Personalized Curiosity Engine (Pique): A Curiosity Inspiring Cognitive

System for Student Directed Learning

Safat Siddiqui

, Mary Lou Maher

, Nadia Najjar

, Maryam Mohseni

and Kazjon Grace

University of North Carolina at Charlotte, NC, U.S.A.

University of Sydney, Sydney, Australia

Keywords:

Personalized Learning, Curiosity, Recommender Systems for Education, Computational Models of Novelty.

Abstract:

Pique is an AI-based system for student directed learning that is inspired by a cognitive model of curiosity.

Pique encourages self-directed learning by presenting a sequence of learning materials that are simultaneously

novel and personalized to learners’ interests. Pique is a web-based application that applies computational

models of novelty to encourage curiosity and to inspire learners’ intrinsic motivation to explore. We describe

the architecture of the Pique system and its implementation in personalizing learning materials. In exploring

the use of Pique by students in undergraduate and graduate courses in Computer Science, we have developed

and implemented two computational models of novelty using Natural Language Processing techniques and

concepts from recommender systems. In this paper, we describe the Pique model, the computational models

for measuring novelty in text-based documents, and the computational models for generating sequences of

personalized curiosity-eliciting learning materials. We report the response from students in the use of Pique in

four courses over two semesters. The contribution of this paper is a unique approach for personalized learning

that encourages curiosity.

1 INTRODUCTION

One of the most signiﬁcant challenges in education

at scale is how to personalize learning (Sampson and

Karagiannidis, 2002). This problem is particularly

prevalent when considering problem- (Wood, 2003),

project- (Krajcik and Blumenfeld, 2006) and studio-

(Carter and Hundhausen, 2011) based learning, in

which content is open ended and students have auton-

omy in deciding how to focus their learning. How can

each student be presented with knowledge and chal-

lenges that ﬁt their interests and encourage curiosity?

One approach is to devote instructor time to provid-

ing personalized learning materials and advice to each

student, which is highly effective in small classrooms

but does not scale. We present the use of AI meth-

ods — speciﬁcally Natural Language Processing—for

personalized learning, based on eliciting curiosity us-

ing computational models of novelty. We contextual-

ize our models of novelty in an interactive system for

recommending course relevant publications to Uni-

versity students.

Curiosity can be deﬁned as the desire to learn

or know. Curiosity can be both a trait and a state

(Berlyne, 1966). Curiosity-as-trait refers to an innate

desire possessed by different people to different de-

grees, while curiosity-as-state refers to a motivation

to seek novel stimuli. This latter deﬁnition is the ba-

sis for encouraging curiosity in Pique. The curiosity

state can arise from exposure to appropriately novel

stimuli, with insufﬁciently novel stimuli being bor-

ing and overly novel stimuli being alienating, a model

ﬁrst proposed by early psychologist Wilhelm Wundt

(Berlyne, 1966). This creates a region of optimal nov-

elty within the space of possible stimuli, in which cu-

riosity is maximally likely to be stimulated. The pa-

rameters of this region are dependent on experiences,

context, and personal preference for novelty (Boyle,

1983; Kashdan and Fincham, 2004). The goal of

Pique is to stimulate the curiosity state of the student

and recommend resources that place them in a state

of maximal curiosity by combining their interests and

novelty in selecting course materials.

The Pique system is built on the principle theo-

rized by Loewenstein (Loewenstein, 1994) that de-

notes curiosity as the result of an ‘information gap’

— the distance between what is known and what is

desired to be known. This leads to a model of curios-

ity as a kind of intellectual hunger, in which small

amounts of new knowledge prime further desire to

Siddiqui, S., Maher, M., Najjar, N., Mohseni, M. and Grace, K.

Personalized Curiosity Engine (Pique): A Curiosity Inspiring Cognitive System for Student Directed Learning.

