). Table 1 above
because the study development process was still is the confusion matrix produced when testing the ML
being carried out. model using test data. This contingency table will then
calculate the six model evaluation metrics reported in
Table 2.
Result and Discussions Second, the accuracy aspect must also be
This research aims to train a prediction model for complemented by a measure of precision, namely the
physics learning achievement with two algorithms: extent to which predictions have been successfully
logistic regression (Problem 1) and random forest made from the total TP and FP predictions in the
(Problem 2). A comparison of prediction performance confusion matrix, as explained in equation (2) above.
between the two ML methods based on six evaluation In contrast to the accuracy defined in measurement
metrics (accuracy, precision, sensitivity, specificity, F1- principles in physics, the precision described in
measure, and AUC) is reported in Table 2. equation (2) is explicitly used to see the proportion of
accurate positive predictions (TP) to false optimistic
Six models have been drilled using logistic regression predictions (FP). The larger the FP results, the smaller
and random forest algorithms. The first step is the precision. The baseline model did not find
establishing a baseline where the target variable precision values because TP and FP were not found in
(student learning achievement) is predicted as "1" or this group. The model precision value in Table 2
more from the minimum completeness during the increases as the variables involved in the model
final test. This model can be used as a comparison of increase. This result indicates good results because it
the fifth model that was explored later. This baseline corresponds to the most significant increase in
model detects the extent of model accuracy in the case accuracy values at BC 3.
of pure guessing. Both logistic regression and random
forest show the same results in the first model.
Table 2
Performance of Predicting Physics Learning Achievement Using Logistic Regression and Random Forest
Logistic Regression Random Forest
Model
Acc Prec Sens Spec F1 AUC Acc Prec Sens Spec F1 AUC
Baseline 0.714 - 0.000 1.000 - 0.500 0.714 - 0.000 1.000 - 0.500
BC 1 0.735 1.000 0.071 1.000 0.133 0.727 0.714 0.500 0.071 0.971 0.125 0.498
BC 2 0.796 0.833 0.357 0.971 0.500 0.862 0.816 0.778 0.500 0.943 0.609 0.872
BC 3 0.857 0.769 0.714 0.914 0.741 0.951 0.878 0.786 0.786 0.914 0.786 0.945
K only 0.857 0.769 0.714 0.914 0.741 0.951 0.878 0.786 0.786 0.914 0.786 0.947
S only 0.878 0.900 0.643 0.971 0.750 0.876 0.816 0.692 0.643 0.886 0.667 0.909
Note: Acc : accuracy, Prec : precision, Sens : sensitivity, Spec : specificity, F1 : F1-measure, AUC : area under curve, BC : basic
competence, K : Knowledge, S : Skills
The TP and TN cases are what we expect in studies Le et al., 2022; Lu et al., 2021). However, cases of FP
predicting student learning achievement. An accurate and FN may occur, and these two results are not
prediction means there is a match between what is expected to predict learning outcomes. Both cases can
predicted and what happens. The good or bad reduce students' learning motivation or lead to
performance predicted by the teacher during the excessive self-confidence (Rubie-Davies, 2006;
learning process will correlate with student learning Shengnan et al., 2018). Therefore, considering only the
achievement at the end of learning (Chen et al., 2018; TP and TN numbers as evaluated through the above

100 https://ejournal.walisongo.ac.id/index.php/perj/index
 Phy. Educ. Res. J. Vol. 5 No. 2 (2023), 97-106

