Towards Visual Explorations of Forums’ Collective Dynamics in

Learning Management Systems

Malik Kon

1,2

, Madeth May

, S

ebastien Iksal

and Souleyman Oumtanaga

LIUM, Le Mans Universit

e, Le Mans, France

LARIT, INP-HB, Yamoussoukro, C

ote d’Ivoire

Keywords: Visualization, Forum, Social Network, Learning Analytics.

Abstract:

Discussion and exchange among peers have being hailed as an essential part of learning since at least, Vygot-

sky’s socio-constructivist theory. There, learning is presented as a subtle and dynamical collective process.

Hence, despite numerous efforts to understand how learners engage and maintain inspiring discussions, re-

searchers continue to question how to effectively reinforce the collective actions. In Learning Management

Systems (LMSs) they propose Learning Dashboards (LDBs) to learners, tutors, and managers to help them

monitor various learning indicators. But only recently have they employed Natural Language Processings

(NLPs) and Social Network Analysis (SNA) techniques to display temporal indicators incorporating the fo-

rums’ content analysis and the learners’ behavioral patterns. In this study, we present our design efforts to

model a scalable and portable indicator of collective actions. We aim to support tutors’ monitoring of fo-

rums’ activities through explorable visualizations. We review previous researches about visual explorations

of Forums’ content and online collaboration’s measures. We expose in progress visualizations built from

three different datasets and propose directions towards further development of indicators to monitor collective

actions.

1 INTRODUCTION

Modern online learning environments often rely on

forums and chats for their students to exchange in-

formation about a subject, get help from peers and

tutors or socialize. When attendance is massive –as

currently in popular Massive Open Online Courses

(MOOCs)–, or if the course session spread over a long

period of time, the forums’ messages rapidly become

intractable. Therefore, the mass of information is de-

valued and can even be a burden on the learners and

tutors as the former struggle to ﬁnd adequate peers to

support learning collectively and the latter are over-

whelmed by numbers.

Efforts have been made to help tutors analyze the

students’ exchanges in forums but, as we will see in

section 3, these analyses are often limited to a small

set of individuals or are topic speciﬁc. Also, the anal-

ysis usually takes place at the end of the learning ses-

sion and does not allow “real-time” guidance.

Meanwhile, interactions taking place in fo-

rums account for essential aspects of the socio-

constructivist learning theory. Vygotsky emphasizes

the importance of discussions as they take learners

at the fringe of their knowledge domain, the proxi-

mal development zone. There, they push back their

knowledge barrier and build new knowledge to ﬁll

the gaps spotted during discussions. Therefore, fa-

cilitating forums’ usages in ways that encourage dis-

cussions and collective actions should theoretically be

beneﬁcial to the learning experience of all actors. But

as far as we know, there is yet no method to come

up with a generic and dynamic way to explore the

social interactions taking place in Learning Manage-

ment System (LMS). Learners often struggle to build

a sense of belonging (Khalil and Ebner, 2014) and

tutors have difﬁculties supporting effective collective

actions (Zheng et al., 2015).

Well-designed learning dashboard could help sup-

port collective actions by enabling the exploration of

the forums’ interactions. They would offer the pos-

sibility for the tutor and administrator to better un-

derstand how students and discussed topics evolve,

as well as facilitate the curriculum design improve-

ments. Learning Dashboards (LDBs) could also help

learners develop reﬂexive competencies and increase

their engagement.

In this paper, we report our ﬁrst step towards facil-

Koné, M., May, M., Iksal, S. and Oumtanaga, S.

Towards Visual Explorations of Forums’ Collective Dynamics in Learning Management Systems.

DOI: 10.5220/0007716000670078

In Proceedings of the 11th International Conference on Computer Supported Education (CSEDU 2019), pages 67-78

ISBN: 978-989-758-367-4

itating collective activities with tools supporting the

tutor in their visual exploration of the Virtual Univer-

sity of C

ote d’Ivoire (VUCI)’s forums. In the next

section, we deﬁne collective dynamic and explain

why we see it differently from collaboration. Then,

in section 3, we recall previous efforts made to mon-

itor collective actions and use LDBs as support tools.

In section 4, we present our model and propose a way

to come up with an interactive visualization (or ex-

plorable) of the forum’s collective dynamic. We illus-

trate it, in the 5

section with our ﬁrsts visualizations

built from different datasets that we will have detailed

in the preceding section. Finally, we discuss our vi-

sualizations’ limitations and future developments in

section 6.

2 VISUALIZATION OF

COLLECTIVE DYNAMICS IN

LMSs

The ﬁrst clariﬁcation we make is between visual-

ization and Learning Dashboard (LDB). A Learning

Dashboard is a “single display that aggregates dif-

ferent indicators about the learner learning process

and/or learning context into one or multiple visual-

izations” (Schwendimann et al., 2017). We will use

the term visualization when the LDB can be made of

single visualization and keep the term LDB when it

incorporates clearly several visualizations. We will

consider discussions taking place in LMSs or in other

online applications if they are used in an educational

context. We understand that LMSs are applications

designed with the intention of being teaching tools,

but sometimes we may refer to Google Hangouts as

a LMS too. If so, it is because we consider the spe-

ciﬁc Hangouts chat from the G Suite for Education

which was built with the intent to support teaching

and learning.

We collected 7 datasets from Coursera, Hangouts,

and Moodle. The datasets are different in size and

quality. Table 1 and Section 4 further details the

means of collection, the datasets speciﬁcities.