DOI: 10.5220/0010883200003182

In Proceedings of the 14th International Conference on Computer Supported Education (CSEDU 2022) - Volume 1, pages 17-28

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

learn, but larger amounts have a satiating effect (Kang

et al., 2009). A related notion from developmental

psychology is Vygotsky’s ‘Zone of Proximal Devel-

opment’ (Vygotsky, 1978): the space of all knowl-

edge that is adjacent to current knowledge and thus

is comprehensible given it. These theories of knowl-

edge acquisition all use spatial metaphors for describ-

ing curiosity’s inﬂuence, which suggests approaches

for how it might be operationalized. Computational

models of curiosity based on these approaches have

led to AI systems that value the unexplored and un-

explained (Schmidhuber, 2010; Merrick and Maher,

2009; Grace and Maher, 2015b; Grace and Maher,

2015a). Pique contributes to this computational mod-

eling of curiosity by proposing to stimulate the cu-

riosity of each individual student. This is then com-

bined with a representation of that student’s prefer-

ences and knowledge to produce personalized recom-

mendations.

Throughout our project, we have explored a va-

riety of algorithms for both novelty and the genera-

tion of sequences of learning resources. In this pa-

per we present two computational models of novelty

and three ways to select learning resources to stim-

ulate students’ curiosity. Each of our models are

based on the concept of unexpectedness as a cause

of novelty and surprise, which consequently leads to

curiosity (Grace et al., 2017; Grace et al., 2018).

For instance, information that students ﬁnd interest-

ing but are not expecting creates a surprising response

when presented as a stimulus to their learning pro-

cess. When students are presented with information

related to their existing knowledge but contain differ-

ent perspectives, those learning materials seem novel

to them. Recently novelty and surprise have been pro-

posed as components of a new kind of recommender

system that attempts to expand its users’ preferences

(Niu et al., 2018; Adamopoulos and Tuzhilin, 2014).

In the Pique system, we use AI-based computational

models to identify novel documents from a data set

of learning resources and develop algorithms to gen-

erate a sequence of learning materials to encourage

curiosity personalized to individuals’ interests. This

approach enables instructors to personalize the learn-

ing experience after identifying the corpus of learning

materials in open ended project based learning.

The four major components of the Pique system

are illustrated in Figure 1: Learning Materials, Artiﬁ-

cial Intelligence Methods (AI), Learner Model, and

User Experience (UX). These components are de-

scribed in Section 3. Pique was applied in undergrad-

uate and graduate courses in Human Centered Design

(HCD) as well as a Graduate Teaching Seminar for

PhD students. The Pique system was used over sev-

eral semesters, throughout which we continually de-

veloped the models of novelty and sequence gener-

ation based on student and instructor feedback. Plat-

form data from the students’ usage of Pique as well as

their reﬂections on the recommended learning content

were used to answer the following research questions:

• RQ1: How does the experience of using Pique

enable self-directed exploration and personalized

learning?

• RQ2: How does the experience of using Pique as-

sist students in expanding their learning interests?

The remainder of this paper describes the related

work and theoretical background (Section 2), the

components of the Pique system (Section 3), the ex-

periences of the students who used it in the classroom

(Section 4), and directions for extending this research

(Section 5).

2 BACKGROUND

Pique contributes to the broad ﬁeld of AI in education.

Baker and Smith (Baker et al., 2019) identify three

perspectives on educational AI: learner-facing (which

focus on assisting students), teacher-facing (which fo-

cus on reducing teachers’ workloads), and system-

facing (which focus on institutions’ administrative

and management capabilities). Zawacki-Richter et al.

(Zawacki-Richter et al., 2019) similarly identify four

areas of AI applications in higher education: adaptive

systems and personalization, assessment and evalua-

tion, proﬁling and prediction, and intelligent tutoring

systems (ITSs). In this paper, we present Pique - a

unique approach to adaptive systems and personaliza-

tion in educational AI as it aims to include curiosity-

inspiring content in the students’ learning process.

