Towards Explainability in Modern Educational Data Mining: A Survey

Basile Tousside, Yashwanth Dama and J

org Frochte

Bochum University of Applied Science, 42579 Heiligenhaus, Germany

Keywords:

Educational Data Mining, Data Mining, Explainable Artiﬁcial Intelligence.

Abstract:

Data mining has become an integral part of many educational systems, where it provides the ability to explore

hidden relationship in educational data as well as predict students’ academic achievements. However, the pro-

posed techniques to achieve these goals, referred to as educational data mining (EDM) techniques, are mostly

not explainable. This means that the system is black-boxed and offers no insight regarding the understanding

of its decision making process. In this paper, we propose to delve into explainability in the EDM landscape.

We analyze the current state-of-the-art method in EDM, empirically scrutinize their strengths and weaknesses

regarding explainability and making suggestions on how to make them more explainable and more trustwor-

thy. Furthermore, we propose metrics able to efﬁciently evaluate explainable systems integrated in EDM

approaches, therefore quantifying the degree of explanability and trustworthiness of these approaches.

1 INTRODUCTION

Data mining, also referred to as Knowledge discov-

ery in data (KDD), has become a standard when deal-

ing with large datasets. It is the process of extract-

ing useful information from a large set of data, tun-

ing those information into valuable insights and pre-

dictions. When data mining is applied to educational

datasets – often collected via a learning management

system (LMS) platform, it is referred to as educational

data mining (EDM).

The growth of online learning over the past two

years due to COVID-19 has made it much easier to

collect a mass of educational data set in universities

and other educational facilities. This mass of data

has whetted the appetite of educational data mining

researchers, producing a boost in the ﬁeld. How-

ever, educational data mining is not a new paradigm

in itself. Its origin goes back to the year 2000

when it was brieﬂy addressed during research on the

intelligence of tutorial systems whose results were

presented during the workshop on “Applying Ma-

chine Learning to ITS Design/Construction” hosted

in Montreal, Canada.

During the 2000s, several other workshops on data

mining in education were held, including the “Educa-

tional Data Mining” workshop created in 2005 and

hosted in Pittsburg, USA. Two years later, a similar

and complementary workshop named workshop on

“Applying Data Mining in e-Learning” was held for

the ﬁrst time in Crete, Greece, in 2007. Most of these

workshops have evolved into conferences and are

known as such nowadays. This is the case of the “In-

ternational Conference on Educational Data mining”,

which takes place this year in London, United King-

dom. In recent years, the topic of EDM has become

increasingly important and several new conferences

have been born. The most prominent among these

conferences are the “International Conference on Ar-

tiﬁcial Intelligence in Education (AIED)”, the “Inter-

national Conference on Learning Analytics & Knowl-

edge (LAK)” and the “International Educational Data

Mining Society” founded in 2011. Over the years,

the idea and purpose of the EDM has evolved signif-

icantly. In its early days, it was limited to predict-

ing students’ performance in speciﬁc courses. During

the last few years, it evolved into improving the edu-

cational process and explaining educational strategies

for better decision-making (Silva and Fonseca, 2017).

This was achieved via the development and adop-

tion of statistical, machine-learning and data-mining

methods to study educational data generated by stu-

dents and instructors.

Currently, the main metric used in EDM is the

overall prediction accuracy. However, educational

data mining exhibits a multi-targeted problem and

only focusing on the accuracy might be misleading.

In this paper, we show that in addition to accuracy,

explainability needs to be considered to better address

problems in educational data mining.

212

Tousside, B., Dama, Y. and Frochte, J.

Towards Explainability in Modern Educational Data Mining: A Survey.

DOI: 10.5220/0011529400003335

In Proceedings of the 14th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2022) - Volume 1: KDIR, pages 212-220

ISBN: 978-989-758-614-9; ISSN: 2184-3228

LMS/CMS

Raw

Education

Data

Target values

Preprocessed

Data

Pattern

knowledge

collect

Choose

Preprocessing

EDM methods

Interpretation /

Evaluation

Figure 1: Educaton Data Mining (EDM) pipeline: Overview of how EDM techniques are applied from data gathering in

learning environments such as learning management system (LMS) platform to knowledge discovery. The dashed lines

indicates revisiting previous state of the process, when the result of the current state is not suitable.

Explainability is a relatively new paradigm in data

mining that aims to shed light on the decision-making

process in predictions made by intelligent systems.