2.1 Actions in Forums

We make an important distinction between collab-

orative actions and collective actions. Dillenbourg

(1999) deﬁne collaboration as a coordinated syn-

chronous activity born from the persistent will to

share a common perception of a problem originat-

ing from people with similar social roles. Taken as

a bottom-up process with the coordination coming up

from the actors themselves, collaboration is difﬁcult

to automatically analyze. It implies that to coordi-

nate, each actor evaluates the intentions of others and,

doing so, each instantiates a theory of mind (Ger-

stenberg and Tenenbaum, 2017) that would be very

challenging to implement artiﬁcially. Therefore we

use the term “collective action” instead of “collabo-

ration” to emphasize the fact that we do not assume

a shared intention or shared goal in the actors’ inter-

actions. We focus on a set of observable actions and

leave the deduction of the interaction’s intent to the

observer. Nevertheless, we use the expression “col-

lective action” instead of the more generic “social in-

teraction” to bear in mind that from the observer point

of view, the studied interactions have a shared goal.

This is not true for all social interactions. So, to wrap

up, in a collaborative interaction the shared goal can

be made explicit by the actors, in a collective inter-

action, it is subjective to the observer, and in social

interactions, it may not exist at all.

The elementary actions we are most concerned

with are those taking place in forum and chats of

LMSs. In general, they are publications and mes-

sages’ comments, but messages’ up and down votes,

publication times are signiﬁcant and integrated into

the model that we elaborate in section 4.

Figure 1: Illustration of how the strength of actors’ ties

(or links) varies as a function of time and topic overlap.

Thread a corresponds to actor-topic dynamic (1a) where B’s

late post after A’s 1

publication does not correlate strongly

enough to create the link from B to A. But A’s 2

post is

timely enough, although not exactly on the same topic as

B’s message, to create the tie A 99K B drawn as a dashed

arrow. In thread b, in addition to the tie B 99K A, we have

a topic overlap and time proximity between C and A. This

makes a the strong tie A → C.

CSEDU 2019 - 11th International Conference on Computer Supported Education

Forums and Chat. Depending on the LMS, dif-

ferent tools exist to facilitate discussions between

peers. The main distinction is historical and separated

asynchronous tools, forums, from synchronous aimed

chats. Hence, forums’ hierarchical structure is of-

ten more ﬁne-grained than that of chats because their

posts were always meant to be persistent. On the other

hand, chats have better online awareness and presence

indicators.

In this study, we consider that a forum is made

of discussions and that each discussion is created

from an initial message followed by other messages.

This sequence of messages is what we call a dis-

cussion thread. A discussion may contain several

threads if comments are allowed or if explicit refer-

ences are made to previous publications. In that case,

the new thread will contain the referenced messages

up to the initial one and all subsequent comments. In

chats, several interrelated threads may also appear in

a unique discussion when one explicitly mentions a

previous message.

Despite their historical differences, today, both fo-

rum and chat can be used for synchronous and asyn-

chronous online discussions. One can subscribe to

a forum’s discussion and receive alerts as soon as a

new message is published. And a history of posts is

kept in modern chats, while the creation of several

co-existing chats has been facilitated. Each chat is,

then, equivalent to a forum discussion. Finally, both

forums and chats display messages’ timestamp and

authors’ information. So for simplicity, we will call

indifferently “forum” any virtual discussion space in

a LMS.

2.2 Collective Dynamics

Collective dynamics are time-dependent interactions

spurring for the messages co-occurrence in the fo-

rums’ discussions.

Actors’ Messages Dynamics. How do actors’ mes-

sages spread over time ? Let (τ

)

1,2,3

be three mes-

sages timestamp and A, B, C three actors. Col-

lectively, the actors’ messages could be distributed

as A

, B

, C

, meaning that A, B and C respectively

posted one message at timestamp τ

, τ

. But an-

other dynamic could be that actor A, alone, published

the three messages, thus A

, A

; or alternatively

that A published a message at τ

and B at timestamp τ

and τ

, thus A

, B

. These denote different dynam-

ics for the actors’ messages. Visualizing them helps to

identify the users posting behavior and distinguishing

learners’ behavior, for example, active from lurkers.

Topics’ Messages Dynamics. How do messages

spread over time and topics ? Or how topics are cov-

ered by messages over time ? Latent Dirichlet Al-

location (LDA) based methods are commonly used

Bayesian parametric methods to approximate a mes-

sage’s topic or topics mixture (Jelodar et al., 2017).

To map each message to a point in topic space Φ we

can also use other statistical non-parametric methods

such as stochastic block model (Gerlach et al., 2018).

The aim is to represent a message M in the topic

space which can be, for example, the set of probabil-

ity distributions over Φ = T × U × V where a point

M = (.7, .1, .2) denotes that the message is made of

topics T , U and V respectively in proportions .7, .1

and .2. To simplify the next examples we suppose that

each message maps to a unique topic. M from topic

T would be the point (1, 0, 0).

A simple topic message dynamics example is

, T

, where all three messages are on the topic

T . At the opposite we could have T

, U

, V

where

each message maps respectively to topics T , U and

V .

This type of dynamic shows the evolution of the

topics’ popularity in the LMS, or the evolution of top-

ics’ interest over time.

Actor-topics’ Dynamics. How do an actor’s mes-

sages cover topics over time? Let (AT )

be the mes-

sage with timestamp τ

posted by actor A and on topic

T . The two threads in Figure 1 illustrate the follow-

ing actor-topics dynamics:

(AT )

, · ·· , (BU)

, (AT )

(1a)

(AT )

, (BV )

, · ·· , (CU)

, (AU)

(1b)

Where (1a) denotes that actor A published two

messages both on topics T , but the 2

came after

B’s message on topic U that was published long after

(·· ·) A’s 1

message.

In (1b), B publishes a message on topic V imme-

diately after A’s message on T . Then, after some time

and other publications, A posts a message on yet a 3

topic U, similar to what C had just published on.