Pique combines elements of the Educational Rec-

ommender Systems (i.e., recommendation of cus-

tomized course materials) and ITS approaches to pro-

viding personalized learning activities. The differ-

ence between Pique and existing educational rec-

ommender systems is that we adopt a cognitively-

inspired model of curiosity as the basis of our recom-

mended sequences of resources, rather than a focus

on explicit learning goals.

2.1 Intelligent Tutoring Systems and

Personalized Learning Technologies

The broad goal of intelligent tutoring is to leverage AI

techniques to provide the instructional capability that

adapts to the needs of individual students. To sup-

port students in effective ways, models have been de-

CSEDU 2022 - 14th International Conference on Computer Supported Education

Figure 1: Architecture of the Pique Cognitive System.

veloped to distinguish students’ search behaviors (An

et al., 2020). Learning Management Systems (LMS)

can monitor students’ learning patterns and categorize

students based on their behavioral patterns (Kuo et al.,

2021; Papamitsiou and Economides, 2014). Ai et al.

(Ai et al., 2019) have applied deep knowledge models

to trace students’ knowledge status and have adopted

reinforcement learning to recommend exercises. In-

telligent tutoring systems have been developed to pro-

vide personalized feedback for learning programming

courses (Keuning et al., 2018) and to support stu-

dents’ mathematical problems solving process (Pozd-

niakov et al., 2021). While we do not present Pique

as an ITS, it can be thought of as analogous to one:

it models student preferences, represents the learning

resources to be recommended, and has a strategy for

composing resource sequences that focus on students’

motivation and maximizing their learning curiosity.

Models of student motivation have previously

been used to augment the instructional capabilities

of ITSs. Del Soldato & Boulay (Del Solato and

Du Boulay, 1995) describe an approach to planning

communication with a student performing a series of

learning tasks based on a model of the student’s mo-

tivations. They base the motivational reasoning on

Keller’s (Keller, 1987) model of motivation as con-

sisting of curiosity, challenge, conﬁdence and control,

previously used in computer-supported collaborative

learning (Jones and Issroff, 2005). The Pique sys-

tem focuses on the ﬁrst of these motivational factors:

curiosity, which is stimulated by surprise and nov-

elty. We explore the computational models of novelty

and recommend resources calculated to increase stu-

dent familiarity with concepts, but additionally aim to

elicit their surprise and curiosity.

2.2 Educational Recommender Systems

(ERS)

The goal of ERSs is to recommend learning resources,

and they can be applied in either formal or informal

educational contexts, and used by either students or

instructors. The Pique system is intended to be ap-

plied to open-ended learning tasks, which — while

part of a formal educational context — share many of

the features of informal learning. Open-ended learn-

ing tasks require students to select a scope of focus for

their work within the proposed problem space (Hill

and Land, 1998), which means that students are self-

directed to a degree, an aspect shared with informal

educational contexts.

Educational recommender systems have been

used to recommend course material for students in

Computer Science (Kose and Arslan, 2016) and Busi-

ness and Administration studies (Hall Jr and Ko,

2008). Cobos et al. (Cobos et al., 2013) have de-

veloped a recommendation system for the instruc-

tors to prepare course content. Educational recom-

mender systems include explainability to justify the

recommendation. Barria-Pineda et al. (Barria-Pineda

et al., 2019) have helped students minimize misun-

derstandings related to solving programming prob-

lems and have explained the reasoning behind the

recommended learning activities. Barria Pineda and

Brusilovsk (Barria Pineda and Brusilovsky, 2019)

identify that students spend more time on the ex-

ploratory interface and suggest the effectiveness of

the transparent recommendation process. In Pique,

we show students the calculated novelty scores of the

papers to facilitate students’ paper selection process.

Personalized Curiosity Engine (Pique): A Curiosity Inspiring Cognitive System for Student Directed Learning

3 THE PIQUE SYSTEM

3.1 The Learning Materials Component

The Learning Materials component is the source

of documents provided by the instructor for a spe-

ciﬁc course. For Pique, these documents are repre-

sented as unstructured text, drawn from articles from

relevant conferences, journals, and digital libraries.