In the context of EDM, it aims to explain for exam-

ple, why a data mining algorithm predicts that a stu-

dent will not pass the mathematics subject. This is a

real challenge since modern algorithms are so sophis-

ticated that their decisions have become too complex

to express in humans understandable terms. However,

signiﬁcant progress has been made in the explain-

able artiﬁcial intelligence community and explain-

ability techniques have been successfully applied to

ML models. The question of whether these tech-

niques can be used in EDM is a major concern in this

paper. A review of those techniques will be presented

in Section 2. Furthermore, section 3 will give a de-

tailed overview of the state of explainability in educa-

tional data mining.

Another artiﬁcial intelligence discipline which is

closely related to EDM is Learning Analytics (LA).

According to the call of the First International Confer-

ence on Learning Analytics and Knowledge (LAK),

LA can be deﬁned as the measurement, collection,

analysis and reporting of data about learners and their

contexts, to understand and optimise learning and the

environments in which it occurs (Calvet Li

an and

Juan P

erez, 2015). The learning analytic process

is the same as the data mining process and can be

summarized as: an iterative process in which data is

extracted from an educational environment and pre-

processed before applying quantitative methods to

help instructors in decisions making. A schematic

representation of this process is shown in Figure 1.

In this paper, we focus on EDM and address

the limitations of the prediction accuracy, which is

the current state-of-the-art metric in evaluating and

analysing educational data mining techniques. Fur-

thermore, we give an overview of the state of explain-

ability in the EDM landscape. As will be seen in Sec-

tion 3, when designing EDM approaches, explainabil-

ity can be combined with prediction accuracy to over-

come the shortcomings of the latter. The main contri-

butions of this paper is as follows:

• We Provide a survey on EDM and compare exist-

ing EDM techniques.

• We deﬁne explainability and review the recent im-

provement made in explainable artiﬁcial intelli-

gence.

• We review the current state of explainability in

modern EDM approaches.

• We discuss the multi-dimensional requirement for

educational data mining, pointing out the limita-

tion of the prediction accuracy metric and propose

to integrate explainability in future EDM tech-

niques in order to ﬁll the accuracy gaps.

2 EXPLANABILTIY IN GENERAL

Despite being opaque and difﬁcult to interpret, data

mining models have spread out in many applications

from different domains such as education, healthcare,

criminal justice or autonomous driving. They help ed-

ucators predicting students’ grade, doctors making di-

agnostics or judges taking decisions. This raises con-

cern and demand for humanly understandable deci-

sions of these artiﬁcial intelligent systems.

To satisfy this demand, the concept of eXplain-

able Artiﬁcial Intelligence (XAI) has emerged with

the goal of producing AI models whose decisions are

interpretable and understandable by a non-AI-expert

human. A doctor should, for example, understand

Towards Explainability in Modern Educational Data Mining: A Survey

213

why an AI model diagnoses a tumour rather than sim-

ply operates on a patient because the model said so.

A judge using AI-based software to pass judgement

should be able to explain the motives and facts be-

hind that decision. An educator or visa ofﬁcer should

understand why a model predicts that a student will

not succeed in his studies, instead of simply rejecting

the application because the model recommends to do

so.

On the other side, there is a growing number of

legal regulations restricting how companies and orga-

nizations make use of AI-based decisions. The EU’s

General Data Protection Rights (GDPR), – which

contains a “right to an explanation” article – for exam-

ple requires a human to examine and conﬁrm the con-

clusions reached by an AI algorithm before applying

them. Another example is the “algorithmic account-

ability bill” of the city of New York, which requires

the fairness and validity of AI-based decisions to be

veriﬁable.

Due to these critical needs for ethics, fairness and

transparency of AI systems, there has been a con-

siderable increase in research interest in explainabil-

ity methods for interpreting and understanding black-

boxed automated decisions. This has led to some con-

vincing model explanation algorithms, especially in

the supervised learning domain (Burkart and Huber,

2021). Some of these algorithms have a promising

future in the EDM ﬁeld. In Section 3, we will elab-

orate on which XAI algorithms we believe could be

integrated to which EDM approach to make it more

explainable.

One problem that can pose difﬁculties in the ac-

ceptance of this explainable methods is that the ex-

plainability capability of these algorithms is some-

times subjective since there is currently no consen-

sus on a deﬁnition of explainability. In 2.1 and 2.2,

we will review some conceptual and formal popular

meaning and deﬁnition behind explainability.

