A Dynamic Indicator to Model Students’ Digital Behavior

Oriane Dermy, Anne Boyer and Azim Roussanaly

Loria, Université de Lorraine, Nancy, France

Keywords:

Students’ Digital Behavior, Time-interval Pattern Mining, Sequential Pattern Mining, Change Mining, Pattern

Histories, COVID-19.

Abstract:

During the ﬁrst French Covid19 lockdown, students had to switch to a fully online learning mode. Therefore,

understanding students’ digital behavior becomes crucial for analysts serving public institution policy. In

particular, they want to determine and interpret the evolution of students’ digital behavior. This paper aims to

offer them indicators. We propose to study generic student logs corresponding to standard digital workspace

services. Therefore, this paper contributes to the scientiﬁc question: Can we give an easy-to-interpret and

visual indicator to model students’ behavior changes from poor and generic data? We ﬁrst verify that we

can extract epidemic-speciﬁc temporal patterns on these logs using Contrast Mining. These patterns represent

students’ behaviors and pace. Then, we propose a new method called Temporal Pattern Histories (TPH),

representing the evolution of the temporal patterns’ over time. It is a dynamic representation of students’

digital behavior. Using this method, we present graphically abrupt changes during the Cov19 lockdown, and

we give some hypotheses about these results. This case study proves the relevance of TPH to detect and

analyze students’ behavioral changes in an interpretive way. This approach has the advantage of representing

the global evolution of students’ behavior without giving students speciﬁc information.

1 INTRODUCTION

From March 2 to August 31, 2020, the French gov-

ernment reported that 25 million people were affected

by COVID-19 (Cov19), with approximately 119,500

people being hospitalized. To cope with this epi-

demic, the government imposed a lockdown from

March 17 to May 20, 2020.

Students had to adapt their learning methods to at-

tend distance learning courses abruptly. This online

education might have changed students’ digital be-

haviors. Many research studies have then used ques-

tionnaires and surveys to analyze and understand how

students and teachers perceive these changes. How-

ever, only a few research works focus on detecting

and understanding these changes, which is what an-

alysts need. Analysts are educational actors working

on what actions (e-service development, speciﬁc sup-

port to local lockdown, etc.) take and how to measure

their impact on the students’ behavior, as presented in

Fig 1. They need indicators to detect and analyze stu-

dents’ digital behavior changes in their Digital Work-

ing Environment. Such indicators must be easily un-

derstandable, interpretive, and respect the anonymity

of the students. As Cov19 was not predictable, no

speciﬁc data has been collected to determine such in-

dicators. It requires working on existing data, mainly

generic (as we want to have a global representation

of the impact of Cov19 on students’ digital behavior)

and poor. Thus, this paper aims at answering the sci-

entiﬁc question: Can we give an easy-to-interpret and

visual indicator to model students’ behavior changes

from poor and generic data?

This paper contributes to this research question

with a model,called Temporal Pattern Histories

(TPH), which represents the evolution of the support

of temporal patterns over time. The contributions of

this paper are (1) Validation of the use of generic and

"poor" data to identify students’ behavioral changes;

(2) Creation of the TPH graphical Data Mining (DM)

method; and (3) Application to the Cov19 case study

to represent its impact on students’ digital behavior.

This paper is organized as follows. Sec. 2 presents

the Cov19 research on students’ digital behavior.

Sec. 3 describes the studied dataset. Then, Sec. 4 de-

velops the ﬁrst approach with a static DM, followed

by the second approach with dynamic DM (Sec. 5).

Finally, Sec. 6 and Sec. 7 present discussions about

the results, conclusions, and perspectives.

Dermy, O., Boyer, A. and Roussanaly, A.

A Dynamic Indicator to Model Students’ Digital Behavior.

DOI: 10.5220/0011039400003182

In Proceedings of the 14th International Conference on Computer Suppor ted Education (CSEDU 2022) - Volume 2, pages 163-170

ISBN: 978-989-758-562-3; ISSN: 2184-5026

163

Data

sep

nov

jan

mar

may

jul

support

200

400

homework -mn- mark

lockdown

students' logs

(service labels & timestamp)

Learning experts needs

- COVID-19 impact

- Workspace service update

- School policy change

Analysis & Detection

of students'

behavioral changes

Context

[...]

Indicator that we propose

Figure 1: Scientiﬁc contribution: A new indicator helping analysts serving the public education policy.

