Sustainable Learning Analytics: Measuring and Understanding the

Drivers of Energy Consumption of AI in Education

Marlene B

ultemann

, Katharina Simbeck

, Nathalie Rzepka

and Yannick Kalff

Hochschule f

ur Technik und Wirtschaft Berlin, Treskowallee 8, 10318 Berlin, Germany

Keywords:

Green IT, Sustainable Learning Analytics, Learning Analytics, Artiﬁcial Intelligence in Education.

Abstract:

As learning analytics increasingly relies on machine learning (ML) to provide insights and enhance educational

outcomes, the environmental impact of these ML-driven tools has become a critical but underexamined issue.

This study aims to ﬁll this gap by investigating the energy consumption of various machine learning models

commonly employed in learning analytics. This is by the execution of four distinct models — Support Vector

Machines (SVM), Multi-Layer Perceptrons (MLP), Decision Trees (DT), and Logistic Regression (LogReg)

— when applied to an educational data set. Our ﬁndings reveal signiﬁcant disparities in energy consumption

between these models, with SVM and MLP models consuming considerably more energy than their simpler

counterparts. This research serves as a call for action for the learning analytics community to prioritize energy-

efﬁcient AI models, thereby contributing to broader sustainability goals in the face of climate change.

1 INTRODUCTION

Climate change is one of the most urgent issues of our

time that threatens humans and the planet and requires

immediate action (IPCC, 2022). Mitigation efforts are

required in all sectors, including the information and

communication technology (ICT) sector (Anser et al.,

2021), which contributed to 9% of global emissions in

2018, estimated to rise to 20% until 2025 (Mancebo

et al., 2021).

While there is growing awareness about hardware

and data center emissions, the environmental impact

of software products remains underappreciated. Soft-

ware is intangible, and no visible waste is generated

when used or disposed of (Naumann et al., 2021). The

environmental costs of software, in general, remain a

niche: “In many software development projects, sus-

tainability is treated as an afterthought, as developers

are driven by time-to-market pressure and are often

not educated to apply sustainability-improving tech-

niques” (Durdik et al., 2012). Nevertheless, “our civ-

ilization runs on software” (Stroustrup, 2014), which

leads to an increased demand for non-sustainable

products (Calero and Piattini, 2017).

Above that, applications utilizing artiﬁcial intelli-

https://orcid.org/0009-0001-6941-1612

https://orcid.org/0000-0001-6792-461X

https://orcid.org/0000-0003-3123-1517

https://orcid.org/0000-0003-1595-175X

gence (AI) have risen signiﬁcantly and will continue

to do so (Wu et al., 2022). AI research is continu-

ously improving the accuracy and capabilities of AI

models, which requires growing computational power

(Henderson et al., 2020; Schwartz et al., 2020).

The increased demand also includes AI for learn-

ing analytics (LA). While positive effects on learn-

ing success or reduced drop-out rates are well-

documented beneﬁts, resource consumption of LA

remains a research gap. In a literature review on

sustainability dimensions of e-learning, most of the

examined studies have dealt with individual and so-

cial requirements (Alharthi et al., 2019). Only 4%

of the publications dealt with environmental aspects

and mainly focused on cloud computing. Despite the

growing integration of AI in education, there has been

no research on its resource consumption.

Our research contributes to this research gap and

compares the energy consumption of different LA AI

models. Our research questions are:

RQ 1: How do different AI models used in edu-

cation vary in terms of energy efﬁciency and environ-

mental impact?

RQ 2: What are the key drivers of energy con-

sumption in AI-based educational systems?

RQ 3: How much energy can be conserved by

choosing the optimal model concerning energy con-

sumption?

We analyze the energy consumption of four alter-

272

Bültemann, M., Simbeck, K., Rzepka, N. and Kalff, Y.

Sustainable Learning Analytics: Measuring and Understanding the Drivers of Energy Consumption of AI in Education.

DOI: 10.5220/0012547600003693

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Computer Supported Education (CSEDU 2024) - Volume 1, pages 272-279

ISBN: 978-989-758-697-2; ISSN: 2184-5026

native ML models designed to predict the likelihood

of correctly solving a task. Training and evaluating

models based primarily on performance metrics like

accuracy and speed is standard practice. Energy con-

sumption usually is not a determining factor. In this

study, we assess the energy consumption of models

for potential integration into a learning platform serv-

ing several thousand users. All models were trained

using the same dataset. The energy consumption of

the models was measured using the PowerJoular li-

brary (Noureddine, 2022).

This study speciﬁcally targets the energy con-

sumption associated with model inference. The en-

ergy needs for training the model, and the learning

platform itself are ignored — so are life-cycle emis-

sions, including hardware production, use and dis-

posal, and resources from software production and

uninstallation (Kern et al., 2018).