Another open challenge in explainable artiﬁcial

intelligence is to quantify explainability. We believe

a standard metric will be difﬁcult to attempt until a

consensus has been made on a concrete deﬁnition of

explainability. However, some promising evaluation

metrics have emerged recently and will be presented

in 2.3.

2.1 Conceptual Deﬁnition of

Explainability

Although the goal of explainable artiﬁcial AI is clear

(making decisions or actions of AI systems – EDM

systems for example – explainable to a human ob-

server), there is currently no consensus on how to

deﬁne explainability. This is emphasized in a re-

cent paper (Miller, 2019), where the authors point

out the fact that most works in eXplainable AI are

based only on the researchers’ intuition of what con-

stitutes a good explanation. However, people select,

represent or comprehend explanations depending on

the discipline they are related to. From psychology,

according to (Lombrozo, 2006), explanations are the

currency in which we exchange beliefs. This def-

inition emphasizes the need for high conﬁdence in

explanation. More recently, (Miller, 2019) formu-

lates explainability as the degree to which a human

can understand the cause of a decision. In a simi-

lar vein, (Kim et al., 2018) deﬁne explainability as

the degree to which a human can consistently pre-

dict the model result. A deﬁnition of explainability

in relation to AI is given by (Islam et al., 2019) as

the extent of transferable qualitative understanding of

the relationship between model input and prediction

in a recipient-friendly manner. (Ribeiro et al., 2016)

link explainability to visual perception. The authors

of the paper argue that explaining a decision means

presenting visual artifacts that provide qualitative un-

derstanding of the relationship between the instance’s

components (e.g. words in text, patches in an image)

and the model’s prediction. We believe a concise deﬁ-

nition of XAI should take these rich and diverse view-

ing angle into consideration.

2.2 Formal Deﬁnition of Explainability

Consider a standard supervised learning problem – as

it is almost always the case in educational data min-

ing – and let denotes f : X 7→ Y , f ∈ F the super-

vised model trained with structured data-set D

train

, y

}

n=1

, which maps input feature x ∈ X to tar-

gets y ∈ Y . In this paper, we focus on local explana-

tion, which is the most common formal deﬁnition of

model interpretability. If F is complex, a local expla-

nation can be generated to understand its behavior in

some neighborhood N

∈ P[X ], where P[X ] denotes

the space of probability distribution over X . In ex-

plainable artiﬁcial intelligence (XAI) literature, sys-

tems that produce local explanations – local explain-

ers – are often denoted as σ : X ×F 7→ ξ, where ξ is

the set of possible explanations (Plumb et al., 2019).

It is worth mentioning that the choice of the expla-

nations system ξ greatly depends on whether or not

feature in D

train

are semantics. In XAI, features are

said to be semantic if one can reason about them and

understand the meaning of change in their values (e.g.

a student’s age, grade or number of siblings).

To understand the decision made by a data mining

model, the vast majority of local explainers try to pre-

KDIR 2022 - 14th International Conference on Knowledge Discovery and Information Retrieval

214

dict how the model’s output would change if its input

were perturbed to some degree (Ribeiro et al., 2016;

Plumb et al., 2018). The explainer output space can

then be formalized as ξ

:= {g ∈ G | g : X 7→ Y },

where G denotes a class of interpretable functions.

Since the vast majority of features encountered in

EDM literature is semantic, we believe integrating lo-

cal explainers into EDM methods should be a fairly

convenient and promising task.

2.3 Quantifying Explainability

While explainability tells the how and why regarding

the decision of an AI model, it does not tell us much

about how much trust we can have in that decision. As

a result, the question of how to quantify explainability

is still an open challenge. In recent years, some met-

rics have been proposed to quantify proposed explain-

ability systems. The rest of this section will present

two of the most popular metrics.

Fidelity Metric. The ﬁdelity metric evaluates ex-

plainability by measuring how accurately g models

f in a neighborhood N

(Ribeiro et al., 2016; Plumb

et al., 2019). It can be formally deﬁned as:

F( f , g, N

) := E

˜x∼N

[(g( ˜x) − f ( ˜x))

]. (1)

In this context, a lower ﬁdelity value (lower is bet-

ter) means that the explainability system is able to

accurately determine which patterns were relevant to

the model for making this prediction. In the context

of EDM, an explanation with good ﬁdelity accurately

identiﬁes which pattern in the data the model used to

predict the student grade or overall student academic

year achievement, for example.