2 LITERATURE REVIEW

The ﬁrst research papers on the impact of the Cov19

epidemic focus on subjective data, interviewing and

analyzing what students and experts think about on-

line learning, either verbally or through surveys.

Thanks to these interviews, researchers identify lim-

itations, critical challenges, or factors that facilitate

the use of e-learning (Abbasi et al., 2020; Almaiah

et al., 2020; Radha et al., 2020). In (Photopoulos

et al., 2021), the author goes further to discuss the

economic and political impact of e-learning develop-

ment.

DM methods began to be used in (Reigal et al.,

2020) with K-Means and clustering to identify

changes in users’ habits of a psycho-social assessment

platform.

Then, some researchers quantitatively analyzed

e-learning usage from more objective data, with a

glance at the dynamic evolution of students’ behav-

ior. In (Favale et al., 2020) the authors study stu-

dents’ learning and campus network usage during an

Italian lockdown. They showed that teams or private

chat messages, calls, and meetings increase notice-

ably during the lockdown. In (Ebner et al., 2020),

they analyze the number of online activities, clicks,

new publications, and hosts and participants at video

conferences. All these activities are studied over pre-

deﬁned sets of periods. Both papers also use dynamic

indicators of students’ behavior without DM methods.

Other works are closer to our paper, as they use

objective data and DM methods. In (Ilieva et al.,

2021), data comes from different countries and cor-

responds to different pandemic periods. They use

many methods (classical statistics, machine learning)

to cluster students and give statistics about their be-

havior. In (Lau et al., 2021), researchers adapted

the school policy to the Cov19 period, using a re-

vised bloom’s taxonomy with ﬂipped classrooms, vir-

tual classroom activities. Then, they analyzed the ef-

ﬁciency of this policy change using machine learn-

ing and students’ attendance, engagement, and global

scores data.

The speciﬁcity of our work is that: a) it uses

generic, objective and poor data; b) it models dynamic

of students’ behavior; c) it uses DM methods; and d)

our new method doesn’t require predeﬁned datasets.

Thus, this research work completes previous ones by

going further with a dynamic representation of stu-

dents’ behavior using generic and objective data and

DM methods.

3 DATASET DESCRIPTION

The experiments below are carried out using real data

corresponding to students’ logs when using their dig-

ital workspace. We had the opportunity to mine

data that has been collected from different digital

workspaces, initially to monitor e-services. It rep-

resents their usage from different middle and high

schools all over France. These digital workspaces per-

form much the same functionality as LMS but only

support face-to-face courses and not fully online ones.

As a result, collected data are less rich than in LMS.

This paper focuses on data collected during the

2018-19 and 2019-20 school years (from September

to August) and on academic institutions located in

two distant areas in France. The two chosen areas

have different granularities. They model students’

behavior independently of the area and granularity.

Thus, we will model students’ behavior and not ed-

ucational school policies. The ﬁrst area is a "departe-

ment", a territorial division in France, and the second

a city.

We represent the evolution of students’ digital be-

haviors monthly to make abrupt changes more visi-

ble and normalize results without altering them with

temporary disruptions. Data are pre-processed and

recorded as temporal or sequential sequences of stu-

dents’ activities. There is one sequence of activities

per month and student. A temporal sequence:

i,m

= h(t

, E

), (t

, E

), . .. , (t

, E

i,m

represents the activities of a student i who clicked on

n identiﬁed workspace services during a month m,

CSEDU 2022 - 14th International Conference on Computer Supported Education

164

where (t

, E

) are the identiﬁed service type E

and its

timestamp (i.e., when the student clicks on it). The

different services are:

«upload; collaborative work (collab); mark; mail;

absence; homework (hwk); pedagogical Itinerary

(pedagIt); school life (SchooL);

time management (timeM) ».

The ﬁrst approach intends to visualize the emerg-

ing / vanishing patterns speciﬁc to the lockdown. This

study would allow analysts to understand students’

behavioral changes. This experiment focuses on the

city and we create 4 sub-datasets (before Cov19, lock-

down, and same periods previous year):

- D

bef

: Before Cov19, 01/11-24/12 2019 (53days) -

cov

: lockdown, 17/03-10/05 2020 (54 days)

- D

bef,18

: 01/11/18 - 24/12/18 period (53 days) -

cov,18

: 17/03/19 - 10/05/19 period (54 days) We col-

lected 20.000 students’ logs per dataset.