2 LITERATURE REVIEW

2.1 Measuring Energy Consumption of

Software

The energy consumption of software can be gauged

through both hardware and software-based method-

ologies. Noureddine et al. (2012) propose a software-

based approach and introduce the PowerAPI library,

which measures the energy consumption on the oper-

ating system level. Raw information is collected from

sensor modules (e.g., central processing unit (CPU) or

network) and used to calculate the system power con-

sumption as well as the power consumption of single

running processes. The authors test their approach by

running different algorithm implementations and con-

clude that an algorithm’s choice affects an applica-

tion’s energy consumption. A software-based tool in-

troduced more recently is PowerJoular (Noureddine,

2022), which uses interfaces provided by hardware

manufacturers.

An example of a hardware-based approach is to

use a measuring station consisting of two PCs - one

to execute the system under test and one to collect

the measured energy data (Junger et al., 2022). The

method is applied by B

ultemann et al. (2023) to mea-

sure the energy usage of different AIED (AI in ed-

ucation) algorithms. Verdecchia et al. (2018) imple-

ment a similar approach and extend it to a spectrum-

based energy hotspot localization to identify energy-

intensive components in the code. Mancebo et al.

(2021) propose a general process that consists of a

hardware component that measures the energy con-

sumption and a software application that analyzes the

results. This proposal is applied in a different study

on the energy utilization of health data (Garcia-Berna

et al., 2021).

Kern et al. (2018) suggest assessment criteria

when measuring resource and energy consumption

of software usage. These include user autonomy

to conﬁgure resource-saving settings, resource man-

agement and default settings, hardware requirements

(e.g., electricity consumption), and backward com-

patibility when releasing new versions. While numer-

ous studies have been conducted in the past years, and

many processes share similarities, there is currently

no universally recognized standard process for energy

assessments of software.

2.2 Measuring Energy Consumption of

Machine Learning

Machine learning is used for various cases like speech

and image recognition, translations, or recommenda-

tions. ML algorithms use large data sets to recog-

nize patterns, learn a behavior, and make predictions

(Helm et al., 2020). Several parameters contribute

to the total energy costs of AI: execution costs of

a model, size of the training data set, and the total

number of hyperparameter experiments (i.e., number

of training iterations) (Schwartz et al., 2020; Sharir

et al., 2020). The increasing size of a model directly

increases the costs for training and inference. In addi-

tion, energy utilization of storage is rising as the data

sets become increasingly larger (Dhar, 2020). Other

drivers of the environmental cost of ML are the lo-

cal energy mix, water demand for data center cooling,

and electronic waste (Patterson et al., 2021; Wu et al.,

2022; Henderson et al., 2020). To reduce the envi-

ronmental impact of ML models, the choice of mod-

els and hardware is essential (Schwartz et al., 2020;

Sharir et al., 2020).

2.3 Artiﬁcial Intelligence and Learning

Analytics

LA is “the measurement, collection, analysis and re-

porting of data about learners and their contexts, for

purposes of understanding and optimising learning

and the environments in which it occurs” (LAK, 2011;

Greller and Drachsler, 2012). Patterns in a learner’s

behavior are recognized, and individual measures

may be drawn (Gobert and Sao Pedro, 2016). LA is

implemented in massive open online courses, intelli-

gent tutoring systems, or learning management sys-

tems (Lu et al., 2018).

Machine learning in learning analytics has grown

considerably in the last decade. AI for Learning

Sustainable Learning Analytics: Measuring and Understanding the Drivers of Energy Consumption of AI in Education

273

Analytics relies on large amounts of data to gener-

ate models and predictions (Brooks and Thompson,

2017). Data is collected from the interaction be-

tween instructors, learners, the educational environ-

ment, and administrative or performance processes

(Brooks and Thompson, 2017; Romero et al., 2014).

Depending on the use case, AI techniques such as ma-

chine learning, natural language processing, or rec-

ommender systems are applied in LA (Brooks and

Thompson, 2017; McNamara et al., 2017; Chen et al.,

2020). Chen et al. (2020) investigate the impact of

AI on the education sector by reviewing 30 stud-

ies. They ﬁnd that AI in education can take differ-

ent forms, e.g., in computers, embedded systems, or

web-based platforms. It primarily supports instruc-

tors pursuing administrative tasks like grading or pro-

viding feedback. On the other hand, students beneﬁt

from AIED as the educational content is tailored to

their individual needs, thus increasing learning suc-

cess. They conclude that AI has a considerable effect

on the educational sector. The use of AI in LA en-

ables adaptivity and individualized interventions. An

adaptive learning environment observes and interprets

the user’s actions to identify and model learners’ indi-

vidual needs and preferences (Paramythis and Loidl-

Reisinger, 2003). AI adapts to a user’s unique level

of skills in real time and supports the personal learn-

ing process using educational data. Interventions take

place in the forms of direct messages to the learner

(i.e., when a learner is identiﬁed as being at risk of

failing), individual content, or actionable feedback

(Wong and Li, 2018).