The preparation of the learning materials for use in

Pique includes data collection, dataset preparation,

and dataset description to prepare the text documents

for applying computational models of novelty.

Pique has been included in two courses in a Com-

puter Science program. The ﬁrst, ‘Human Centered

Design’, has a focus on human computer interac-

tion. The learning materials are papers published in

the ACM Digital Library under the classiﬁcation of

Human Centered Computing. The second, ‘Gradu-

ate Teaching Seminar’, has a focus on educational re-

search in computer science, and the relevant learning

materials are papers published in the ACM SIGCSE

(Special Interest Group on Computer Science Educa-

tion) proceedings.

For the Human Centered Design Course we ex-

tracted 9,452 conference, journal and magazine pa-

pers with publication dates from 2008-2018. For each

publication we collected the following metadata: ti-

tle, ISSN, location, abstract, publisher, address, ACM

ID, journal, URL, volume, issue date, DOI, number,

month, year, pages, and tags/keywords. For the Grad-

uate Teaching Seminar we extracted 1172 papers with

publication dates from 2008-2018, with the following

metadata: title, author, conference, year, DOI, key-

words, and abstract.

3.2 The AI Component

The AI component of the Pique system applies NLP

and novelty detection algorithms to rank the novelty

of items in the learning content module as a basis for

generating a personalized sequence of learning ma-

terials. The AI component has two subcomponents:

the Model of Novelty and the Sequence Composition.

The Model of Novelty module receives a set of doc-

uments from the Learning content module and gener-

ates novelty scores for each of the learning items in

the corpus. This subcomponent uses a list of features

to represent each learning item based on topic models

or keywords associated with each item. These fea-

tures provide the basis for computing a novelty score

for each item. The Sequence Composition subcom-

ponent uses the novelty score, information about the

learner’s interests, and the information from their in-

teraction with Pique to generate a sequence of learn-

ing materials. These materials are personalized to en-

courage the learner’s curiosity and their intrinsic mo-

tivation to explore the learning resources.

3.2.1 Models of Novelty

In developing Pique, we implemented two compu-

tational models of novelty, each based on the prob-

abilities of keywords and topic models associated

with each document. We refer to these models as

the ‘Keyword co-occurrence model’ and the ‘Topic

co-occurrence model’. The keyword co-occurrence

model represents each item in the learning materials

as a bag of keywords. The topic co-occurrence model

represents each item in the learning materials as a vec-

tor of topic distributions by applying a topic modeling

algorithm (Blei and Lafferty, 2007) to the corpus of

learning materials. This section describes how these

computational approaches generate novelty scores for

the corpus of learning materials provided by the in-

structors.

Keyword Co-occurrence Model. The keyword co-

occurrence model is based on the probability of the

co-occurrence of each pair of keywords in the cor-

pus. The challenge in this model is to identify the

keywords for the learning materials. In our corpus,

each paper has two ﬁelds in the metadata that can

serve as the keywords for this model: the keywords

selected from the ACM’s Computing Classiﬁcation

System (CCS) and the author deﬁned keywords. The

ACM Computing Classiﬁcation System has been de-

veloped as a poly-hierarchical ontology that results in

common topics relevant to all papers, but they do not

speciﬁcally represent the content in each paper. Con-

versely, author-deﬁned keywords are deﬁned speciﬁ-

cally for each paper, but do not follow any standard

representation.