accuracy and precision metrics can be problematic in performance relevant to the three previous measures:
their interpretation. We need to review other numbers accuracy, precision, and sensitivity.
that can understand the extent to which the model
produces FP and FN results in predicting student Fifth, we will use the proposed F1 measure to improve
physics learning achievement, namely through the ML classification performance measures described
sensitivity and specificity measures. by the four previously used measures. F1 measure can
overcome uneven class distribution in the training
Motivation learning can be inaccurate, and prediction dataset (Luo et al., 2021; Yang et al., 2017). For
reports influence them. For example, the FN results in example, the predicted learning outcomes of students
this case cause students who are high performers (+) categorized as above the minimum completeness value
to be predicted as students with poor learning or "1" are much higher than those classified as "0".
achievement (-). Then, excessive self-confidence can Based on this measure, the BC3 model is still superior
be caused by FP prediction reports, which report that to the other two. The fifth measure offered by the F1
students who perform poorly (-) are instead predicted measure further strengthens the argument provided by
to have exemplary learning achievements (+) as has the four previous evaluation metrics that the more
been found (Jeong et al., 2021; Vasalou et al., 2021), formative assessment variables are implemented, the
interpreting the metrics involving the number of FN more the prediction performance provided by ML will
and FP above must be adjusted to the context of improve.
predicting learning outcomes aimed in this research.
Lastly is the area under curve (AUC), calculated via
Nevertheless, researchers argue that if FP creates receiver operating characteristic (ROC). As a general
higher student self-confidence, it should have a better rule, the higher the AUC, the more the ML prediction
impact on the physics learning process than the model can differentiate the class of each target. This
negative effect caused by the FN case. This argument means that the prediction model is increasingly
can be supported by the self-regulated learner (SRL) accurate in predicting “1” as “1” or “0” as “0”. A good
theory proposed by (Kind, 2013; Zimmerman et al., ML model has an AUC close to 1. Based on the results
2014). Positive feedback teachers give can increase shown in Table 2, the BC 3 (vector material) model is
students' self-confidence, so it impacts students' the best model according to AUC.
learning independence according to the SRL
theoretical framework. Therefore, the case of Based on the six measures reported by the evaluation
predicting physics learning achievement should place metrics above, it can be concluded that the ML
more emphasis on prediction performance, which can prediction model, which involves all formative
minimize cases of FN, which can reduce students' assessments up to BC 3, is the most optimal in
confidence in learning physics. Sensitivity or recall predicting student physics learning outcomes for both
measures are more suitable to pay attention to in cases the logistic regression and random forest algorithms.
where we want to minimize FN cases in our prediction Researchers found no significant difference in the
results. The sensitivity value will be more significant prediction performance results reported by logistic
when the FN value is smaller, or what we expect in regression and random forest.
Equation 3. Thus, reviewing a more considerable These results are from previous research they are
sensitivity value can be an option in evaluating the ducted by (Semenikhina et al., 2020). The five
model with the most optimal prediction performance assessment points found that they could produce
to predict student learning achievement. prediction models with up to 80% accuracy with just
Third, BC 3 in Table 2 shows satisfactory prediction three formative assessment data. These results have
performance based on sensitivity. Using the skill implications for learning purposes in that the
aspect can have a sensitivity of 90% in the logistic predictive information provided by ML can become
regression algorithm. These results exceed those found input or feedback that students can immediately use to
by Zabriskie et al. (2019), whose model only achieved improve learning. Simultaneously, teachers can also
an accuracy level of 80% using only five variable involve these results as a reflection of the learning they
assessment points. carry out in improving the effectiveness of physics
learning. Through ML support, physics classes can be
Fourth, specificity measures should be expected to be accepted by students more effectively. However,
the opposite of those found by sensitivity. The focus experimental studies must be conducted to test this
on predicting learning outcomes is more aimed at conjecture.
minimizing expectation errors, which reduces student
learning motivation in the case of FN. Based on Table Apart from looking at the performance of ML
2 above, we can see that BC 3 is a model with good prediction results between logistic regression and

https://ejournal.walisongo.ac.id/index.php/perj/index 101
 Phy. Educ. Res. J. Vol. 5 No. 2 (2023), 97-106