3 CASE STUDY: THE LEARNING

PLATFORM “Orthograﬁetrainer”

Orthograﬁetrainer

(OT) is a learning platform for

students from grades 5 to 10 to acquire and consol-

idate German spelling skills. It was established in

2008 and has since been accessed by more than 1.25

million. users. The platform provides more than 5000

tasks, which have been solved over 12 million times.

The training platform offers exercises in various

areas of competencies, such as comma formation or

capitalization. Each set of exercises bundles ten sen-

tences that have to be completed. In its default setup,

the platform works as follows: if a learner enters

a sentence incorrectly, two additional sentences will

be added to the exercise set before the previously

wrongly submitted sentence is displayed again. It pro-

vides immediate feedback and dynamic adjustments.

In the scope of an experiment, OT was transformed

https://orthograﬁetrainer.net/

into an adaptive platform using a machine learning

model to predict the probability of solving a task sen-

tence correctly (Rzepka et al., 2022, 2023). The de-

ployed adaptive learning interventions on the plat-

form outperform the original setup regarding learn-

ing gains and drop-outs (Rzepka, 2023). Before con-

ducting the experiment, four different ML algorithms

were trained on the data set: a decision tree, a logis-

tic regression, SVM, and MLP. The dependent vari-

able is the success of solving a task correctly (1) or

incorrectly (0). The data set consists of (1) demo-

graphic data (e.g., grade and gender), (2) task data

(e.g., difﬁculty, previous attempts), and (3) a learn-

ing history of submitted exercises over the past three

months (Rzepka, 2023). The models differ slightly in

accuracy (96.62 - 97.09), precision (97.38 - 97.97),

and recall (95.82 - 96.31) (Rzepka, 2023).

4 EXPERIMENTAL SETUP

4.1 Experimental Setup

In this study, we measure the energy consumption

of model inference, i.e., when the trained model is

provided with new data to predict the probability of

solving a sentence correctly. In each experiment, the

Python code predicts the probability of doing the task

using one of the models. Each model is executed 30

times, and an average is calculated for the measured

values. The averages of the four models are compared

to the baseline (BL), which is the metered energy con-

sumption of the system without inference running.

Before the experiments, the computer was newly set

up with the operating system, required software, and

dependencies for the measurement. Because of the

model requirements, Python is installed in two differ-

ent versions: version 3.7 for the decision tree, logistic

regression, and SVM, and version 3.10 for the MLP.

Only local resources are measured in this setup, as

it is impossible to access any further infrastructure.

The model is saved locally, so the differences across

models can be estimated. The PowerJoular software

(Noureddine, 2022) measures energy and system re-

source consumption. It is deployed on a Raspberry Pi.

PowerJoular reads resource consumption on the CPU

cycles using a power polynomial regression model

read from the /proc/stat system ﬁle. Its error rate

is estimated between 0.3% and 3.83%. PowerJoular

can monitor the entire system or processes based on a

process ID.

Measurements are conducted on a Raspberry Pi

4b with 4 GB RAM with Broadcom BCM2711, a 64-

bit system on chip, and a processor sub-architecture

CSEDU 2024 - 16th International Conference on Computer Supported Education

274

at 1.5GHz. Raspberry Pi devices have been used for

various studies in the past (Kumar et al., 2018; Saha

et al., 2018; Umarghanis et al., 2020) as they are in-

expensive and compatible with many other hardware

devices.

In this setup, PowerJoular is used to measure the

system’s CPU usage and electric power. Memory us-

age is captured by the Linux command line tool PS,

as suggested by Wagner et al. (2023). PS measures

the relative memory usage of the system on a process

basis (Balister et al., 2007). In this experiment, the

processes of each model execution (i.e., the execution

of the respective Python script) are monitored, and

memory is reported. In the baseline measurement, the

entire system is monitored. Time stamps are logged in

the code to get an accurate start and stop time for each

of the 30 measurement cycles. The variables gathered

are CPU usage as a percentage, memory consumption

as a percentage, and electrical power consumption in

watts.

4.2 Experimental Hypotheses

The objective of the following measurements is to an-

alyze the effect of the independent variables (the mod-

els, i.e., decision tree, logistic regression, SVM, and

MLP) on the dependent variables (CPU usage, mem-

ory usage, power usage) to answer RQ1 and RQ2.