To calculate the probability of keyword co-

occurrence, we synthesized the list of keywords from

each paper into a master list of keywords for the cor-

pus. Due to the large number of keywords in the mas-

ter list of keywords, we manually curated a reduced

set that can be used to create a mapping from a user’s

interests to the concepts in the learning materials. In

reducing the number of keywords we tried to target

the largest number of keywords that are reasonable

to present to students for selection: too many key-

words would be overwhelming, and not enough key-

words would not represent the dataset with enough ﬁ-

delity. In our pilot study of using Pique in the Human-

Centered Design course, when using the CCS classi-

ﬁcation as keywords, we gave students 118 different

CSEDU 2022 - 14th International Conference on Computer Supported Education

keywords to select from. This was perceived by the

students as too many, and it resulted in a very sparse

user preference vector. As a result, we reduced this

set in subsequent semesters. We manually replaced

keywords that were not in the reduced list to be the

most relevant keyword in the reduced set. Across the

semesters, feedback from students indicated that our

reduced set of 35-55 keywords was sufﬁcient for stu-

dents to express their interests.

With the keywords for each paper, we created a

bag-of-keywords representation to calculate the co-

occurrence of keywords for measuring novelty. First

we eliminated papers that had fewer than two key-

words. We then measured the probability of each pair

of keywords appearing together in the corpus. We

used these probabilities (see eq. 1, eq. 2) to calcu-

late the probability of keywords x

and x

occurring

together in the corpus (eq. 3) and took its logarithm as

the novelty score for that pair of keywords (as in (Niu

et al., 2018), (Bouma, 2009)). We prepared a novelty

matrix, NM (eq. 4) by applying this process to all

pairs of keywords in the corpus. This matrix serves

as the look-up table for identifying the novelty scores

among the keyword pairs in the papers. To convert

from this keyword-pair novelty score to the score for

a paper, we took the highest value of all keyword pairs

present in the paper (eq. 5) as surprising combinations

stand out (Grace et al., 2017).

prob(x

) =

# o f papers have x

# o f total papers

(1)

prob(x

) =

# o f papers have x

# o f total papers

(2)

prob(x

, x

) =

# o f papers have both x

and x

# o f total papers

(3)

NM(x

, x

) =

log

(prob(x

, x

))

prob(x

) ∗ prob(x

)

(4)

NoveltyScore P

= max(NM(x

, x

), NM(x

, x

), ...)

(5)

Topic Co-occurrence Model. The Topic Co-

occurrence Model calculates the novelty score based

on frequency and proportion of topics present in the

corpus. We applied the topic modeling approach on

the abstract of the papers. Topic modeling produces a

set of ‘topics’ each comprising a distribution over all

the words in the corpus (Grace et al., 2017). The ad-

vantage to using topic modeling compared to author

deﬁned keywords is that there is consistency in the

identiﬁcation of features across the entire data set in

topic modeling, where author deﬁned keywords pro-

vide features relevant to the author of a single item in

the corpus.

We used the R package ‘STM’ (Structural Topic

Model) (Roberts et al., 2019), a topic model extension

that is equivalent to CTM (Correlated Topic Model)

(Blei and Lafferty, 2007). Correlated topic models

relax the assumption made by earlier topic modeling

algorithms that all the topics in a corpus are indepen-

dent and therefore no one pair of topics is more likely

to occur together in a document than any other. The

STM algorithm was run on the dataset to obtain a vec-

tor of topic proportions for each paper and a topic

correlation matrix. Each paper in the corpus is rep-

resented with a 20-dimensional vector containing the

prevalence of topics in that document. The correlation

matrix is a 20 × 20 matrix including the correlation

coefﬁcient for all topic pairs.

We calculate the novelty of a document as equal

to the most novel concept or combination of concepts

within that material (Grace et al., 2017). The novelty

of a document is the highest negative correlation co-

efﬁcient among all pairs of topics present in that doc-

ument (above a certain threshold), weighted by the

proportion of the document which contains that pair

(Grace et al., 2017). In order to determine whether

a topic is signiﬁcantly present in a document a topic

proportion threshold of 0.1 is used (i.e. the document

should be at least 10% of that topic). This novelty

formula is based on previous work in topic-model

approaches to novelty (Grace et al., 2017). Equa-

tion 6 shows the formula for a paper p given p =

, t

, . . . , t

] as the set of topics signiﬁcantly present

in p. The pair of topics with the highest negative cor-

relation coefﬁcient are denoted by t

and t

. This co-

efﬁcient is normalized against the most novel pair of

topics in the whole corpus (here denoted t

and t

) and

then weighted by the proportions of each topic in p to

calculate the novelty score.