From the above observable dynamics, we can de-

ﬁne two more dynamics: the actor-actor and topic-

topic dynamics.

Actor-actor’s Dynamics. The actor-actor’s dynam-

ics can be taken as the evolution of the actor’s social

network where the actors are linked by message close-

ness and past interactions. We suppose that messages

posted on overlapping topics in the same discussion

Towards Visual Explorations of Forums’ Collective Dynamics in Learning Management Systems

by different actors potentially indicate some interac-

tions between the authors. This may not always be the

case.

Figure 1 exempliﬁes how the strength of mes-

sages correlation is made of, topic, temporal and actor

closeness. What is not shown in this ﬁgure is that the

strength of the tie may also depend on previous ac-

tors’ interactions, that is the social network built from

previous messages correlations. We will detail this in

section 4 and explain how we avoid inferring causal

interactions from the messages’ correlations.

If a message is published shortly after another then

their correlation should be strong. In example (1b),

actors A and C would have a strong interaction be-

cause they published on the same topic U and their

respective messages’ timestamp are close. An inter-

action also exists between actors A and B but proba-

bly weaker because it is only based on the messages’

timestamp and not the topic overlaps. The relation-

ship between actors B and C would be even weaker if

it existed at all. Their messages are far apart and not

on the same topic.

So, from the message topic, time and actor cor-

relations we build a directed graph representing the

social network of messages exchanged between the

LMS actors. The evolution of that network is what

we call the actor-actor dynamic.

Topic-topics’ Dynamics. This type of dynamic

concerns the way the topic’s correlations evolves over

time. For example, if at the beginning of a course,

topics T and U tend to be closely connected because

students often mix them up in discussions, we hy-

pothesize that as the course’s concepts disambiguate,

the relationship between the two topics will likely de-

crease because fewer students will publish messages

mixing both topics in the same discussion.

As for the actor-actor dynamic, the topic-topic dy-

namic can be expressed as a temporal network (or

temporal graph), but where nodes are topics and links

between them represent relationships whose strength

depend on a topic, time and actor’ correlation. This

can be seen as a message based distance.

In Figure 1a), some relation between T and U

would occur for the same reason than that of A 99K B.

Incidentally, in the 2

thread, we would have V 99K

T based on time proximity, but also U 99K T based

on the shared author A.

As we will explain in the literature review, the

topic-topic network may be best suited for student tar-

geted visualization than the actor-actor network.

Finally, let’s recall that for us collective dynamic

is the evolution of relationships between topics and

actors spurring from the messages’ co-occurrences in

a LMS’s forum without making any assumption about

the actors’ intentions.

3 RELATED WORKS

We review previous works pertaining to collective ac-

tions and those about supporting learning with visual-

izations.

3.1 Analyzing Collective Actions

Studies that consider collaboration usually try to iden-

tify the intentions by studying the student’s behavior

and message publications.

Detecting Collective Action. In a 250 strong Com-

munity of Learning (COL), Rehm et al. (2015) com-

pare on-task users, those showing engagement and

high performance, with off-task users. They use

questionnaires to relate the different behaviors to the

users’ hierarchical position in the COL. They com-

pared the actors’ social network position and their en-

gagements in learning discussions. They found a pos-

itive correlation between social position and engage-

ment, therefore the authors did not invalidate their hy-

pothesis that the social position inﬂuenced the learn-

ing behavior. Social Network Analysis (SNA) tech-

niques are also used to collect statistical measures

from the social network of messages’ exchange in or-

der to automatically operationalize collaborative in-

dicators. In (Lobo et al., 2016), for example, SNA

is used to compute initiative, activity, regularity of

activity, regularity of initiation, and reputation, that

permits identifying isolated learners and learners with

potentially assistant roles.

Figure 2: iForum’s Dashboard (Fu et al., 2017) showing (a)

overall changes of post in the forum, (b) a thread represen-

tation, (c) discussions in packed forms.

CSEDU 2019 - 11th International Conference on Computer Supported Education

Wang et al. (2016) are other researchers interested

in the learning differences between off-topic and on-

topic users. After a detailed analysis of forum mes-

sages, they demonstrate that on-topic users, that is

high-order thinking users displaying constructive and

interactive behaviors in the forums, have more learn-

ing gain than off-topic learners. In conclusion, the

authors advocate for an off-topic discussion detector

mechanism to guide users back on more constructive

grounds.

In (Chua et al., 2017), the authors’ approach is

to study the turn-taking in discussions. They iden-

tify different types of conversation and categorize the

users as: loner, replier, initiator without reply, initiator

who respond, active social learner, active social with-

out turn-taking, reluctant active social learners. Be-

side their valuable categorization, an important result

is that they observe more engagement from recurrent

posters, that is poster replying to comments made to

their initial posts.

If the importance of collective action for learning

is agreed upon, the difﬁculty is identifying it at scale.

It is a complex process needing content, temporal and

social network analysis. The previous studies justify

the importance of a detailed content analysis but they

relied on human intervention limiting their potential

to scale up.

Scaling Up. Dascalu et al. (2017); Boroujeni et al.

(2017) are the ﬁrst big scale attempts that we found

taking into account time, message content, social and

dialogue structure. Each model the students’ dynam-

ics with a mixture of Natural Language Processing

(NLP) techniques such as LDA or Latent Semantic

indexing (LSI), and SNA (eg. block models) ap-

plied to big temporal datasets. Dascalu et al. (2017)’s

dataset contains 3,685 contributions from 179 partic-

ipants and spans 2 years. Boroujeni et al. (2017) use

2 datasets of respectively 7,699 and 12,283 messages

written by 1,175 and 1,902 participants.