The second experiment consists of visualizing and

analyzing the history of s/ti-patterns (i.e., sequential

and temporal patterns respectively). We worked on

10.000 logs per month, when enough logs were avail-

able, for the 4 datasets:

• D

c18

, city, 2018, 111.992

• D

c19

, city, 2019, 120.000

• D

d18

, departement, 2018, 120.000

• D

d19

, departement, 2019, 120.000

3.1 Algorithms Selection

Pattern mining (PM) methods are numerous and solve

different challenges, as presented in (Han et al.,

2012). We choose methods that mine sequential

databases (our data type).

3.1.1 Sequential Pattern Mining (SPM)

SPM has been extensively used in Educational Data

Mining (Anjum and Badugu, 2020). With the same

objective than us, the authors in (Gutierrez-Santos

et al., 2010; Poon et al., 2017) want to identify fre-

quent patterns of students’ activities. This ﬁeld of re-

search aims to discover frequent s-patterns (Agrawal

and Srikant, 1995). A sequential database is com-

posed of a set of sequences s="E

. . .E

", with E

∈

E being the set of ordered events. An s-pattern is a

sub-sequence, say p = "E

" contained in at least

k sequences (k is the minimum support

). In this pa-

per, we select the PreﬁxSpan method, which is one

The support is the sequence number in which the pat-

tern appears.

of the most efﬁcient and commonly used algorithms,

based on the pattern-growth method (Pei et al., 2001).

However, we also want to use the timestamping of

logs to improve patterns.

3.1.2 Temporal Pattern Mining (TI-PM)

Other approaches add to classical s-patterns contex-

tual information (Wang et al., 2018; Dong et al., ).

In our case, we use logs’ timestamps. Many meth-

ods add temporal information to discovered patterns,

as reviewed in (Dermy and Brun, 2020). The au-

thors’ conclusions show that to model students’ be-

havior and pace, we can complete s-patterns with ti-

patterns using the TI-PM method. This takes into

account gap values between events and groups them

in predetermined time intervals. A ti-pattern is de-

ﬁned as: α ="E

. . .E

", where E

∈ E is the set

of events for 1 ≤ i ≤ l and τ

∈ T I the set of time-

intervals. The sequence α is a time-interval pattern if

support(α) >= δ. Two algorithms have been devel-

oped by (Chen et al., 2003) to mine ti-patterns. For

our study, we select the I-PreﬁxSpan algorithm.

TI-PM and SPM have the disadvantage of stat-

ically modeling students’ behavior. We decide to

model students’ behavior dynamically to represent the

evolution of s/ti-patterns as a function of time, thanks

to Contrast and change DM.

4 CONTRAST DM APPROACH

Contrast DM methods aim to ﬁnd "contrast patterns"

describing signiﬁcant differences in patterns found

between datasets. Datasets differ temporally, locally,

or through different contrasting conditions (e.g., user

groups). Many algorithms exist for Decision Trees,

Clustering, or PM (Boettcher, 2011).

This section aims to answer the research question:

Can We Automatically Extract from Poor and

Generic Data Students’ Digital Behavior Speciﬁc to

Cov19?

Proposal: Dataset Comparison using Contrast

TI-Pattern Mining.

To provide an answer, we statically compare the re-

sults of DM performed on the pre-lockdown dataset

bef

) with the one recovered during the lockdown

cov

). To differentiate changes related to Cov19 with

the ones related to the school period, we compare

the same periods during the previous year (D

bef,18

and D

cov,18

). The comparison focuses on ti/s-pattern

changes detected in students’ logs. We also check

if Cov19 impacts students’ pace by calculating, for

A Dynamic Indicator to Model Students’ Digital Behavior

165

each time interval, the proportion of frequent patterns

which contain this time interval.

These proportions are compared across datasets.

The following experiment intends to discover stu-

dents’ behaviors speciﬁc to the lockdown: patterns

from the D

cov

dataset should have more differences

than between the other datasets.

4.1 Experiment

Experiment: Research of Students’ Digital Behav-

iors Speciﬁc to the Lockdown.