The null hypothesis states that there is no signiﬁcant

difference between the independent variables (m) and

the dependent variables. The alternative hypothesis

states that a statistically signiﬁcant difference be-

tween a dependent and an independent variable is ob-

served.

: µ

= µ

: µ

̸= µ

RQ3 aims to determine how much energy could

be conserved by choosing a model. We assume that

the amount of energy saved in comparison to the base-

line is not signiﬁcant (H

), as opposed to statistically

signiﬁcant savings (H

: µ

= µ

: µ

̸= µ

5 DATA AND RESULTS

5.1 Measured System Demand of ML

Models and Baseline System

We measure system demand using CPU utilization

[%], memory utilization [%], and power in Watt

(W)

. The values of these measures for all four

models and the baseline system are available in ta-

ble 1. Compared to the baseline average of 1.8%,

CPU utilization measured by PowerJoular of the four

models ranges between 24.47% for the decision tree

(DT) model and 33.59% for the multilayer percep-

tron (MLP) model. The lowest mean value for

memory utilization is also measured in the decision

tree with 29.47%, the highest in MLP (41.97%).

Baseline memory utilization is 7%. For electric

power, mean values range from 3.33W/3.38W (deci-

sion tree/logistic regression) to 3.91W (MLP), in con-

trast to a baseline metric of 3.24W.

Table 1: Descriptive variables for system demand.

Min Max Mean Median Std

CPU Utilization

DT [%]

1.00 33.84 24.47 26.75 0.07

Memory

Utilization DT

[%]

1.20 41.8 29.74 31.8 8.12

Electric Power

DT [W]

3.03 3.65 3.33 3.37 0.1

CPU Utilization

LogReg [%]

1.5 52.12 26.1 26.82 0.06

Memory

Utilization

LogReg [%]

15.5 41.9 32.72 33.75 7.19

Electric Power

LogReg [W]

3.03 4.70 3.38 3.36 0.19

CPU Utilization

SVM [%]

1.25 79.34 28.36 28.25 0.02

Memory

Utilization SVM

[%]

1.20 43.9 30.92 30.9 1.13

Electric Power

SVM [W]

3.03 5.86 3.42 3.41 0.08

CPU Utilization

MLP [%]

0.0 50.50 33.59 45.89 0.17

Memory

Utilization MLP

[%]

8.5 63.0 41.97 43.05 10.01

Electric Power

MLP [W]

3.03 4.61 3.91 4.33 0.6

Memory

Utilization BL

[%]

7.0 8.0 7.00 7.0 0.01

Electric Power

BL [W]

3.03 3.41 3.24 3.26 0.05

In the next step, the mean values are compared

to the baseline values (table 2). All models require

more resources in terms of CPU utilization, memory

utilization, and electric power than the baseline.

Cf. the GitHub repository https://github.com/

marlenebuelt/messung-2 for results.

Sustainable Learning Analytics: Measuring and Understanding the Drivers of Energy Consumption of AI in Education

275

Table 2: Relative difference to baseline values (pp.: per-

centage points).

Delta

DT %

Delta

LogReg

Delta

SVM %

Delta

MLP %

CPU

Uti-

lization

[pp.]

1260.49 1350.90 1476.93 1767.92

Memory

Uti-

lization

[pp.]

324.90 367.49 341.72 499.53

Electric

Power

[pp.]

2.90 4.44 5.65 20.63

The data undergoes a Shapiro-Wilk test to as-

sess normality. However, the results indicate that the

data is not normally distributed. Therefore, a Mann-

Whitney U test compares the mean system demand

parameters between the models and baseline. The ef-

fect sizes are estimated using Cliff’s delta. It takes

on a value between -1 and 1, with 0 indicating no ef-

fect and |1| a very strong effect. The threshold for a

medium effect is set at a value larger than |0.28|, and

a large effect is considered |0.43| and above (Vargha

and Delaney, 2000). Table 3 shows that almost all dif-

ferences between models and between baseline and

models are highly signiﬁcant with a p-value of 0.00.

Only the p-values for the correlation of CPU utiliza-

tion and power conversion between DT and LogReg

models are not close to zero but are still very low at

0.01. As indicated by Cliff’s delta (refer to Table 4),

the effect size consistently demonstrates a high mag-

nitude when comparing the ML models to the base-

line. Additionally, there is a notable high effect size

between the decision tree and logistic regression mod-

els on the one hand and SVM and MLP on the other.

5.2 Estimation of the Annual Energy

Consumption

To interpret the system demand, the model runtimes

need to be considered. The models differed in mean

power and average run time per inference (table 5).

To calculate the energy consumption per year, we re-

quire an estimation of the total number of inferences

per year. We base our assumption on the experiment

described in Rzepka (2023). In four months, 38,000

sessions took place on the platform (114,000 sessions

per year). One session is assumed to include an exer-

cise set of ten sentences. As the ﬁrst sentence is pre-

deﬁned, nine inferences are undertaken to determine

the second sentence, eight inferences for the third, and

Table 3: Mann–Whitney U test.