Dascalu et al. (2017) operationalize collaboration

with a Cohesion Network Analysis score applied to

synchronous chat discussions. It correlates signiﬁ-

cantly with human discussions’ analysis but it is not

tested to identify collective actions as the learners

were forced by pedagogical design in collaborative

groups.

Boroujeni et al. (2017) analyze the inﬂuence of

the course structure (timing and # of the staffs’ pub-

lications) on the forum structure, content and on the

social network of learners. They show that the course

structure correlates to the forum structure (timing and

# of students’ posts), but not to the forum content or

to the social network grown from the students’ in-

Figure 3: Convis Dashboard (Hoque and Carenini, 2016)

helps explore conversations. On the left, the topics found in

the forum using LDA and organized hierarchically. In the

middle, the colored rectangles show a sentiment analysis for

each message. Each message is linked to his author placed

on a semi-circle and thus creating a social network. Finally,

on the right, Convis display the detail of the conversation.

teractions in the forum. They report that although

some learners do not publish often, those still have

an important impact in forums because they some-

times trigger long discussions on course’s topics. Fi-

nally, the authors recommend combining forum activ-

ity prediction model with content analysis to support

instructors focusing on important discussions.

These two studies exemplify the possibility to get

collective actions indicators based on content, struc-

ture and time, and as in (Ezen-Can et al., 2015) they

push for better support tools for tutors.

3.2 Visualizations as Support Tools

Visualization as supporting tools have been used suc-

cessfully in teaching contexts. Heer and Shneider-

man (2012) classify the important visualization, while

Emmons et al. (2017) and citeLeeuwen2014 advocate

their generalization to support all LMS users.

Exploration and Awareness Tools. Medina et al.

(2016) use a LDB to provide quick and precise feed-

back with their Behavioral Awareness Mechanism.

They tackle the problem of portability and provide

dynamic feedback across several platforms to 24

students working on collaborative projects. They

use communication, coordination, motivation, perfor-

mance and satisfaction indicators to operationalized

collective actions. Other studies showed how group

awareness, i.e. rapid feedback about collective ac-

tion, and visual narratives (Yousuf and Conlan, 2015)

are beneﬁcial to the students’ engagement, even if no

content analysis is done (May et al., 2011). Davis

Towards Visual Explorations of Forums’ Collective Dynamics in Learning Management Systems

et al. (2017) evaluated the impact of a radar type visu-

alization given to students in a two MOOCs providing

them with awareness on what previous sessions learn-

ers had done at the same time of the course. Their vi-

sualization had a positive impact on older students but

it did not improve signiﬁcantly the younger students’

activity in the forums.

Generally, as Qu and Chen (2015) note, there is

a “need to develop both advanced data mining meth-

ods to reveal patterns from MOOC data and visual-

ization techniques to convey the analytical results to

end users and allow them to freely explore the data

by themselves”. VisForum Fu et al. (2017) answers

Qu and Chen (2015)’s call. This LDB (Figure 2) pro-

vide a visual analytic system to interactively explore

the forum of a LMS. The complex interface helps the

tutor group users and compare them to gain insights

on the general forum’s dynamic in terms of structure

but also in terms of sentiment analysis. In G

omez-

Aguilar et al. (2015), an original spiral visualization

of Moodle’s activities enabled the authors to spot that

students’ activities peaked on the same days for stu-

dents with similar grades. The visualization cleared

helped the author identify this interesting behavioral

trend. Convis, as an exploratory visualization, also

satisﬁes Bull and Kay (2016)’s recommendation for

negotiable user models (Figure 3). It is a forum explo-

ration tool designed around a topic model that evolves

with users’ feedback. It facilitates ﬁnding insight-

ful messages from discussions crammed within hun-

dreds. Finally, Guo et al. (2017) propose TieVis, an

original scalable visualization speciﬁcally tailored to

track explore and analyze dynamics in interpersonal

links, like those we could have between the actors

mentioned in the previous section.

The limit of these researches is their complex vi-

sualizations. iForum required an extensive explana-

tion from the designers to help the instructor grasp

what was shown. Similarly, in Convis some users re-

ported difﬁculties to use the visualization efﬁciently

and in TieVis the authors recognized that some of

their visualizations were not intuitive at all.

It is clear that visualizations can help creativity

and holistic thinking; improve the ability to make ef-

fective inferences; that translating or making visual

analogies reinforces conceptual development; im-

pacts cognition, helps sense-making and understand-

ing (Klerkx et al., 2014). Without contradicting the

later, Twissell (2014) clarify the visualizations lim-

its : a) different learning styles, natural differences

in learners have a signiﬁcant impact on the way dia-

grams are perceived, visualized and understood b) vi-

sualizations do not equally affect all types of learning

activities.

Visualizations’ Effectiveness. The study from

Anaya et al. (2016) investigates how to reinforce stu-

dent collective actions in the LMS dotLRN. Notice-

ably they focus on directly helping the students, by

designing a comprehensible tree-based visualization

explaining to the student why they received a recom-

mendation to act more collectively and how having a

higher-order thinking behavior would be beneﬁcial to

them. The engineer students working on a collabora-

tive project reported to understand the tool and gener-

ally found it useful.

Nevertheless, precaution has to be taken with vi-

sualizations for young learners because, for the least,

Lonn et al. (2015) found that a LDB could have unde-

sired side effects on teenagers. Analyzing students in

a summer camp remedial program, they reported that

extensive referral to a LDB downgraded some mas-

tery goal willing students to students only willing to

show proof of competence only. The opposite was not

witnessed. Therefore, some students abandoned their

initial motivation to understand and aimed to trick the

system so that the LDB showed that they understood.

This last experiment justiﬁes that although we aim

to support tutors with explorable visualizations, we

will consider topic based visualizations rather than

user-based ones. Topic-based visualizations do not

emphases individual actions and, therefore, should

damper the motivations to trick the system by adopt-

ing a superﬁcial behavior.