Stability Metric. The stability metric has been pro-

posed by (Alvarez Melis and Jaakkola, 2018) and

consists of evaluating the degree to which the expla-

nation changes between points in the neighborhood

S( f , σ, N

) := E

˜x∼N

[ ||(σ(x, f ) −σ( ˜x, f ))||

]. (2)

In this setting, the lower the stability value (lower is

better), the more reliable and trustworthy the model

explainability.

In the next section, we will analyse explainabil-

ity in EDM. Doing so, we will give some suggestions

on how to integrate explainability system in EDM ap-

proaches where it is missing. For the evaluation of

such explainability system, the tools described in this

section can be used.

3 EXPLAINABILITY IN

EDUCATIONAL DATA MINING

Explainable AI (XAI) is a relatively new ﬁeld that has

not yet really penetrated the EDM ﬁeld, although it

would be of great importance in achieving the EDM

objectives. Generally, the objective of EDM is to im-

prove the educational system by assisting adminis-

trators and educators in decision-making tasks, like,

grades prediction (Bhopal and Myers, 2020), predict-

ing student dropout (Gardner et al., 2019), education

admission decisions (Marcinkowski et al., 2020) or

forecasting on-time graduation of students (Hutt et al.,

2019). In all these tasks, the current state of research

in EDM considers accuracy as the main metric to

evaluate the quality of a model. This allows to gain

great insight since accuracy is important. However,

EDM often infers a multi-target problem, therefore,

just focusing on accuracy is misleading and will not

be sufﬁcient to reach the goal. This accuracy short-

coming can be overcome by integrating explanability

in EDM approaches, as will be shown in this section.

The rest of this section will explore if and how

explainability is integrated into state-of-the-art educa-

tional data mining techniques. A summary of this ex-

ploration on the most prominent state-of-the-art EDM

techniques is given in Table 1. Before delving into

explainability in the current state of EDM research

landscape, let brieﬂy discuss the motivation for using

explainability in EDM.

3.1 Advantages of using Explainability

in EDM

To maintain the integrity of the data and validate the

ethical decision made by a model, explainable AI

should be used in EDM for deriving an understand-

ing of the important features that might be directly

effecting the students’ performance and help them

in choosing suitable courses of interest for example.

Furthermore, instead of focusing on the students’ per-

formance and result-oriented examinations, the con-

sideration and understanding of various attributes af-

fecting students’ performance and behavioural areas

is an important aspect for instructors. This is not a

trivial task since the educational dataset is dynamic,

as it consists of varying attributes such as location,

courses chosen, behavioural factors, historical perfor-

mance records and many more for each and every in-

dividual student. There is a massive necessity for sup-

porting students’ achievement, whether good or weak,

answering the question why and also discern the in-

formation for betterment of students’ performance.

Explainability also determines the model’s prediction

Towards Explainability in Modern Educational Data Mining: A Survey

215

strategy’s intrinsic bias (Gaur et al., 2020). Moreover,

explainability opens new research areas and ensures

decision-making based on domain knowledge.

In a nutshell, eXplainable Artiﬁcial Intelligence

(XAI) would not only provide the user with the per-

formance predictions but also necessitates the evalu-

ation of underlying features which play the medium

for reaching out to speciﬁc predictions.

3.2 EDM Approaches Adressing

Explainability

Recent research in educational data mining used at-

tention models to leverage a certain degree of in-

terpretability and visibility of words in sentences to

which the model predicts responses. This is of-

ten achieved by translating the model into responses

to questions where sentences are separated by OR

clauses. In (Gaur et al., 2020) for example, the

authors describe possible ways of infusing knowl-

edge graphs in deep learning (DL) models using

knowledge-aware loss function or knowledge-aware

propagation function. The former function can be

used to calculate the deviations in the information at

each and every layer of the DL model and for every

epoch during learning. The latter can interpret the

lost information and transfers the lost information us-

ing mathematical procedures. The paper describes a

use case on students’ academic performance with aca-

demic domain knowledge as knowledge graph. This

allows to infuse deep learning to the model for not

only making predictions but also estimate the incor-

rect decisiveness of meta concepts when solving a test

and provide us with the students’ knowledge. For ex-

ample, a student is asked to solve a problem encap-

sulating concepts of quadratic equations and physics,

the correctness to the answer explains his domain

knowledge in either one of the concepts or both. This

method provides a certain level of explanation of the

model decision making process.