NoveltyScore P

CovMat

)

CovMat

)

×2(min(prop(d, t

), prop(d, t

))) (6)

CovMat is the covariance matrix obtained from

the STM model. CovMat

)

is the correlation of the

document’s most atypical topic combination (t

, t

and CovMat

)

is the correlation of the most atypi-

cal topic combination among the whole corpus (t

and

). prop(d, t) is the proportion of document d that

consists of topic t. This corresponds to the novelty

of the document’s most novel topic combination, rela-

tive to the corpus’s most novel combination, weighted

by how much of the document consists of that com-

bination. The reason for using the minimum of the

Personalized Curiosity Engine (Pique): A Curiosity Inspiring Cognitive System for Student Directed Learning

two topic proportions rather than their average is to

prevent favoring documents that just passed the sig-

niﬁcance threshold with one topic, and were thus not

particularly novel in combining it with another, much

more weighted topic (Grace et al., 2017). In the Top-

ics Co-occurrence Model, the novelty rating for docu-

ments containing many relatively novel topic combi-

nations will be higher than for documents containing

only a little of the most novel pair of topics (Grace

et al., 2017). Equation 6 assigns a novelty score of

1 to a document that is made up of 50% of each t

and t

. The equation assigns 0 as the novelty score

to the document when the rarest topic combination

in it is independent, and assigns negative scores when

the only topic combinations above the threshold in the

document are positively correlated.

3.2.2 Sequence Composition

The purpose of the Sequence Composition subcom-

ponent is to take the novelty ratings of each document

in the corpus and construct a sequence of learning re-

sources that will maximize the chance of a student

experiencing optimal novelty. The Sequence Compo-

sition subcomponent generates a sequence of learn-

ing resources to support student-directed learning and

to stimulate students’ curiosity about learning. Dur-

ing the course of our project we explored three se-

quence generator models, which we call the ‘Origin-

Destination model’, the ‘Destination model’, and the

‘User-Directed model’.

Pique generates a personalized sequence of nine

documents in sets of three papers from the corpus of

learning resources in the Learning Materials compo-

nent, based on information from the Learner Model.

Students choose one paper from each set of three, read

it, and reﬂect on it before they are presented with the

next set. The different sequence generator models are

based on different representations of student interests.

The Destination Model uses a single set of

student-speciﬁed interests as the input to the algo-

rithm, while the Origin-Destination Model uses two

student-speciﬁed sets of keywords: one that they self-

report as already knowing about (the ‘origin’ set) and

one that they want to learn more about (the ‘desti-

nation’) set. The User-Directed Model builds on the

Origin-Destination Model to include other keywords

from the papers most recently selected by the student.

The sequence generator use these keywords to rep-

resent student preference, and combine that with the

novelty score for each paper to select and sequence

learning resources with the goal of inspiring curios-

ity.

Destination Model. The Destination Model priori-

tizes what students desire to learn. It recommends a

set of nine novel documents containing information

related to their stated desires. When used with our

keyword co-occurence novelty model, the student in-

terests can be directly mapped to corpus keywords.

When using the topic co-occurrence model, a map-

ping was manually built between the keyword set we

had constructed and the automatically generated topic

model topics. Here we refer to ‘novel documents’

generally, without specifying which of the novelty

models labeled them as such.

Students select their learning interests, which be-

come the destination set, D. The destination model

then identiﬁes candidate documents from the learning

materials corpus for which the top N topics within

that document include at least one of the user’s se-

lections. After some experimentation we decided on

N=3, as most documents in our corpus included at

least this many topics at reasonable proportions. From

this set of candidate documents the nine most novel

papers are selected and sorted based on their novelty,

with the most novel last. In this way the Destination

Model returns nine documents as output containing

information that students’ want to learn, starting with

a document of moderate novelty but then scaling up

to highly-novel documents as the student reads more

and learns about the topics they are interested in.