4 THE EXPLORABLE

COLLECTIVE DYNAMIC

MODEL

In this section, we detail our datasets and present the

model that we are going to use to analyze the collec-

tive dynamics from them.

4.1 Dataset Collection

Table 1 lists our seven datasets collected in 2017 and

2018. They are organized into four groups based

on the datasets’ origins. The datasets contain fo-

rum information from the following online courses:

1) Coursera (2018), a database extraction for a Hu-

man Right (HR) and Understanding Financial Mar-

kets (UFM) MOOCs. 2) Moodle (2009), a database

extraction without message content for a French as a

Foreign Language (FFL) course. 3) Hangouts (2018),

a JSON export from the VUCI’s G Suite for Educa-

tion with the data from 5 chats setup-up by the uni-

versity staff for the staffs or the students. 4) Coursera

CSEDU 2019 - 11th International Conference on Computer Supported Education

Table 1: List of datasets and courses that we use to experiment our LDB.

The forum structure is given by the existence of

different forum types: Weekly (W), General (G), Technical Support (S), Thematic (T) or Assignment (A) related.

Extra

information is often available, such as messages up votes (V), comments (C), subscription (Sub.) or ﬁle attachment (Att.).

iii

When data was scrapped, the dates were in humanized format (e.g. 6 month ago, 23 minutes ago), therefore the precision

varies with the posts’ ages. Recent posts can be compared with greater precision than older ones. We give the intervals in

which the precision varies in hours (h), days(d), month (m) and year (y).

Dataset Discussion Message Author

Time

Structure

Extra

source (#) (#) (#) Span (d) Precision

Moodle (2009)

FFL 348 1490 19 78 s T Active time,

Citations, Att.

Coursera (2017)

PP 868 2548 1112 365 [1 h; 2 m]

iii

W, G & A V & C

PML 1135 4157 982 240 [5 h; 1 m]

iii

W, G & A V & C

AT 248 549 311 728 [1 d; 1 y]

iii

W & A V & C

Coursera (2018)

HR 499 1004 638 989 ms W, G & S V, C & Sub.

UFM 1318 9460 4609 1022 ms W, G & S V, C & Sub.

Hangouts (2018)

VUCI 5 7297 96 327 s G & S -

(2017), a saving of online courses using selenium web

scraper. The dataset is then transformed and stored

as a CSV ﬁle to be processed by a Python engine.

It has messages’ content but approximates the times-

tamp. Three courses are available: Python Plotting

(PP), Python Machine Learning (PML) and African

Towns (AT) an urban planning course.

4.2 Collective Dynamic Model

Implicit Relationship’S Strength. Figure 1 gives

a ﬁrst example of messages’ correlation, or close-

ness, translated as interaction’s strength between their

authors. There, the strength is either high, low or

null. In practice the strength I, that we will refer to

as implicit relationship’s strength, could be anything

in [0;1]. We propose to deﬁne the implicit relation-

ship strength between two messages as a function of

time, topics and actors. This translates the idea that

the messages relationship depends on :

Who: wrote them. Do the messages’ authors have

already interacted together before ? Is the author

a super poster, a lonely lurker ? One will probably

consider a message differently depending on his

relationship to the message’s author.

Time: Obviously, the delay (or time delta) between

two messages inﬂuence the strength of their rela-

tionship. The quicker the response, the stronger

the link.

Content: should also play an import role in the way

messages relate to one another. The difﬁculty is to

reliably automate content analysis, but NLP tech-

niques exist to advance in that direction.

Required Message Correlation. As we said ear-

lier, deducting the causal relationship between mes-

sages is difﬁcult because one has to guess what is the

real intention of the message’s author. To avoid er-

rors, we do not directly deduce the messages real cor-

relation based uniquely on the implicit strength. In-

stead, we propose to make the relationship strength

Figure 4: Graphical proposition for the function E(I, r),

the explicit messages’ interaction strength. I is the “im-

plicit” strength based on content, time and social network

structure. r is “requirement” set by the observer.

Towards Visual Explorations of Forums’ Collective Dynamics in Learning Management Systems

Figure 5: Interactions cycles built from a bi-party actor-

topic graph Thread c and d are transformed to an actor-actor

graph. Dotted arrows denotes weaker links.

dependent on a parameter set interactively by the ob-

server. We call it the required message correlation r

(r ∈ [0; 1]). If the observer set r ≈ 0 then, for him, the

requirement for a message to be linked to other mes-

sages is weak and the interactions between messages,

actors and topics will be common. Although that

would probably lead to overly complex and unusable

dynamics metrics and visualizations. Conversely, if

r ≈ 1, the requirement is high, meaning that the ob-

server wants linked messages to be close in time, have

a lot of topic overlap and be written from closely con-

nected authors. In that case messages, actor and top-

ics will hardly have any relations with one another and

dynamic will probably be invisible.

Explicit Relationship’S Strength. The graph of

the function E could be like the one pictured in Fig-

ure 4. If the observer’s requirement is high, r ≈ 1,

then, for most Is, the interaction E should be low, and

conversely, if the observer’s requirement is low, r ≈ 0,

then the interaction E should be high, for most Is.