Another paper by (Hasib et al., 2022) approaches

the explainability problem in EDM in a very different

way. The authors described the approach of utilizing

ﬁve different data mining algorithms such as Logis-

tic Regression, K-Nearest Neighbors (KNN), Support

Vector Machine (SVM), Naive Bayes and XGBoost

for classiﬁcation task to predict the accurate categor-

ical estimations of ﬁve classes (Excellent, Good, Sat-

isfactory, Poor and Failure) of high-school student

dataset on two courses. The proposed experiments

showcase an excellent accuracy in utilizing the SVM

approach and provide a comparative analysis of the

ﬁve algorithms used. To explain the integrity of pre-

dicted results, the author trained LIME (Local Inter-

pretable Model-agnostic Explanations), a popular lo-

cal explainer model, to understand the changes in the

data and the effect of these changes on the predic-

tion results for ﬁve classiﬁcation algorithms. How-

ever, LIME is better suited for image data-set and has

shown some limitations when applied to EDM data.

3.3 EDM Approaches Lacking

Explainability

The vast majority of approaches in educational data

mining concentrate on the accuracy of the model be-

ing deployed, therefore largely neglecting its explain-

ability. This result in EDM approaches performing

well in regard of their accuracy, but they are not ex-

plainable. This lack of explainability goes against re-

cent legislation like the EU’s ”General Data Protec-

tion Right” (GDPR), for example, which requires the

decision making process of an AI algorithm to be hu-

manly understandable.

One of such EDM technique exclusively focusing

on accuracy has been proposed in (Asif et al., 2017),

where the authors concentrated on predicting the aca-

demic achievements of students at the end of a four

year study program. This is a multi-target problem

since it includes predicting the student grade at each

course for example. The main ﬁnding of the paper

was the importance for the educators to concentrate

on a small number of courses showing particularly

good or poor performance. Doing so, support can

be offered to under-performing students as well as

advice or opportunities to high-performing students.

Their model which mainly built on decision tree and

random forest does not tackle the explainability is-

sue. However, this will be an interesting feature, es-

pecially in the case of a multi-target problem as de-

scribed in their paper. Fortunately, their method is

lightweight and integrating an explainability module

like EXpO (Explanation-based Optimization), which

has been presented in (Plumb et al., 2019), should be

fairly straightforward.

An approach with a similar objective has been

proposed in (Burgos et al., 2018). The goal here was

to predict the overall grade a student would get in the

subsequent semester. Here again, the approach fo-

cuses on the model prediction accuracy at the expense

of its explainability. Although this is a single-target

problem in contrast to the previously mentioned ap-

proach by (Asif et al., 2017), which was a multi tar-

get problem, understanding the algorithm’s decision-

making process would be a valuable asset. This

should be conveniently achievable with an explainer

like FTSD (Functional Transparency for Structured

Data) proposed in (Lee et al., 2019).

KDIR 2022 - 14th International Conference on Knowledge Discovery and Information Retrieval

216

Table 1: State-of-the-art Approaches for Educational Data Mining. For each method, we describe the task tackled, the data

mining methods used, and whether or not explainability has been addressed.

Methods Tasks Methods

Explainability

(Hasib et al., 2022)

Regression,

Classiﬁcation

KNN, XG-Boost, SVM, Logistic

Regression and Naive Bayes

√

(Hoffait and Schyns, 2017)

Classiﬁcation RF, LR, ANN

(Fernandes et al., 2019) Classiﬁcation

Gradient Boosting

(Gaur et al., 2020) Regression

Knowledge Graphs

infused DL model

√

(Cruz-Jesus et al., 2020) Classiﬁcation

DT, RF, SVM

(Toivonen and Jormanainen, 2019) Classiﬁcation DT ×

(Kaur et al., 2015) Regression Classiﬁcation Algorithms ×

(Asif et al., 2017) Regression DT, RF ×

(Waheed et al., 2020) Classiﬁcation ANN ×

(Xu et al., 2019) Regression DT, SVM ×

(Burgos et al., 2018) Regression SVM, FNN ×

(Rebai et al., 2020)

Regression,

Classiﬁcation

RT, RF ×

Another view of the EDM problem has been pro-

posed by (Waheed et al., 2020). They suggest a model

analyzing students’ records related to their naviga-

tion through a learning management system (LMS)

platform. The authors demonstrate that a student’s

clickstream activity has a signiﬁcant impact on its

performance. More precisely, students who navigate

through courses achieve better results. Their approach

which is based on deep neural networks is complex

and its decision is difﬁcult to explain or interpret. In-

corporating a neural network based local explainers

like RRR (Right for the Right Reason) (Shao et al.,

2021) – an explainer, which regularizes a black-box

model via a regularizer that involves a model’s expla-

nations – into their model could enhance the trustwor-

thiness of the results.