We perform the following experiment, with a min-

imum support variable equal to 30 and using the

PreﬁxSpan and TI-PreﬁxSpan algorithms (Gao et al.,

2008; Fournier-Viger et al., 2016). This experi-

ment statically compares ti-patterns of each dataset

(c.f., Table 1). For the sake of clarity, this Table 1

presents only some representative patterns (most fre-

quent/longer patterns), but changes between other

patterns for all datasets follow the same trend. Results

are presented as follows: each cell comprises the two

most frequent or longer patterns, followed by ":" and

its support (e.g., "mark mn mark:2014"). Long and

repetitive patterns (rows 4 and 5) are abbreviated with

dots (e.g., "mail mn mail mn . . . mn mail:38 (s7)",

where "(s7)" is the pattern-size). Students’ pace is

approximated in the 6

row (proportion of time inter-

vals in the ti-patterns). The last row presents statistics

about each service page.

Results: Yes, We Can Extract Students’ Digital Be-

havior Speciﬁc to the Covid Period.

For each dataset, we ﬁrst notice that for each frequent

temporal pattern "E

", its symmetric "E

" is

also frequent with similar support. It is visible in Ta-

ble 1 by patterns presented in row 3, and it is also

valid for other patterns. For example, the symmetric

of "mail mn Ped.It:495" exists with similar support:

"Ped.It mn mail:402". Thus, the service order doesn’t

seem important for students even if the pace (here I

)

is. During Cov19, students’ digital behavior is sud-

denly changing:

1. The most frequent pattern goes from "mark

mn mark" (3 other datasets) to "mail mn mail", and

the support of "mark mn mark" became 210, which is

much smaller (not shown there). The lockdown might

cause these changes since teachers communicated by

mail. Moreover, analysts interpret and conﬁrm that

teachers decided to give no marks to students during

the lockdown. The second more important pattern

is "mail mn mail mn mail", which conﬁrms that se-

quence of mails is really frequent. It could suggest

to analysts that i) students need to check their mails

often to follow instruction mails sent by teachers; ii)

students were a lot stressed and clicked many times

on mails. This second hypothesis is rejected because

students’ pace (row 6) didn’t accelerate.

2. The most frequent patterns without duplicates

go from "mark mn hwk" and its symmetric to "mail

mn Ped.It" and "hwk mn mail". This result conﬁrms

that students have few marks during this lockdown. It

might also be because, when students began to work

in a fully online mode, teachers give instruction to use

the Pedagogic Itinerary service by mails with some

additional information and because students return

their homework by mails.

3. The two longest patterns without duplicates

highlight again that the Pedagogical Itinerary service

is a lot used during Cov19.

4. Regarding students’ pace (row 6), the com-

parison of the Cov19 period with others shows that

the proportion of "second" interval decreases and the

"hour" one increases. About the hour interval, it

might suggest that students check services between

each online course (one or more hours). The decrease

of the "second" interval may be because schools’ In-

ternet connections are sometimes saturated, forcing

students to click on the same pages several times.

For each result, we gave some reasons that could

explain students’ behavioral change. However, they

require validation by analysts. To highlight the in-

crease of pattern change during Cov19, we study the

percentage of "similar s/ti-patterns" between D

bef,18

and D

cov,18

and that between D

bef

and D

cov

We consider that two patterns (p1 and p2) from two

different datasets are similar if (p1 = p2) and

|support (p1) − support (p2)|

support(p1)+support(p2)

< 0.5

Table 2 shows that the percentage of "similar" ti-

patterns is littler between D

bef

and D

cov

than between

bef,18

and D

cov,18

. Thus, the change rate related to

Cov19 is more signiﬁcant than we could expect from

2018.

5 CHANGE DM APPROACH

The previous Contrast Mining approach only allows

comparing statically several databases that must be

pre-deﬁned. Thus they can neither detect qualitative

leap about students’ behavior

nor identify the dy-

namics of behavioral changes since dynamic analysis

A qualitative leap is when there is a change in behavior

signiﬁcant enough to consider that students have changed

their strategy.

CSEDU 2022 - 14th International Conference on Computer Supported Education

166

Table 1: Contrast ti-pattern mining.