BL DT LogReg SVM

CPU

DT 0.00

LogReg 0.00 0.01

SVM 0.00 0.00 0.00

MLP 0.00 0.00 0.00 0.00

Memory

DT 0.00

LogReg 0.00 0.00

SVM 0.00 0.00 0.00

MLP 0.00 0.00 0.00 0.00

El. Power

DT 0.00

LogReg 0.00 0.01

SVM 0.00 0.00 0.00

MLP 0.00 0.00 0.00 0.00

Table 4: Cliff’s delta.

BL DT LogReg SVM

CPU

DT -0.98

LogReg -0.98 −0.11

SVM -1.00 -0.76 -0.70

MLP -0.91 -0.41 -0.37 −0.21

Memory

DT -0.97

LogReg -1.00 −0.23

SVM -1.00 0.12 -0.7

MLP -1.00 -0.75 -0.37 -0.81

El. Power

DT -0.69

LogReg -0.85 −0.11

SVM -0.99 -0.76 -0.70

MLP -0.62 -0.44 -0.38 −0.21

so forth. This sums up to 44 inferences. However, if

the prediction for a learner to correctly solve a sen-

tence is determined to be less than 50%, the proce-

dure changes, and additional inferences are executed

to provide training sentences. Hence, the total number

of inferences in a session varies. This paper assumes

64 inferences per session, thus 7,296,000 inferences

per year.

Table 5 summarizes the use of electric power in a

single inference in watt-seconds (Ws), the estimated

usage of power per year in watt hours (Wh) as ex-

plained in chapter 5.2, and the costs of electric power

in euros per year. We assume the average market price

for electricity in Germany for the year 2022 of 0.3279

EUR per kWh Eurostat (2023). The average usage

per inference ranges between 45.77 Ws in the deci-

sion tree model and 3,783.27 Ws when executing an

SVM. This sums up to 92,760 Wh in a decision tree

CSEDU 2024 - 16th International Conference on Computer Supported Education

276

Table 5: Estimation of annual energy consumption.

Time per

Inference [s]

Mean

Power [W]

Energy per

Inference [Ws]

DT 13.8 3.33 45.77

LogReg 15.97 3.38 51.64

SVM 4085.00 3.42 3783.27

MLP 34.83 3.91 129.19

Energy

per Year [kWh]

Costs per

Year [ C]

DT 92.8 30.43

LogReg 104.7 34.33

SVM 7667.4 2514.14

MLP 261.8 85.84

and to 7,667,428 Wh in an SVM. The energy costs of

a decision tree amount to 30.43 Eur, of a logistic re-

gression 34.33 Eur, of an SVM 2,514.14 Eur, and of

an MLP 85,84 Eur.

6 DISCUSSION

Answering RQ1, we ﬁnd that ML models vary in en-

ergy consumption. The data indicates signiﬁcant dif-

ferences between all models and the baseline as well

as between the models. The low basic load of a Rasp-

berry Pi contributes to the signiﬁcance and the ef-

fect sizes when compared against the baseline. Our

data indicates that the decision tree is slightly more

efﬁcient than the logistic regression model, while

MLP and especially SVM require much more elec-

tric power. The similar signiﬁcance values and effect

sizes of CPU utilization and electric power consump-

tion in the PowerJoular measurement indicate that the

parameters correlate. Thus, CPU usage could be con-

sidered as a proxy for power use.

The general ﬁndings of this experiment coincide

with other ﬁndings, e.g., of Strubell et al. (2019) and

Frey et al. (2022). The study reveals that differences

between models are discernible and exhibit a broad

range. These ﬁndings align with the authors’ conclu-

sions drawn in a prior experiment, as documented in

ultemann et al., 2023).

Our results demonstrate that the complexity of

a model contributes to the consumption of com-

putational resources and, thus, energy consumption

(RQ2). The execution of all models resulted in a

signiﬁcant increase in the load on the observed pa-

rameters (CPU utilization, memory utilization, elec-

tric power) compared to the baseline. This effect

was most pronounced for SVM and MLP. Less com-

plex models, such as decision trees and logistic re-

gressions, could make predictions more rapidly than

SVM and MLP. Several layers and the number of in-

terconnected neurons contribute to the increased en-

ergy consumption of an MLP. SVMs require more

runtime and energy due to the complex optimization

process of ﬁnding the largest margin between many

data points. The training kernel matrix squares with

the size of the dataset, thus making it inefﬁcient on

large data sets like the ones used in this experiment.