Tackling a similar problem as the previously men-

tioned approach, (Xu et al., 2019) proposed to model

the relationship between Internet usage behaviors and

academic performance of undergraduate students via

decision tree and SVM. The ﬁnal model therefore pre-

dicts student academic performance from Internet us-

age data. The authors found a positive correlation be-

tween academic performance and internet connection

frequency, whereas internet trafﬁc volume was neg-

atively correlated with academic performance. Al-

though they do not mention the explanation of the

model in their article, their method is already intu-

itive in itself and its result should be fairly explain-

able. Nonetheless, the integration of a fairly simple

explainer should not be a complicated task.

From a different perspective, (Rebai et al., 2020)

proposed to identify the key factors that impact

schools’ academic performance and to explore the re-

lationships between these factors. Their regression

tree model featured that factors, which are the most

correlated to high student performance are compe-

tition, school size, class size and parental pressure.

Their model does not provide any explanation on

its decision-making process. Moreover, because we

could not ﬁnd any implementation of their method,

it is difﬁcult to assess whether the integration of an

explainability module would be good achievable or a

relatively complex task.

Towards Explainability in Modern Educational Data Mining: A Survey

217

In a related move, a model trained with demo-

graphic characteristics of students of a federal district

of Brazil was proposed by (Fernandes et al., 2019).

The approach which aims at predicting student’s aca-

demic performance uses classiﬁcation models based

on the Gradient Boosting Machine (GBM). The re-

sults indicate that demographic features like neigh-

bourhood, school and age are crucial indicators of

student success or failure. In their paper, the au-

thors provide usecase to understand the functioning of

their model. However, understanding how the model

works do not help to understand the model decision

process when predicting a speciﬁc class. We sug-

gest the integration of a model agnostic method like

SHARP (Lundberg and Lee, 2017) to make their ap-

proach more interpretable.

A similar and more recent study based on stu-

dent’s demographic characteristics has been proposed

in (Cruz-Jesus et al., 2020), where the authors develop

a model able to predict the academic achievement of

public high school students in Portugal. The dataset

consisted of 16 demographics on 110, 627 students

among which gender, age, internet access, computer

possession or class attendance. The authors imple-

ment many variants of their approach using various

data mining techniques such as K-nearest neighbours,

logistic regression, support vector machine or artiﬁ-

cial neural networks. These implementations achieve

a student performance prediction accuracy ranging

from 51.2% to 81.1%. Interestingly, the variants of

their implementation resulting in low accuracies of-

fer a greater ﬂexibility regarding the integration of an

explainability module. On the other side, the imple-

mentations that demonstrate the best accuracies are

more complex and integrating an explainability mod-

ule without compromising the models accuracies is

not a trivial task. This difﬁculty is well illustrated

in their ANN model, which is a complex black box

model, thus making the integration of external com-

ponent difﬁcult. A technique that could be used to

infuse explainability into their neural network is the

one proposed in (Plumb et al., 2019), which regular-

izes an ANN for explanation quality at training time.

To wrap up the exploration and importance of ex-

plainability in EDM, Figure 2 compares the neighbor-

hood ﬁdelity metric (presented in 2.3) of the simple

FTSD explainer, which was particularly designed for

structured data, and the predictive mean square error

of several models trained on the UCI Educational Pro-

cess Mining (EPM) regression dataset.

0.15

0.2

0.25

0.3

0.2

0.4

0.6

MSE

FTSD Neighborhood Fidelity

SVM

ANN

Random

Forest

•

Decision

Trees



Linear

Regression

Figure 2: Neighborhood Fidelity of the FTSD explaina-

tions (lower is better) vs. predictive MSE of several model

trained on the UCI “EPM” regression dataset.

3.4 Discussion and Outlook

An empirical conclusion that can be made out of this

section is that when designing EDM approaches, the

educational data mining community should try to ﬁnd

a trade-off between prediction accuracy and model

explanation. Good reasons for this are on one side the

constantly growing demands and legislation for hu-

manly understandable decision of artiﬁcial intelligent

system and on the other side the easier debugging and

enhancement of the trustworthiness of those systems.