random forest, this research also investigates what aspects. These three variables are displayed in Table 3
factors contribute most to the optimal prediction below, which displays the coefficients from the logistic
results. The last two models were deliberately created regression equation, which has been simplified using
to look at each aspect of the assessment carried out by the 'stepAIC' method. The explanation of the variable
teachers, namely knowledge and skills. Based on the code 'P.PH3.1' is that the leftmost letter 'P' represents
six evaluation metrics, the performance of the logistic the knowledge aspect, then 'PH' means the method
regression and random forest models did not show used, namely daily assessment, and '3.1' is the basic
significant differences when predicting student physics competency code that is being assessed by the
learning outcomes using aspects of knowledge or teacher, namely the vector by Indonesian physics
knowledge alone. Except that AUC reported a higher learning class X curriculum. We can see that all aspects
difference in logistic regression performance for the of knowledge are reported from the three variables
knowledge aspect than for the skill aspect. Even by summarized by 'stepAIC'. These results can confirm
using the features indicated by the knowledge aspect and strengthen what was reported by the previous
alone, the ML model can show performance in AUC measure that the knowledge aspect is a
predicting learning outcomes with performance that is determining factor in predicting physics learning
close to the AUC value shown by the BC 3 model. This achievement in this study. In addition, basic vector
initial information illustrates that the knowledge aspect competence is the most critical aspect in this case.
still significantly contributes to the success of studying
Table 3
high school physics found in this research.
Estimation Results of KD3 Logistic Regression Model Coefficient
This finding may be challenged by several opinions of Parameters in Final 'stepAIC'
previous researchers who suggested the importance of Intercept P.PH3.1 P.TG3.1 P.TG2.1
aspects of science process skills in physics learning, -11.1255 0.07216 0.07039 0.02768
including what was found (Lin et al., 2023; Yang et al.,
2017), which still uses non-cognitive variables in its As a reinforcement of the AUC and stepAIC findings
physics learning outcomes prediction model. above, the 'varImpPlot' function is used in Figure 2
However, this problem is outside the research above to extract the most essential variables in the
question, which is the aim of this research. Critical model by visualizing the decrease in the mean of the
studies with more comprehensive data are highly accuracy metric and Gini index. Both relative report
recommended to open up space for scientific results that are not that different from the visualization
discussion in the physics learning assessment field. of a horizontal bar chart arranged from the most
The 'stepAIC' function is a method that is often used critical variables to the least important. Based on
to select features in classification models. By what has previous findings, three knowledge variables were
been described in the method above, this function has accurately reported as having the most significant
been used to investigate the contribution of each contribution to model 3. Both AUC, stepAIC, and
feature involved in the model. In this presentation, feature importance from random forest concluded
'stepAIC' is applied to the BC 3 model because we that these three aspects were determining factors in
have found that this model shows the most optimal this study's predicted results of physics learning
performance in predicting physics learning achievement.
achievement according to the previous discussion. The research results reported in this study have built
However, it should be emphasized that 'stepAIC' is two ML models to predict students' physics learning
not intended to improve model fit to data or more achievement with two algorithms that achieved
optimal performance of ML prediction results. This entirely satisfactory performance. The next stage of
function is used to simplify the model without this research is further testing the ML model through
significantly affecting the prediction performance of training data with different contexts or setting several
the ML model. Therefore, AIC only measures the hyperparameters of the logistic regression and random
amount of information loss when one of the features forest models, especially for data up to KD3, which is
is omitted from the model. AIC is an abbreviation of proven to perform best according to the above
Akaike Information Criteria. findings. Apart from that, deployment has not been
carried out in this research, which should be an
The final stage of the 'stepAIC' calculation concluded inseparable part of a data science project, according to
that three variables could be simplified for the KD3 CRISP-DM. The opportunity to continue the studies
model with a lower AIC value of 76.56 from the designed in this research is still open to carry out these
previous one of 82.17 by involving all class assessment stages or test other ML models.
data including assessment of knowledge and skills

102 https://ejournal.walisongo.ac.id/index.php/perj/index
Figure 2
Feature Importance of the KD3 Model with the Random Forest Algorithm

References
Conclusions Aikenhead, G. S. (2023). Humanistic school science:
Research, policy, politics and classrooms. Science
The development of technology and information- Education, 107(2), 237–260.
based learning is a channel that can produce large and https://doi.org/https://doi.org/10.1002/sce.2
complex data. One data that physics teachers often 1774
measure is the class assessment reported to students at
the end of learning in the knowledge and skills aspects Albreiki, B., Zaki, N., & Alashwal, H. (2021). A
of the Indonesian physics learning curriculum. In this Systematic Literature Review of Student’
research, two supervised ML models, namely logistic Performance Prediction Using Machine
regression and random forest, have been trained using Learning Techniques. Education Sciences, 11(9).
academic data from physics lessons conducted by a https://doi.org/10.3390/educsci11090552
teacher. Six ML models were created by considering
the development of student learning in the three Atmam, P. L., & Mufit, F. (2023). Using Adobe Animated
essential physics competencies and investigating the CC in Designing Interactive Multimedia Based on
contribution of each assessment data used to predict Cognitive Conflict on Parabolic Motion Materials. 8(1),
the most optimal physics learning achievement. 64–74.
Logistic regression and random forest did not show Bates, D., Mächler, M., Bolker, B., & Walker, S. (2015).
significant differences in performance to predict Fitting Linear Mixed-Effects Models Using
learning achievement in physics classes. The three lme4. Journal of Statistical Software, 67(1), 1–48.
assessment points of the basic physics competency of https://doi.org/10.18637/jss.v067.i01
class X high school can be predicted optimally if there
is an assessment of the final physics competency, Bloor, T., & Santini, J. (2023). Modeling the Epistemic
namely vector. The knowledge assessment aspect is Value of Classroom Practice in the Investigation
proven to substantially influence the performance of of Effective Learning. Science & Education, 32(1),
predicting physics learning outcomes using logistic 169–197. https://doi.org/10.1007/s11191-021-
regression and random forest. The results of this 00298-9
research can recommend one method teachers can use
to design physics learning with student feedback. Chapman, P. (2000). CRISP-DM 1.0: Step-by-step data
mining guide.
https://api.semanticscholar.org/CorpusID:597
77418
 Phy. Educ. Res. J. Vol. 5 No. 2 (2023), 97-106