On the other hand, decision trees and logistic regres-

sions have a relatively simple prediction process that

is boolean or mathematically explainable. Memory

usage reveals the highest values in MLP. This indi-

cates that a large amount of data is processed in the

prediction process of the models. Regarding RQ3, our

results indicate that the SVM model would substan-

tially increase energy consumption by 30-80. Real-

life implications could differ based on the hardware

used in the actual data center, but we expect an effect

of the same magnitude.

It is essential to acknowledge certain limitations

that reduce the generalizability of our ﬁndings. This

study only analyses ML model inference’s energy de-

mand and neglects model training’s energy consump-

tion and life-cycle emissions. Further, results depend

on the deployed hardware; data centers usually uti-

lize different hardware. Finally, models should also

be trained on other datasets to validate results.

As we reveal differences in the resource consump-

tion of different ML models, future research in ML

implementations for LA should consider the resource

impact of the chosen models in line with parameters

such as accuracy, fairness, and explainability. Both

energy consumption and learning outcomes should

be considered when comparing AI-related emissions

and learning outcomes to ﬁnd a well-balanced, cost-

efﬁcient solution. Energy costs, in particular, are

an essential economic determinant in times of per-

sistently high energy prices. If the performance of

models reveals similar results, environmental aspects

should be considered. This conﬁrms the ﬁndings in

Schwartz et al. (2020), who propose to focus AI re-

search on efﬁciency measures.

7 CONCLUSION

This study aimed to ﬁll a critical research gap by

investigating the environmental impact of machine

learning models on educational platforms, speciﬁ-

cally focusing on their energy efﬁciency and con-

sumption. Our ﬁndings indicate signiﬁcant variations

in energy consumption across different ML mod-

els. Decision trees and logistic regression models are

more energy-efﬁcient than SVMs and MLPs. The

study also reveals that the model’s complexity is pro-

Sustainable Learning Analytics: Measuring and Understanding the Drivers of Energy Consumption of AI in Education

277

portional to its energy consumption, with SVM and

MLP requiring considerably more computational re-

sources and computing time. These insights answer

our research questions, providing a better understand-

ing of the critical drivers of energy consumption in

AI-based educational systems (RQ2) and the poten-

tial energy savings by selecting more efﬁcient models

(RQ3).

From a practical perspective, our ﬁndings indi-

cate that choosing an energy-efﬁcient model could

reduce energy consumption signiﬁcantly compared

to more resource-intensive models. This is partic-

ularly relevant for educational platforms that oper-

ate at scale, where even marginal improvements in

energy efﬁciency can lead to substantial reductions

in operational costs and carbon footprint. A key

driver is a model’s complexity caused by the train-

ing algorithms. A logistic regression is mathemati-

cally explainable, while a decision tree uses a boolean

decision process. SVMs use complex optimization

processes, and MLPs consist of many layers of in-

terconnected neurons, which makes prediction pro-

cesses more complex and the model larger in storage.

This causes the processes’ resource consumption to

rise and increases the runtime. Our work has shown

that choosing resource-efﬁcient ML models for edu-

cational purposes is complementary to achieving ob-

jectives related to explainability and speed of infer-

ence.

Future research could extend this work by consid-

ering a broader range of ML models and algorithms.

Studies could also incorporate energy consumption

during the training phase and other life-cycle emis-

sions to provide a more comprehensive view of the

environmental impact. This study shows options for

incorporating sustainability considerations into de-

veloping LA systems. Performance predictions im-

ply that signiﬁcantly lower energy consumption can

achieve the same results. In conclusion, educational

achievements and environmental sustainability are

not mutually exclusive.

REFERENCES

(2011). LAK ’11: Proceedings of the 1st International Con-

ference on Learning Analytics and Knowledge, New

York. ACM.

Alharthi, A. D., Spichkova, M., and Hamilton, M. (2019).

Sustainability requirements for eLearning systems: a

systematic literature review and analysis. Require-

ments Engineering, 24:523–543.

Anser, M. K., Ahmad, M., Khan, M. A., Zaman, K., Nas-

sani, A. A., Askar, S. E., Abro, M. M. Q., and Kab-

bani, A. (2021). The role of information and commu-

nication technologies in mitigating carbon emissions:

evidence from panel quantile regression. Environmen-

tal Science and Pollution Research, 28:21065–21084.

Balister, P. J., Dietrich, C., and Reed, J. H. (2007). Mem-

ory usage of a software communication architecture

waveform. In 2007 Software Deﬁned Radio Technical

Conf. Product Exposition. Citeseer.

Brooks, C. and Thompson, C. (2017). Predictive modelling

in teaching and learning. Handbook of learning ana-

lytics, pages 61–68.

ultemann, M., Rzepka, N., Junger, D., Simbeck, K., and

uller, H.-G. (2023). Energy consumption of ai in

education: A case study. In 21. Fachtagung Bildung-

stechnologien (DELFI), pages 219–224. Gesellschaft

ur Informatik e.V., Bonn.