Improving the trustworthiness of EDM approaches by

making their decision process understandable is of

crucial importance since EDM decisions are applied

to humans, for example when deciding to accept or

reject an application for a study visa or an application

for a speciﬁc program.

In view of this, more attention should be paid on

providing EDM approaches that are explainable, even

at the accuracy cost. In designing these human under-

standable models, the EDM community could inspire

from the few literatures in the XAI community (Al-

varez Melis and Jaakkola, 2018; Plumb et al., 2019;

Al-Shedivat et al., 2020) that have thematized on this

accuracy-explainability trade-off.

4 CONCLUSION

In this work, we addressed the explainability problem

in educational data mining. We made several obser-

vations, both conceptual and empirical, which open

future research directions. In particular, we noticed

that explainability is not well addressed in EDM al-

though it would be of great importance in satisfying

regulations about artiﬁcial intelligent systems as well

as making EDM more trustworthy. Furthermore, we

took a look at the current state of explainable artiﬁ-

KDIR 2022 - 14th International Conference on Knowledge Discovery and Information Retrieval

218

cial intelligence (XAI) research both in terms of con-

ceptual formalism and evaluation metrics and propose

some XAI methods that could be incorporated into

speciﬁc state-of-the-art EDM approaches.

In the long horizon, achieving explainable EDM

would have a great impact on education and should

be among signiﬁcant goals for a healthier and trust-

worthy EDM practice. Our work is a foundational

research and does not lead to any direct applications.

ACKNOWLEDGEMENTS

This work is part of the Digital Mentoring project,

which is funded by the Stiftung Innovation in der

Hochschullehre under FBM2020-VA-219-2-05750.

REFERENCES

Al-Shedivat, M., Dubey, A., and Xing, E. P. (2020). Con-

textual explanation networks. J. Mach. Learn. Res.,

21:194–1.

Alvarez Melis, D. and Jaakkola, T. (2018). Towards robust

interpretability with self-explaining neural networks.

Advances in neural information processing systems,

31.

Asif, R., Merceron, A., Ali, S. A., and Haider, N. G. (2017).

Analyzing undergraduate students’ performance using

educational data mining. Computers & Education,

113:177–194.

Bhopal, K. and Myers, M. (2020). The impact of covid-19

on a level students in england.

Burgos, C., Campanario, M. L., de la Pe

na, D., Lara, J. A.,

Lizcano, D., and Mart

ınez, M. A. (2018). Data min-

ing for modeling students’ performance: A tutoring

action plan to prevent academic dropout. Computers

& Electrical Engineering, 66:541–556.

Burkart, N. and Huber, M. F. (2021). A survey on the ex-

plainability of supervised machine learning. Journal

of Artiﬁcial Intelligence Research, 70:245–317.

Calvet Li

an, L. and Juan P

erez,

A. A. (2015). Educa-

tional data mining and learning analytics: differences,

similarities, and time evolution. International Jour-

nal of Educational Technology in Higher Education,

12(3):98–112.

Cruz-Jesus, F., Castelli, M., Oliveira, T., Mendes, R.,

Nunes, C., Sa-Velho, M., and Rosa-Louro, A. (2020).

Using artiﬁcial intelligence methods to assess aca-

demic achievement in public high schools of a euro-

pean union country. Heliyon, 6(6):e04081.

Fernandes, E., Holanda, M., Victorino, M., Borges, V., Car-

valho, R., and Van Erven, G. (2019). Educational data

mining: Predictive analysis of academic performance

of public school students in the capital of brazil. Jour-

nal of Business Research, 94:335–343.

Gardner, J., Brooks, C., and Baker, R. (2019). Evaluating

the fairness of predictive student models through slic-

ing analysis. In Proceedings of the 9th international

conference on learning analytics & knowledge, pages

225–234.

Gaur, M., Faldu, K., and Sheth, A. (2020). Semantics of

the black-box: Can knowledge graphs help make deep

learning systems more interpretable and explainable?

Hasib, K. M., Rahman, F., Hasnat, R., and Alam, M. G. R.

(2022). A machine learning and explainable ai ap-

proach for predicting secondary school student per-

formance. In 2022 IEEE 12th Annual Computing and

Communication Workshop and Conference (CCWC),

pages 0399–0405.