We care about this interaction because it’s based

on its value that we display the actor-actor or topic-

topic dynamics. To further illustrate our intent, we

complete our ﬁrst example with longer threads (Fig-

ure 5). We set four strengths to I: high, medium, low

or ≈ 0. Each time or topic delta decreases the implicit

strength by one unit. Therefore, in our case, keeping

the topic unchanged, the maximum number of rela-

tions a message can have with “previous” messages is

3. We call this connection pattern previous-3. But, the

I function could have other forms. For example, the

star pattern, i.e., I been strong for the initial message

but null for the others; or the previous-∞ pattern, also

called “total co-presence” in Wise et al. (2017), de-

noting existing correlations with all the previous mes-

sages. In Figure 5c), not only does actor D connects

to actor C but he also connects to actor B. He po-

tentially could have been connected to actor A but in

addition to their time delta of 3, there a is topic delta

dropping the strength of I nearly to 0. Hence, only an

exceptionally low requirement r would explicit E the

interaction between the messages of actors A an D. In

our case, only if D had published on the same topic

as A would their implicit relationship be high enough

to gain visibility in our sociogram since this is not the

case, D is not directly connected to A in the associated

sociogram.

The results of these manipulations are a temporal

actor-actor network and a temporal topic-topic net-

work that can be represented by weighted oriented

graphs. Snapshots of an actor-actor network from the

PP dataset is presented in Figure 9.

Identifying Collective Actions. Once we built the

actor-actor and topic-topic networks, we want to iden-

tify potential collective actions. Those are derived

from the actor-actor network structure. We make the

hypothesis that collective actions need the presence of

a recurrent actor, that is an actor replying to one of his

replier (Chua et al., 2017). We take it as the evidence

that at least one of the actors has potentially assimi-

lated someone else’s message before acting, therefore

initiating a collective action. Structurally, recurrent

actors form cycles in the sociogram. For example,

in Figure 5c), actor A who posted twice the thread,

close the cycle A ← B ← C L99 D L99 E L99 A. Fur-

thermore, since actor E is in cycle with A and F, and

F with H, we will consider that actor A and all of

the above are engaged in a common collective action

with H. G on the other side is not part of any cycle

and, therefore, does not participate in collective ac-

tion. In fact, G published a message and disappeared.

We have no evidence that his message had an impact

on others or others on his. That is why we consider

necessary (but not sufﬁcient) that the recurrent inter-

actions conditions collective action. In that sense, we

further Chua et al. (2017)’s ﬁndings that recurrent in-

teractions are important for discussions.

5 FIRSTS VISUALIZATIONS

We now introduce three visualizations from our work

in progress. They were made separately, each testing

some elements of the global conception model pre-

sented in Figure 8.

CSEDU 2019 - 11th International Conference on Computer Supported Education

Figure 6: Visualizations from the FFL dataset.

5.1 Moodle Dataset

We used the FFL dataset to sketch our ﬁrst visual-

izations. Our Moodle dataset has the particularity to

contains detailed information about the actors’ activ-

ity type and duration. It enabled us to distinguish the

users’ active time from their idle times. With that

information, we came up with the visualizations of

Figure 6, built as part of a standalone web applica-

tion. Once we selected a user, we see the visualiza-

tions corresponding to his active time. The top chart

presents, in yellow, the total time spent in the forums

for a user, day per day and relates it with his active

time, shown in blue. To handle the scaling problem,

we implemented a monthly, weekly and daily based

grouping criteria. On the lower chart, we compared

the activity time of two users. We see that, although

they display similar activity patterns over time, one is

generally more active than the other.

This was a successful test to sketch our ﬁrsts vi-

sualizations, but the FFL dataset lacked content and

its size did not create a huge monitoring problem for

the tutors. Albeit, this dataset is interesting because it

emphasizes the importance of the activity type. The

following step is to scale up with a bigger dataset that

includes content.

5.2 Hangouts Dataset

The hangouts dataset is slightly different from the

other datasets. It is less structured because it comes

from chats and not forums. Figure 7 displays 6 274

messages from one chat, gathering several threads

of exchanges between the VUCI’s administration and

their tutors, from March 2018 to December 2018.

We started from a manual, but automatable, export

of Google’s hangouts services. It gave us a JSON ﬁle

that we preprocessed in python and fed to d3, a visu-

alization JavaScript library, via a standalone Django

Figure 7: Detail of VUCI conversation taking place in 2

hours. Each circle is a message, and hovering over them

brings up its content. The circle’s size is proportional to

their content’s length. Message are layered vertically by ac-

tors. On the left is an indication of the actors total messages

count.

web application. We built a LDB, testing on that

larger dataset, interactive features such as zooming,

panning, data point selection. The ﬁgure’s pane (a)

contains a bird’s eye view of all messages. Users

are represented vertically, in the middle, with their

names. On the left is a histogram of message count for

each user. On the right, along with the time axis, we

plotted the messages as discs whose areas are propor-

tional to the messages’ length. Activating the mouse

wheel while on the time axis zoom in and out. Bel-

low, is the 3 hours period framed in red, zoomed. It

shows a message pop-up, with full detail, activated by

the mouse hovering a data point.

In this investigation, we did not include the SNA

and NLP analysis because our objective was the vi-

sualization of a large dataset with content and the

handling of the scaling problem. It proved that our

technology choices were sound. We transformed the

dataset of several thousand messages, from the JSON

ﬁle to the HTML rendering, in a few seconds. But

it also pinpointed the importance to implement many

interactive exploration functions to alleviate the scal-

ing problem. For examples, ways to quickly zoom on

a few data points without losing the overall picture, as

well as ways to ﬁlter and order data point or axis la-

bels, and also enabling more intricate communication

between the different LDB’s charts. In addition to all

Towards Visual Explorations of Forums’ Collective Dynamics in Learning Management Systems

this, a major drawback to our visualization is that it

did not take into account, yet, the ﬁnal users. Finally,

besides deﬁning visual modalities, we still need to test

the algorithm to compute the collective activity indi-

cator.