Calero, C. and Piattini, M. (2017). Puzzling out software

sustainability. Sustainable Computing: Informatics

and Systems, 16:117–124.

Chen, L., Chen, P., and Lin, Z. (2020). Artiﬁcial In-

telligence in Education: A Review. IEEE Access,

8:75264–75278.

Dhar, P. (2020). The carbon impact of artiﬁcial intelligence.

Nature Machine Intelligence, 2(8):423–425.

Durdik, Z., Klatt, B., Koziolek, H., Krogmann, K., Stam-

mel, J., and Weiss, R. (2012). Sustainability guide-

lines for long-living software systems. In 2012 28th

IEEE International Conference on Software Mainte-

nance (ICSM), pages 517–526.

Eurostat (2023). Electricity prices by type of user. Data

Browser.

Frey, N. C., Li, B., McDonald, J., Zhao, D., Jones, M.,

Bestor, D., Tiwari, D., Gadepally, V., and Samsi, S.

(2022). Benchmarking resource usage for efﬁcient

distributed deep learning. In 2022 IEEE High Per-

formance Extreme Computing Conference (HPEC),

pages 1–8. IEEE.

Garcia-Berna, J. A., Fernandez-Aleman, J. L., de Gea, J.

M. C., Toval, A., Mancebo, J., Calero, C., and Garcia,

F. (2021). Energy efﬁciency in software: A case study

on sustainability in personal health records. Journal

of cleaner production, 282:124262.

Gobert, J. D. and Sao Pedro, M. A. (2016). Digital as-

sessment environments for scientiﬁc inquiry practices.

The Wiley handbook of cognition and assessment:

Frameworks, methodologies, and applications, pages

508–534.

Greller, W. and Drachsler, H. (2012). Translating learning

into numbers: A generic framework for learning ana-

lytics. Journal of Educational Technology & Society,

15(3):42–57.

Helm, J. M., Swiergosz, A. M., Haeberle, H. S., Karnuta,

J. M., Schaffer, J. L., Krebs, V. E., Spitzer, A. I., and

Ramkumar, P. N. (2020). Machine learning and ar-

tiﬁcial intelligence: deﬁnitions, applications, and fu-

ture directions. Current reviews in musculoskeletal

medicine, 13:69–76.

Henderson, P., Hu, J., Romoff, J., Brunskill, E., Jurafsky,

D., and Pineau, J. (2020). Towards the systematic re-

porting of the energy and carbon footprints of machine

learning. The Journal of Machine Learning Research,

21(1):10039–10081.

CSEDU 2024 - 16th International Conference on Computer Supported Education

278

IPCC (2022). Summary for Policymakers, page In Press.

Cambridge University Press, Cambridge, UK. Ed-

itors: P

ortner, H. O. and Roberts, D. C. and Tig-

nor, M. and Poloczanska, E. S. and Mintenbeck, K.

and Alegr

ıa, A. and Craig, M. and Langsdorf, S. and

oschke, S. and M

oller, V. and Okem, A. and Rama,

Junger, D., Wohlgemuth, V., and Kammer, E. (2022). Con-

ception and test of a measuring station for the anal-

ysis of the resource and energy consumption of ma-

terial ﬂow-oriented environmental management infor-

mation systems (EMIS). In Wohlgemuth, V., Nau-

mann, S., Arndt, H.-K., Behrens, G., and H

ob, M., ed-

itors, EnviroInfo 2022, page 211, Bonn. Gesellschaft

ur Informatik e.V.

Kern, E., Hilty, L. M., Guldner, A., Maksimov, Y. V., Filler,

A., Gr

oger, J., and Naumann, S. (2018). Sustain-

able software products—Towards assessment criteria

for resource and energy efﬁciency. Future Generation

Computer Systems, 86:199–210.

Kumar, A., Chattree, G., and Periyasamy, S. (2018). Smart

healthcare monitoring system. Wireless Personal

Communications, 101:453–463.

Lu, O. H., Huang, A. Y., Huang, J. C., Lin, A. J., Ogata, H.,

and Yang, S. J. (2018). Applying learning analytics

for the early prediction of Students’ academic perfor-

mance in blended learning. Journal of Educational

Technology & Society, 21(2):220–232.

Mancebo, J., Garcia, F., and Calero, C. (2021). A process

for analysing the energy efﬁciency of software. Infor-

mation and Software Technology, 134:106560.

McNamara, D. S., Allen, L., Crossley, S., Dascalu, M., and

Perret, C. A. (2017). Natural language processing and

learning analytics. Handbook of learning analytics,

93.