City datastets

Dbefore18

01/11/18-24/12/18

Dbefore19

01/11/19-24/12/19

Dcov18

17/03/19–10/05/19

Dcov19

17/03/20–10/05/20

2 most frequent

mark mn mark:2014

hwk mn hwk:766

mark mn mark:1252

mail mn mail:684

mark mn mark:1476

mail mn mail:679

mail mn mail:1782

mail mn mail mn mail:678

2 most frequent

without duplicates

mark mn hwk:530

hwk mn mark:467

mark mn hwk:418

hwk mn mark:340

hwk mn hwk:651

mark mn hwk:417

mail mn Ped.It.:495

hwk mn mail:412

2 longuests

(and most

frequent)

mark mn mark [ ] …

[mn mark:49

mark mn mark wk [...]

mn mark:37

mark mn mark wk

[...] mn mark:36 (s6)

mark mn mark wk

[...] mn mark:35 (s6)

mark mn mark [ ] …

[mn mark:46 (s6)

mark mn mark [ ]…

[mn mark:53 (s5)

mail mn mail [...]

[mn mail:38 (s7)

mail mn mail [ ]…

[mn mail:63 (s6)

2 longuests

without duplicates

mark mn hwk mn mark:99

hwk mn mark mn hwk:69

mark mn hwk mn mark:89

mark mn mail mn mark:58

mark mn hwk mn mark:69

mark mn mail mn mark:61

mail mn Ped.It. mn mail:112

mail mn hwk mn mail:99

Proportion

of time

intervals

(%)

sec

hour

day

mth

7.1%

59.5%

8.0%

5.9%

12.9%

6.5%

6.2%

60.1%

8.6%

5.9%

12.7%

6.5%

9.4%

56.4%

7.2%

5.6%

9.0%

12.4%

6.8%

54.4%

12.3%

5.8%

9.7%

11.0%

Proportion

of service

page

usage

(% [nb])

upload

collab

mark

abs

schooL

timeM

hwk

mail

ressOL

PedaIt

1.7% [111]

5.2% [331]

65.3% [4193]

6.2% [395]

2.6% [165]

0.6% [38]

31.9% [2048]

26.3% [1690]

0.0% [1]

3.1% [201]

0.5% [33]

7.1% [438]

44.5% [2715]

3.8% [234]

9.5% [580]

0.2% [14]

30.6% [1871]

31.4% [1918]

5.7% [349]

27.1% [1655]

2.0% [120]

7.4% [448]

54.1% [3282]

6.4% [388]

2.5% [153]

0.4% [24]

30.0% [1816]

31.4% [1902]

0.0% [3]

22.5% [1364]

0.9% [49]

7.8% [439]

13.8% [771]

0.8% [46]

5.2% [292]

0.3% [19]

27.2% [1524]

64.4% [3605]

2.6% [144]

31.7% [1773]

One dataset per column. Rows 2 to 5 summarize the most representative patterns. Row 6 gives the proportion of time

intervals in all discovered patterns, followed by the dataset proportion of page usage per year (row 7). Minimum

support = 30.

Table 2: Percentage of ti/s-patterns contained both in peri-

ods before & during Cov19 during School year.

common D

bef18

− D

cov18

bef

− D

cov

ti-patterns 45.3% 3.4%

s-patterns 90.3% 16.6%

requires observing an event over time. Therefore,

we now integrate temporal information into our

algorithm. This section aims to answer the research

question: How to Represent the Dynamic Evolution

of the Students’ Digital Behavior in an Interpretive

Way?

Proposal: Using Pattern Change Histories.

Change Mining algorithms represent the evolution of

models made by different DM algorithms. (Boettcher,

2011) presents an overview of Change Mining meth-

ods. In our case, we are interested in analyzing the

temporal evolution of patterns’ supports, called "pat-

tern histories" (Wang, 2011; Chen et al., 2004). This

paper uses ti-pattern histories, which have never been

done before.

We want to represent the evolution of the most

frequent patterns that stay frequent during some

months. Such patterns are called "change patterns".

To discover them, we compute frequent s/ti-patterns

per month and select the frequent ones for at least

3 months. Finally, we plot TPH (graphical repre-

sentations of the support evolution of each change

pattern). On these graphs, we represent covid-19 year

c19

or D

d19

) and previous one (D

c18

or D

d18

) to

compare them. For this experiment, we use a relative

minimum support of 20 and the (TI)-PreﬁxSpan

algorithms to mine s/ti-patterns.

Experiment: Representation of the Dynamic Stu-

dents’ Digital Behavior.

To represent the dynamic evolution of students’

digital behavior in an interpretative way, we look at

the 10 most frequent s-patterns and ti-patterns for

each month and dataset. Then, we select the ones

that are change patterns. Finally, we represent the

s/ti-pattern change histories (resp. SPH and TPH)

where support histories are computed each month.