5.3 Coursera’s Dataset

We used the PP dataset from Coursera 2017 to in-

vestigate the construction of a collective activity in-

dicator with SNA techniques. We came up with the

sociogram (Figure 9) illustrating an actor-actor dy-

namics. Nodes are all learners. We removed the

tutors and the course’s mentors to approach Dillen-

bourg’s collaboration deﬁnition that we gave in Sec-

tion 2. The users are linked based on their proximity

to the previous-3 actors who published in the same

discussion. The arrows’ width depends on the mes-

sages’ timestamp closeness, # of co-occurrences of

the authors and # of votes that the message collected.

We colored them by discussion, and order them by

age, the oldest being the lightest. This is an interme-

diate visualization that is used for analysis purposes

and will not be presented to the end users for two

reasons: a) it gives abstruse information for some-

one that does not have access to the raw data, b) it

represents the actors as nodes and as we noted in sec-

tion 3, we would rather have visualizations showing

topics than persons. We present it to illustrate what

a social network from our dataset looks like, and be-

cause, zoomed and reduced to the three snapshots (b),

(c), (d), representing three successive yearly quarters,

we distinguish an evolving pattern. In particular a de-

tailed analysis of actor 642, circled in red, show that

he started to designate two other messages probably

because they were meaningful to him (a), then he en-

gaged in several message exchanges designating oth-

ers’ message meaningful as his where also bringing

Figure 8: Data Analysis cycle.

attention (c), and in the most recent quarter someone

commented one of his earlier message (d). It is not

clear from Figure 9 if actor 642 was part of cycles.

Testing our hypothesis that actors in cycles share a

collective dynamic is in our perspectives.

6 PERSPECTIVES AND

CONCLUSION

6.1 Perspectives

The three test and visualizations presented are prepar-

ing further work to build: a collective activity indica-

tor taking in account the social network structure, the

content of messages and their evolution in time, and

visualizations for that indicator.

To further our effort to ﬁnd visual modalities for

the indicator, we conducted a small survey in July

2018 with 48 tutors from the VUCI to introduce our

project and start engaging them in a co-construction

process. 27 were not satisﬁed or unsure about their

current tools’ effectiveness to monitor their students’

work. 39 agreed or strongly agreed that ICTs could

help their students better collaborate and 40 that col-

laboration was, indeed important for learning. In a

second survey, we plan to ask the tutors the kind of

visual representation they would ﬁnd useful to moni-

tor the collective activities of their students, but also

what dimensions of the relationships between mes-

sages they would like to leverage to explicit visually

those relationships. This is part of a co-construction

approach that should engage the tutors, facilitating the

adoption of our visualization while increasing its us-

ability and impact.

Concerning the portability and validation of our

pipeline stages (Figure 8), we will need to continue

working with several datasets, extending our uniﬁed

model to incorporate SNA and NLP statistics for all

datasets. The portability assumption rests on our

capacity to extract periodically, every few minutes,

hours or days, data from the main LMS. This we will

necessitate proper APIs authorizations to make our

LDB communicate effectively with the LMS.

6.2 Conclusion

In this paper, we presented a model to detect collec-

tive activities from the forums’ discussions. We based

our model on an implicit message relationship and an

external parameter set by an observer to explicit that

relationship. Doing so, we hope to facilitate the porta-

bility of our indicator to courses covering different

CSEDU 2019 - 11th International Conference on Computer Supported Education

Figure 9: At the top (a) is half of compound yearly actor-actor network. The three bottom images (b), (c) and (d) are

closeup around actor 642 during the quarters of the year. Higher resolution images for all ﬁgures are available at https:

//git-lium.univ-lemans.fr/mkone/csedu-2019

domains, spanning various time periods and having

different populations. Learning Analytics (LA) is at

a turning point where lots of attention is moving to

support tools rather than fully automated learning so-

lutions (Kon

e et al., 2018; Baker, 2016). Therefore,

we started investigating ways to give a visual feed of

the collective dynamics spurring from the LMSs fo-

rums, back to the tutors. We illustrated our work in

progress with three visualizations pushing for more

data analysis and data visualization for learning. We

further hope that this work will help tutors and, even-

tually, students discover the topic-topic dynamics in

their MOOCs and support collectives activities, so

beneﬁcial to learning.

REFERENCES

Anaya, A. R., Luque, M., and Peinado, M. (2016). A visual

recommender tool in a collaborative learning experi-

ence. Expert Systems with Applications, 45:248–259.

Baker, R. S. (2016). Stupid Tutoring Systems, Intelligent

Humans. International Journal of Artiﬁcial Intelli-

gence in Education, 26(2):600–614.

Boroujeni, M. S., Hecking, T., Hoppe, H. U., and Dillen-

bourg, P. (2017). Dynamics of MOOC discussion fo-

rums. In LAK, pages 128–137.

Bull, S. and Kay, J. (2016). SMILI: a framework for inter-

faces to learning data in open learner models, learning

analytics and related ﬁelds. International Journal of

Artiﬁcial Intelligence in Education, 26(1):293–331.

Chua, S. M., Tagg, C., Sharples, M., and Rienties, B.

(2017). Discussion Analytics: Identifying Conversa-

tions and Social Learners in FutureLearn MOOCs.

Dascalu, M., McNamara, D. S., Trausan-Matu, S., and

Allen, L. K. (2017). Cohesion network analysis of

CSCL participation. Behavior Research Methods,

pages 1–16.

Davis, D., Jivet, I., Kizilcec, R. F., Chen, G., Hauff, C., and

Houben, G.-J. (2017). Follow the successful crowd:

raising MOOC completion rates through social com-

parison at scale. In LAK, pages 454–463.

Dillenbourg, P. (1999). What do you mean by collabora-

tive learning. Collaborative-learning: Cognitive and

computational approaches, 1:1–15.