Chen, J., Wang, M., Grotzer, T., & Dede, C. (2018). Analytics in Higher Education---A Literature
Using a three-dimensional thinking graph to Review. In A. Peña-Ayala (Ed.), Learning
support inquiry learning. Journal of Research in Analytics: Fundaments, Applications, and Trends: A
Science Teaching, 55. View of the Current State of the Art to Enhance e-
https://doi.org/10.1002/tea.21450 Learning (pp. 1–23). Springer International
Publishing. https://doi.org/10.1007/978-3-
Darling-Hammond, L., Flook, L., Cook-Harvey, C., 319-52977-6_1
Barron, B., & Osher, D. (2020). Implications for
educational practice of the science of learning Lin, R., Yang, J., Jiang, F., & Li, J. (2023). Does
and development. Applied Developmental Science, teacher’s data literacy and digital teaching
24(2), 97–140. competence influence empowering students in
https://doi.org/10.1080/10888691.2018.15377 the classroom? Evidence from China. Education
91 and Information Technologies, 28(3), 2845–2867.
https://doi.org/10.1007/s10639-022-11274-3
Domínguez Romero, E., & Bobkina, J. (2021).
Exploring critical and visual literacy needs in Lu, K., Yang, H. H., Shi, Y., & Wang, X. (2021).
digital learning environments: The use of memes Examining the key influencing factors on college
in the EFL/ESL university classroom. Thinking students’ higher-order thinking skills in the smart
Skills and Creativity, 40, 100783. classroom environment. International Journal of
https://doi.org/https://doi.org/10.1016/j.tsc.2 Educational Technology in Higher Education, 18, 1.
020.100783 https://doi.org/10.1186/s41239-020-00238-7
Fynn, A., Adamiak, J., & Young, K. (2022). A Global Luo, M., Sun, D., Zhu, L., & Yang, Y. (2021).
South Perspective on Learning Analytics in an Evaluating scientific reasoning ability: Student
Open Distance E-learning (ODeL) Institution. performance and the interaction effects between
In P. Prinsloo, S. Slade, & M. Khalil (Eds.), grade level, gender, and academic achievement
Learning Analytics in Open and Distributed Learning: level. Thinking Skills and Creativity, 41, 100899.
Potential and Challenges (pp. 31–45). Springer https://doi.org/https://doi.org/10.1016/j.tsc.2
Nature Singapore. 021.100899
https://doi.org/10.1007/978-981-19-0786-9_3
Narvaez Rojas, C., Alomia Peñafiel, G. A., Loaiza
Hochberg, K., Kuhn, J., & Müller, A. (2018). Using Buitrago, D. F., & Tavera Romero, C. A. (2021).
Smartphones as Experimental Tools – Effects Society 5.0: A Japanese Concept for a
on Interest, Curiosity, and Learning in Physics Superintelligent Society. Sustainability, 13(12).
Education. Journal of Science Education and https://doi.org/10.3390/su13126567
Technology, 27. https://doi.org/10.1007/s10956-
018-9731-7 Ndoa, Y. A. A., & Anastasia, D. P. (2022). Development
of An Android-Based Physics E-Book with A Scientific
Jeong, J. S., González-Gómez, D., & Cañada-Cañada, Approach to Improve The Learning Outcomes of Class
F. (2021). How does a flipped classroom course X High School Students on Impulse and Momentum
affect the affective domain toward science Materials. 18(December), 107–121.
course? Interactive Learning Environments, 29(5), https://doi.org/10.15294/jpfi.v18i2.30824
707–719.
https://doi.org/10.1080/10494820.2019.16360 Romero, C., & Ventura, S. (2020). Educational data
79 mining and learning analytics: An updated
survey. WIREs Data Mining and Knowledge
Kind, P. M. (2013). Establishing Assessment Scales Discovery, 10(3), e1355.
Using a Novel Disciplinary Rationale for https://doi.org/https://doi.org/10.1002/widm
Scientific Reasoning. Journal of Research in Science .1355
Teaching, 50(5), 530–560.
Rubie-Davies, C. M. (2006). Teacher expectations and
Le, B., Lawrie, G. A., & Wang, J. T. H. (2022). Student student self-perceptions: Exploring
Self-perception on Digital Literacy in STEM relationships. Psychology in the Schools, 43(5), 537–
Blended Learning Environments. Journal of Science 552.
Education and Technology, 31(3), 303–321. https://doi.org/https://doi.org/10.1002/pits.2
https://doi.org/10.1007/s10956-022-09956-1 0169
Leitner, P., Khalil, M., & Ebner, M. (2017). Learning Santoso, P. H., Istiyono, E., & Haryanto. (2022).