Results: SPH and TPH Give an Easy-to-Interpret

Representation.

Fig. 2 presents TPH appearing in all datasets:

c18

c19

d18

d19

}. For each represented TPH,

we have the corresponding SPH, with curves that

follow the same dynamic (to facilitate the readabil-

ity, we don’t represent them). Moreover, whatever

the dataset, frequent time-interval change patterns are

generally symmetrical.

For example, the ﬁrst graph presents the histo-

ries of "homework -mn- mail", but there also exist

A Dynamic Indicator to Model Students’ Digital Behavior

167

200

400

support

100

200

300

400

sep oct nov dec jan feb maraprmayjun jul aug

100

200

sep oct nov dec jan febmaraprmayjun jul aug

100

200

lockdown

c18

c19

d18

d19

symmetric*

lockdown

* Similar TPH for both

(a -I

- b) & (b -I

- a)

homework -mn- mail

homework -s- mail

homework -mn- mark

mail -mn- collab

Figure 2: Display of TPH that change over time similarly in the two study locations. Note: corresponding SPH (homework-

mail, homework-mark, or mail-collab) are similar (not displayed here).

graphs "mail -mn- homework", "mail-homework", and

"homework-mail", with a similar support dynamic.

Those graphs show that students often check the

homework, mail, and mark services successively

without a speciﬁc order, but with a particular pace: a

few minutes delay before changing services. In the

departement area, we even have SPH of size 3, based

on the "mark-homework" pattern and its symmetric,

which follow the same support’s trend (not presented

here). This shows how concerned the students were

about this sequence. Independently of the lockdown,

curves show some students’ behavioral change

between 2018 and 2019. For example, in the city on

11/2018, there was a spike in the use of "homework

-mn- mail" and "homework -mn- mark" patterns that

don’t appear in 2019. However, these changes are

lighter than the ones during the lockdown.

The Comparison of the Four Datasets Suggests

That during the Lockdown:

(1) Students brutally stopped using mark and home-

work services successively, within the hour. This may

be because, during the lockdown, students weren’t

graded, as analysts told us.

(2) Students tend to use homework and mail ser-

vices more often (Fig. 2, left), and they move more

and more quickly from one service to another (bot-

tom left).

(3) The pattern "mail -mn- collab" (bottom-right)

has a spike of activities in July 2020 for both areas.

This might correspond to a ﬁnal collaborative project

linked to Cov19. Speciﬁc to the departement area,

2019 School year, this pattern appears in February,

just before the lockdown, with a little support (around

25). This same pattern remains frequent about the city

area during the whole 2018 School year, which seems

to correspond to a learning method based on projects.

This behavior is still present during the 2019 School

year, with a lower frequency.

The interpretation of these results is not the focus

of this paper. Our goal is to provide this new indicator

to analysts. This indicator can also be helpful for

teachers’ dashboards.

Patterns Speciﬁc to Students of the City (c.f.,

Fig. 3), Give Other Information:

(1) They often check mail (or pedagogical Itinerary)

with mark services successively, without temporal

regularity. Since the lockdown, this behavior is less

frequent than during the previous year. pedagogical

Itinerary - mark pattern ends up disappearing. Again,

this might be because students have no marks during

the outbreak.

(2) During the 2019 School year, they often per-

form the "mail -mn- pedagIt" pattern, with this spe-

ciﬁc order and with a regular gap. This behavior

might result from a policy of one or more educational

institutions in the city. The lockdown reinforces this

behavior since the support of this ti-pattern increases

from 100 (February) to around 275 (May).

However, we gave some clues to explain students’

behavioral changes. These results validate the ability

of TPH graphs to visually detect signiﬁcant students’

behavioral changes that correspond to the lockdown

period. Some of these behavioral changes are com-

mon to students located in the two analyzed areas

of France, and others are location-speciﬁc. Finally,

we saw that the lockdown also impact students’ pace.

Thus, we can conﬁrm that SPH and TPH can model

the dynamic of students’ behavioral changes in an in-

terpretative way.

CSEDU 2022 - 14th International Conference on Computer Supported Education

168

c18

c19

symmetric

100

150

200

pedagIt - mark

Lockdown

Support

sep

nov

jan

mar

may

jul

Lockdown

100

200

300

Support

mail -mn- pedagIt

sep

nov

jan

mar

may

jul

200

400

mail - mark

Support

sep

nov

jan

mar

may

jul

Lockdown

Figure 3: Some TPH and SPH speciﬁc to city datasets. Dot lines represent the baseline (2018-19 School year), not impacted

by the outbreak. A change history aI

b is said "Symmetric" if the change history bI

a is similar.