Naumann, S., Guldner, A., and Kern, E. (2021). The eco-

label blue angel for software—Development and com-

ponents. In Advances and New Trends in Environ-

mental Informatics: Digital Twins for Sustainability,

pages 79–89. Springer.

Noureddine, A. (2022). PowerJoular and JoularJX: multi-

platform software power monitoring tools. In 2022

18th International Conference on Intelligent Environ-

ments (IE), pages 1–4. IEEE.

Noureddine, A., Bourdon, A., Rouvoy, R., and Seinturier,

L. (2012). A Preliminary Study of the Impact of Soft-

ware Engineering on GreenIT. In 2012 First Interna-

tional Workshop on Green and Sustainable Software

(GREENS), pages 21–27. IEEE.

Paramythis, A. and Loidl-Reisinger, S. (2003). Adaptive

learning environments and e-learning standards. In

Second european conference on e-learning, volume 1,

pages 369–379.

Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L.-

M., Rothchild, D., So, D., Texier, M., and Dean, J.

(2021). Carbon emissions and large neural network

training.

Romero, C., Romero, J. R., and Ventura, S. (2014). A sur-

vey on pre-processing educational data. Educational

data mining: applications and trends, pages 29–64.

Rzepka, N. (2023). Transforming ﬁrst language learning

platforms towards adaptivity and fairness: models, in-

terventions and architecture. PhD thesis, Humboldt

Universi

at zu Berlin, Berlin.

Rzepka, N., Simbeck, K., M

uller, H.-G., and Pinkwart, N.

(2022). Keep it up: In-session dropout prediction to

support blended classroom scenarios. In CSEDU (2),

pages 131–138.

Rzepka, N., Simbeck, K., M

uller, H.-G., and Pinkwart, N.

(2023). Go with the Flow: Personalized Task Se-

quencing Improves Online Language Learning. In

International Conference on Artiﬁcial Intelligence in

Education, pages 90–101. Springer.

Saha, A. K., Sircar, S., Chatterjee, P., Dutta, S., Mitra, A.,

Chatterjee, A., Chattopadhyay, S. P., and Saha, H. N.

(2018). A raspberry Pi controlled cloud based air and

sound pollution monitoring system with temperature

and humidity sensing. In 2018 IEEE 8th Annual Com-

puting and Communication Workshop and Conference

(CCWC), pages 607–611. IEEE.

Schwartz, R., Dodge, J., Smith, N. A., and Etzioni, O.

(2020). Green AI. Communications of the ACM,

63(12):54–63.

Sharir, O., Peleg, B., and Shoham, Y. (2020). The cost

of training nlp models: A concise overview. arXiv

preprint arXiv:2004.08900.

Stroustrup, B. (2014). Programming: principles and prac-

tice using C++. Pearson Education.

Strubell, E., Ganesh, A., and McCallum, A. (2019). Energy

and policy considerations for deep learning in NLP.

Umarghanis, S. A., Darari, F., and Wibisono, A. (2020).

A Low-Cost IoT Platform for Crowd Density Detec-

tion in Jakarta Commuter Line. In 2020 International

Conference on Advanced Computer Science and In-

formation Systems (ICACSIS), pages 121–128. IEEE.

Vargha, A. and Delaney, H. D. (2000). A critique and im-

provement of the CL common language effect size

statistics of McGraw and Wong. Journal of Educa-

tional and Behavioral Statistics, 25(2):101–132.

Verdecchia, R., Guldner, A., Becker, Y., and Kern, E.

(2018). Code-level energy hotspot localization via

naive spectrum based testing. In Advances and New

Trends in Environmental Informatics: Managing Dis-

ruption, Big Data and Open Science, pages 111–130.

Springer.

Wagner, L., Mayer, M., Marino, A., Nezhad, A. S., Zwaan,

H., and Malavolta, I. (2023). On the Energy Con-

sumption and Performance of WebAssembly Binaries

across Programming Languages and Runtimes in IoT.

In Proceedings of the 9th International Conference on

Evaluation and Assessment on Software Engineering

(EASE).

Wong, B. T. and Li, K. C. (2018). Learning analytics inter-

vention: A review of case studies. In 2018 Interna-

tional Symposium on Educational Technology (ISET),

pages 178–182. IEEE.

Wu, C.-J., Raghavendra, R., Gupta, U., Acun, B., Ardalani,

N., Maeng, K., Chang, G., Aga, F., Huang, J., Bai,

C., et al. (2022). Sustainable ai: Environmental impli-

cations, challenges and opportunities. Proceedings of

Machine Learning and Systems, 4:795–813.

Sustainable Learning Analytics: Measuring and Understanding the Drivers of Energy Consumption of AI in Education

279