104 https://ejournal.walisongo.ac.id/index.php/perj/index
 Phy. Educ. Res. J. Vol. 5 No. 2 (2023), 97-106

Physics Teachers’ Perceptions about literature over the past 30 years. Educational
Their Judgments within Differentiated Learning Research and Evaluation, 24(3–5), 124–179.
Environments: A Case for the Implementation https://doi.org/10.1080/13803611.2018.15487
of Technology. Education Sciences, 12(9). 98
https://doi.org/10.3390/educsci12090582
Starr, C., Hunter, L., Dunkin, R., Honig, S., Palomino,
Santoso, P. H., & Munawanto, N. (2020). R., & Leaper, C. (2020). Engaging in science
Approaching Electrical Circuit Understanding practices in classrooms predicts increases in
with Circuit Builder Virtual Laboratory. Jurnal undergraduates’ STEM motivation, identity, and
Ilmiah Pendidikan Fisika Al-Biruni, 9(2), 259–269. achievement: A short-term longitudinal study.
https://doi.org/10.24042/jipfalbiruni.v9i2.5976 Journal of Research in Science Teaching, 57.
https://doi.org/10.1002/tea.21623
Sari, D., Herlina, K., Viyanti, V., Andra, D., & Safitri,
I. (2023). E-module Newton's Law of Gravity Susilawati, S., Azizah, N. A. N., & Kusuma, H. H.
based Guided Inquiry to Train Critical Thinking (2021). Investigating differences in project
Skills. Physics Education Research Journal, 5(1), 13- activities and student digital literacy between
20. learning through electronic workbench and
doi:https://doi.org/10.21580/perj.2023.5.1.116 PhET Simulation. Jurnal Ilmiah Pendidikan Fisika
57 Al-Biruni, 10(2), 299–311.
https://doi.org/10.24042/jipfalbiruni.v10i2.100
Sekeroglu, B., Abiyev, R., Ilhan, A., Arslan, M., & 08
Idoko, J. B. (2021). Systematic Literature Review
on Machine Learning and Student Performance Vasalou, A., Benton, L., Ibrahim, S., Sumner, E., Joye,
Prediction: Critical Gaps and Possible Remedies. N., & Herbert, E. (2021). Do children with
Applied Sciences, 11(22). reading difficulties benefit from instructional
https://doi.org/10.3390/app112210907 game supports? Exploring children’s attention
and understanding of feedback. British Journal of
Semenikhina, O., Yurchenko, A., & Udovychenko, O. Educational Technology, 52(6), 2359–2373.
(2020). Formation of skills to visualize of future https://doi.org/https://doi.org/10.1111/bjet.1
physics teacher: results of the pedagogical 3145
experiment. Physical and Mathematical Education,
23(1), 122–128. Yang, C., Lan, S., Shen, W., Huang, G. Q., Wang, X.,
https://doi.org/10.31110/2413-1571-2020- & Lin, T. (2017). Towards product
023-1-020 customization and personalization in IoT-
enabled cloud manufacturing. Cluster Computing,
Shafiq, D. A., Marjani, M., Habeeb, R. A. A., & 20. https://doi.org/10.1007/s10586-017-0767-
Asirvatham, D. (2022). Student Retention Using x
Educational Data Mining and Predictive
Analytics: A Systematic Literature Review. IEEE Zimmerman, L., Spillane, S., Reiff, P., & Sumners, C.
Access, 10, 72480–72503. (2014). Comparison of Student Learning about
https://doi.org/10.1109/ACCESS.2022.318876 Space in Immersive and Computer
7 Environments. Journal Review Astronomical
Education and Outreach, 1(1), A5–A20.
Shengnan Wang, C. M. R.-D., & Meissel, K. (2018). A
systematic review of the teacher expectation

https://ejournal.walisongo.ac.id/index.php/perj/index 105
 Phy. Educ. Res. J. Vol. 5 No. 2 (2023), 97-106

106 https://ejournal.walisongo.ac.id/index.php/perj/index