6 DISCUSSION

The ﬁrst experiment validates the relevance of data to

detect an abrupt change in students’ digital behavior

caused by Cov19. It shows that students use the mail

service a lot more during the lockdown and are more

likely to follow an hour space gap between explored

services, probably because students use e-services be-

tween classes. However, the most classic gap between

services remains the second gap. Results highlight

that the order in which the students use the different

services seems irrelevant: s/ti-patterns are "symmetri-

cal" (Sec.4.1). To conclude this experiment, we show

there are more ti/s-pattern changes during the lock-

down than in the previous year (baseline). It would

be beneﬁcial to evaluate the amount of pattern change

between datasets based on contrast-sets speciﬁc mea-

sures in future works, as proposed in (Magalhães and

Azevedo, 2015).

The method of the 2

experiment allows to dy-

namically represent students’ behavioral changes, vi-

sually and understandably, intending to facilitate the

analysis of the evolution of students’ digital behavior.

This method highlights the lockdown impact on stu-

dents’ digital behavior, allowing analysts to interpret

it. The results highlight a qualitative leap in students’

behavior during the lockdown, even without a-priori

knowledge. So, this approach enables analysts to de-

tect important students’ e-learning changes. More-

over, since this graphical method allows for overlap-

ping different pattern histories, we can easily compare

the dynamic of students’ behaviors, which is efﬁcient

to perform analysis. Hence, results show that some

patterns are more and more followed by students dur-

ing the lockdown (e.g., "homework-s-mail"), while

others stop being followed suddenly by students (e.g.,

"homework-mn-mark").

7 CONCLUSION AND FUTURE

WORKS

This paper proposes two approaches allowing an-

alysts to detect and analyze students’ behavioral

changes. They have been experimented on the Cov19

case study. Considering generic, objective, and poor

data, we succeed in detecting a global trend and visu-

ally representing the dynamic of students’ behavior in

an easy-to-interpret way, thanks to the new approach:

TPH. Thus, we answer the global research question

and extend the Cov19 state-of-the-art research.

The 1

approach makes a temporal pattern com-

parison between datasets. It is a static comparison

(pre-deﬁned datasets), useful for learning experts to

compare a speciﬁc period with others. In this 1

study, we explore students’ behavioral changes during

the lockdown with other temporal periods, thanks to

Contrast Mining. We discovered the emergence and

vanishing of some temporal patterns and the modiﬁ-

cation of students’ pace, and this, only with e-services

data information.

The 2

approach allows analysts to detect and in-

terpret the dynamic of students’ behavioral changes,

thanks to TPH. This new method successfully rep-

resents clearly, thanks to graph representations, the

dynamic of students’ behavior. On these graphs,

the Cov19 impact was visible even for non-experts.

These methods succeed in representing the general

trend of a group of students rather than targeting a

speciﬁc student.

In the future, we will apply these methods to sup-

port teachers. We also want to work with analysts to

analyze deeper the Cov19 impact on students’ digi-

tal behavior and to have clues to improve and create

dynamical indicators. We will ﬁnally search methods

that automatically detect signiﬁcant students’ behav-

ior change periods.

A Dynamic Indicator to Model Students’ Digital Behavior

169

REFERENCES

Abbasi, S., yoob, T., Malik, A., and Memon, S. I. (2020).

Perceptions of students regarding e-learning during

covid-19 at a private medical college. Pakistan J. of

Med. Sci.

Agrawal, R. and Srikant, R. (1995). Mining sequential pat-

terns. In Yu, P. S. and Chen, A. L. P., editors, Proc.

11th Int. Conf. on Data Eng., Taipei. IEEE Comp. Soc.

Almaiah, M. A., Al-Khasawneh, A., and Althunibat, A.

(2020). Exploring the critical challenges and fac-

tors inﬂuencing the e-learning system usage during

COVID-19 pandemic. Educ. Inf. Technol.

Anjum, N. and Badugu, S. (2020). A study of different

techniques in EDM. In ICETE. Springer, online.

Boettcher, M. (2011). Contrast and change mining. Wiley

Int. Reviews: DM & KD, 1(3):215–230.