Emmons, S. R., Light, R. P., and B

orner, K. (2017). MOOC

visual analytics: Empowering students, teachers, re-

searchers, and platform developers of massively open

online courses. Journal of the Association for Infor-

mation Science and Technology, 68(10):2350–2363.

Ezen-Can, A., Boyer, K. E., Kellogg, S., and Booth, S.

(2015). Unsupervised modeling for understanding

MOOC discussion forums: a learning analytics ap-

proach. In Proceedings of the ﬁfth international con-

ference on learning analytics and knowledge, pages

146–150. ACM.

Towards Visual Explorations of Forums’ Collective Dynamics in Learning Management Systems

Fu, S., Zhao, J., Cui, W., and Qu, H. (2017). Visual Analysis

of MOOC Forums with iForum. IEEE Transactions

on Visualization and Computer Graphics, 23(1):201–

210.

Gerlach, M., Peixoto, T. P., and Altmann, E. G. (2018). A

network approach to topic models. Science Advances,

4(7).

Gerstenberg, T. and Tenenbaum, J. B. (2017). Intuitive the-

ories. In The Oxford Handbook of Causal Reasoning.

omez-Aguilar, D. A., Hern

andez-Garc

ıa, A., Garc

ıa-

nalvo, F. J., and Ther

on, R. (2015). Tap into visual

analysis of customization of grouping of activities in

eLearning. Computers in Human Behavior, 47:60–67.

Guo, F., Chen, W., Lin, T., Zhu, B., Zhang, F., Wu, Y.,

and Qu, H. (2017). TieVis: visual analytics of evo-

lution of interpersonal ties. Journal of Visualization,

20(4):905–918.

Heer, J. and Shneiderman, B. (2012). Interactive dynamics

for visual analysis. Queue, 10(2):30.

Hoque, E. and Carenini, G. (2016). Multiconvis: A vi-

sual text analytics system for exploring a collection

of online conversations. In Proceedings of the 21st In-

ternational Conference on Intelligent User Interfaces,

pages 96–107. ACM.

Jelodar, H., Wang, Y., Yuan, C., Feng, X., Jiang, X., Li,

Y., and Zhao, L. (2017). Latent Dirichlet Allocation

(LDA) and Topic modeling: models, applications, a

survey. arXiv:1711.04305 [cs].

Khalil, H. and Ebner, M. (2014). MOOCs completion rates

and possible methods to improve retention-A litera-

ture review. In World Conference on Educational Mul-

timedia, Hypermedia and Telecommunications, vol-

ume 1, pages 1305–1313.

Klerkx, J., Verbert, K., and Duval, E. (2014). Enhancing

Learning with Visualization Techniques. In Spector,

J. M., Merrill, M. D., Elen, J., and Bishop, M. J., ed-

itors, Handbook of Research on Educational Commu-

nications and Technology, pages 791–807. Springer

New York.

Kon

e, M., May, M., and Iksal, S. (2018). Towards a Dy-

namic Visualization of Online Collaborative Learning.

In CSEDU (1), pages 205–212.

Lobo, J. L., Santos, O. C., Boticario, J. G., and Del Ser, J.

(2016). Identifying recommendation opportunities for

computer-supported collaborative environments. Ex-

pert Systems, 33(5):463–479.

Lonn, S., Aguilar, S. J., and Teasley, S. D. (2015). Inves-

tigating student motivation in the context of a learn-

ing analytics intervention during a summer bridge pro-

gram. Computers in Human Behavior, 47:90–97.

May, M., George, S., and Pr

ot, P. (2011). TrAVis to

enhance online tutoring and learning activities: Real-

time visualization of students tracking data. Interac-

tive Technology and Smart Education, 8(1):52–69.

Medina, E., Meseguer, R., Ochoa, S. F., and Medina, H.

(2016). Providing Behaviour Awareness in Collabora-

tive Project Courses. Journal of Universal Computer

Science, 22(10):1319–1338.

Qu, H. and Chen, Q. (2015). Visual analytics for MOOC

data. IEEE computer graphics and applications,

35(6):69–75.

Rehm, M., Gijselaers, W., and Segers, M. (2015). The

impact of hierarchical positions on communities

of learning. International Journal of Computer-

Supported Collaborative Learning, 10(2):117–138.

Schwendimann, B. A., Rodriguez-Triana, M. J., Vozniuk,

A., Prieto, L. P., Boroujeni, M. S., Holzer, A., Gillet,

D., and Dillenbourg, P. (2017). Perceiving learning

at a glance: A systematic literature review of learning

dashboard research. IEEE Transactions on Learning

Technologies, 10(1):30–41.

Twissell, A. (2014). Visualisation in applied learning con-

texts: a review. Journal of Educational Technology &

Society, 17(3).

Wang, X., Wen, M., and Ros

e, C. P. (2016). Towards trig-

gering higher-order thinking behaviors in MOOCs.

In Proceedings of the Sixth International Conference

on Learning Analytics & Knowledge, pages 398–407.

ACM.

Wise, A. F., Cui, Y., and Jin, W. Q. (2017). Honing in on

social learning networks in MOOC forums: examin-

ing critical network deﬁnition decisions. In Proceed-

ings of the Seventh International Learning Analytics

& Knowledge Conference, pages 383–392. ACM.

Yousuf, B. and Conlan, O. (2015). VisEN: Motivating

Learner Engagement Through Explorable Visual Nar-

ratives. In Design for Teaching and Learning in a Net-

worked World, pages 367–380. Springer.

Zheng, S., Rosson, M. B., Shih, P. C., and Carroll, J. M.

(2015). Designing MOOCs as interactive places for

collaborative learning. In Proceedings of the Second

(2015) ACM Conference on Learning@ Scale, pages

343–346. ACM.

CSEDU 2019 - 11th International Conference on Computer Supported Education