Hoffait, A.-S. and Schyns, M. (2017). Early detection of

university students with potential difﬁculties. Deci-

sion Support Systems, 101:1–11.

Hutt, S., Gardner, M., Duckworth, A. L., and D’Mello,

S. K. (2019). Evaluating fairness and generalizability

in models predicting on-time graduation from college

applications. International Educational Data Mining

Society.

Islam, S. R., Eberle, W., and Ghafoor, S. K. (2019).

Towards quantiﬁcation of explainability in explain-

able artiﬁcial intelligence methods. arXiv preprint

arXiv:1911.10104.

Kaur, P., Singh, M., and Josan, G. S. (2015). Classiﬁcation

and prediction based data mining algorithms to predict

slow learners in education sector. Procedia Computer

Science, 57:500–508. 3rd International Conference on

Recent Trends in Computing 2015 (ICRTC-2015).

Kim, B., Wattenberg, M., Gilmer, J., Cai, C., Wexler, J.,

Viegas, F., et al. (2018). Interpretability beyond fea-

ture attribution: Quantitative testing with concept ac-

tivation vectors (tcav). In International conference on

machine learning, pages 2668–2677. PMLR.

Lee, G.-H., Jin, W., Alvarez-Melis, D., and Jaakkola, T.

(2019). Functional transparency for structured data: a

game-theoretic approach. In International Conference

on Machine Learning, pages 3723–3733. PMLR.

Lombrozo, T. (2006). The structure and function of ex-

planations. Trends in cognitive sciences, 10(10):464–

470.

Lundberg, S. M. and Lee, S.-I. (2017). A uniﬁed approach

to interpreting model predictions. Advances in neural

information processing systems, 30.

Marcinkowski, F., Kieslich, K., Starke, C., and L

unich, M.

(2020). Implications of ai (un-) fairness in higher ed-

ucation admissions: the effects of perceived ai (un-)

fairness on exit, voice and organizational reputation.

In Proceedings of the 2020 conference on fairness, ac-

countability, and transparency, pages 122–130.

Miller, T. (2019). Explanation in artiﬁcial intelligence: In-

sights from the social sciences. Artiﬁcial intelligence,

267:1–38.

Plumb, G., Al-Shedivat, M., Cabrera, A. A., Perer, A.,

Xing, E., and Talwalkar, A. (2019). Regularizing

black-box models for improved interpretability. arXiv

preprint arXiv:1902.06787.

Towards Explainability in Modern Educational Data Mining: A Survey

219

Plumb, G., Molitor, D., and Talwalkar, A. S. (2018). Model

agnostic supervised local explanations. Advances in

neural information processing systems, 31.

Rebai, S., Yahia, F. B., and Essid, H. (2020). A graph-

ically based machine learning approach to predict

secondary schools performance in tunisia. Socio-

Economic Planning Sciences, 70:100724.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). ” why

should i trust you?” explaining the predictions of any

classiﬁer. In Proceedings of the 22nd ACM SIGKDD

international conference on knowledge discovery and

data mining, pages 1135–1144.

Shao, X., Skryagin, A., Stammer, W., Schramowski, P., and

Kersting, K. (2021). Right for better reasons: Training

differentiable models by constraining their inﬂuence

functions. In Proceedings of the AAAI Conference on

Artiﬁcial Intelligence, volume 35, pages 9533–9540.

Silva, C. and Fonseca, J. (2017). Educational data mining: a

literature review. Europe and MENA Cooperation Ad-

vances in Information and Communication Technolo-

gies, pages 87–94.

Toivonen, T. and Jormanainen, I. (2019). Evolution of de-

cision tree classiﬁers in open ended educational data

mining. In Proceedings of the Seventh International

Conference on Technological Ecosystems for Enhanc-

ing Multiculturality, pages 290–296.

Waheed, H., Hassan, S.-U., Aljohani, N. R., Hardman, J.,

Alelyani, S., and Nawaz, R. (2020). Predicting aca-

demic performance of students from vle big data using

deep learning models. Computers in Human behavior,

104:106189.

Xu, X., Wang, J., Peng, H., and Wu, R. (2019). Predic-

tion of academic performance associated with internet

usage behaviors using machine learning algorithms.

Computers in Human Behavior, 98:166–173.

KDIR 2022 - 14th International Conference on Knowledge Discovery and Information Retrieval

220