Chen, C. F., Yang, S.-H., Falsaﬁ, B., and Moshovos, A.

(2004). Accurate and complexity-effective spatial pat-

tern prediction. In HPCA, Madrid. IEEE.

Chen, Y.-L., Chiang, M.-C., and Ko, M.-T. (2003).

Discovering time-interval sequential patterns in se-

quence databases. Expert Systems with Applications,

25(3):343–354.

Dermy, O. and Brun, A. (2020). Can we take advantage

of time-interval pattern mining to model students ac-

tivity? In Proc. 13th Int. Conf. on EDM, online. Int.

EDM Soc.

Dong, W., Fan, W., Shi, L., Zhou, C., and Yan, X. A gen-

eral framework to encode heterogeneous information

sources for contextual pattern mining. In Proc. ACM

Int. CIKM, Maui.

Ebner, M., Schön, S., Braun, C., Ebner, M., Grigoriadis, Y.,

Haas, M., Leitner, P., and Taraghi, B. (2020). Covid-

19 epidemic as e-learning boost? Future Internet,

12(6):94.

Favale, T., Soro, F., Trevisan, M., Drago, I., and Mellia, M.

(2020). Campus trafﬁc and e-learning during covid-19

pandemic. Computer Networks.

Fournier-Viger, P., Lin, J. C.-W., Gomariz, A., Gueniche,

T., Soltani, A., Deng, Z., and Lam, H. (2016). The

spmf open-source data mining library version 2. In

Joint European conf. on ML and KD in DB. Springer.

Gao, C., Wang, J., He, Y., and Zhou, L. (2008). Efﬁcient

mining of frequent sequence generators. In Huai, J.,

Chen, R., Hon, H., Liu, Y., Ma, W., Tomkins, A.,

and Zhang, X., editors, Proc. of the 17th Int. Conf.

on WWW. ACM.

Gutierrez-Santos, S., Mavrikis, M., and Magoulas, G.

(2010). Sequence detection for adaptive feedback

generation in an exploratory environment for mathe-

matical generalisation. In AIMSA, Varna. Springer.

Han, J., Kamber, M., and Pei, J. (2012). 7 - advanced pat-

tern mining. In Data Mining, The Morgan Kaufmann

Series in DMS. Boston, 3rd edition.

Ilieva, G., Yankova, T., Klisarova-Belcheva, S., and

Ivanova, S. (2021). Effects of covid-19 pandemic on

university students’ learning. Information.

Lau, E., Chai, K. K., Goteng, G. L., and Wijeratne, V.

(2021). A neural network modelling and prediction of

students’ progression in learning: A hybrid pedagogic

method. In CSEDU.

Magalhães, A. and Azevedo, P. J. (2015). Contrast set min-

ing in temporal databases. Expert systems.

Pei, J., Han, J., Mortazavi-Asl, B., Pinto, H., Chen, Q.,

Dayal, U., and Hsu, M.-C. (2001). Preﬁxspan: Min-

ing SP efﬁciently by preﬁx-projected pattern growth.

In Proc. 17th Data Engineering, Heidelberg. IEEE.

Photopoulos, P., Stavrakas, I., and Triantis, D. (2021).

Post-covid-19 education: A case of technology driven

change? In CSEDU.

Poon, L. K., Kong, S.-C., Wong, M. Y., and Yau, T. S.

(2017). Mining sequential patterns of students’ access

on learning management system. In DM & big data,

Fukuoka. Springer.

Radha, R., Mahalakshmi, K., Kumar, V. S., and Sara-

vanakumar, A. (2020). E-learning during lockdown

of covid-19 pandemic: A global perspective. Int. jour.

of control & automation.

Reigal, R. E., Pastrana-Brincones, J. L., González-Ruiz,

S. L., Hernández-Mendo, A., Morillo-Baro, J. P., and

Morales-Sánchez, V. (2020). Use of data mining to

determine usage patterns of an online evaluation plat-

form during the covid-19 pandemic. Front. in Psycho.

Wang, J.-D. (2011). A novel approach to compute pattern

history for trend analysis. In 8th Int. FSKD, Shanghai.

IEEE.

Wang, R., Ji, W., Liu, M., Wang, X., Weng, J., Deng, S.,

Gao, S., and Yuan, C. (2018). Review on mining data

from multiple data sources. Pattern Reco. Letters.

CSEDU 2022 - 14th International Conference on Computer Supported Education

170