Origin-destination Model. The Origin-

Destination model intends to inspire students to

explore learning materials that contain some infor-

mation that they already know, combined with some

new information that they don’t. This is based on

ideas from educational psychology like Vygotsky’s

Zone of Proximal Development (Vygotsky, 1978), in

which new material is only learnable if it is at least

somewhat connected to topics already known. The

model selects a sequence that interpolates from what

the student already knows to what they want to know.

This recommendation generator is inspired from the

surprise walks algorithm (Grace et al., 2018) that

similarly tries to interpolate from an unsurprising

source to a surprising destination.

The Origin-Destination Model stimulates individ-

uals’ curiosity by presenting learning materials in

three steps: ‘close’, ‘far’, and ‘farther’. By rec-

ommending the learning materials step by step, the

model assists students to learn new materials similar

to what they already know and inspire them to explore

without recommending materials that are so novel as

to be unfamiliar or alien for them (Berlyne, 1966).

In the ﬁrst step, the model recommends papers that

are similar to what students already know and labels

CSEDU 2022 - 14th International Conference on Computer Supported Education

those papers as the ‘close’ category of learning ma-

terials for that student. In the second step, the model

recommends papers that are similar to both what the

students already know and what they want to learn,

and labels those papers as the ‘far’ category. Finally,

in the third step, the model recommends papers that

contain materials related only to what students want

to learn, and labels those papers as the ‘farther’ cate-

gory.

In the ‘close’ category, the model identiﬁes can-

didate papers that contain at least one common key-

word (or topic) from the students’ ‘source’ interest

set. The model uses the k-means algorithm and clus-

ters the candidate papers based on their novelty scores

to distinguish papers with three novelty levels: high,

medium, and low. The model also calculates the pa-

per’s familiarity score, which is the number of key-

words in common between the paper and the ‘origin’

set of topics/keywords the student already knows. The

papers with highest familiarity scores in each novelty

level are selected. The algorithm recommends one

low, one medium, and one high novelty paper

In the ‘far’ category step, the model recommends

another three papers intended to extend students’

learning from what they are familiar with to the new

topics they desire to learn. The candidate papers of

this category contain at least one common keyword

from the ‘origin’ keywords set (O) and at least one

common keyword from the destination keywords set

(D). The model uses the same clustering approach to

identify low, medium, and high novelty candidate pa-

pers, and identiﬁes the candidate paper in each level

with the highest number of common keywords.

In the ‘farther’ category step, the model presents

papers that contain information that students desire to

learn. The candidate papers of this category contain at

least one keyword from the destination keywords set

(D), and are categorized into three levels of novelty

just like the other two sets.

User-directed Model The User-Directed model is

an adaptation of the Origin-Destination Model that

considers students’ decisions during the recommen-

dation process to recommend materials aligned with

their evolving interests. The model recommends

papers step by step (close, far, and farther) as in

the Origin-Destination model, but additionally keeps

track of students’ selections of papers from the previ-

ous step. The keywords in the papers from the previ-

ous step are used to prioritize similar resources in the

recommendations of the next step.

Speciﬁcally, the User-Directed model ﬁlters the

candidate papers for the far step to those that share

at least one keyword papers selected in the close step.

Likewise, the model ﬁrst identiﬁes candidate papers

for the farther step that contain at least one keyword

match with the keywords of the paper selected in the

far step. Other than this additional ﬁltering, the model

is identical to the Origin-Destination model: it recom-

mends one low, one medium, and one high novelty

paper in each of the close, far, and farther steps.

3.3 The Learner Model Component

The Learner Model in Pique is primarily focussed

on collecting information about the learner to sup-

port the selection and presentation of learning ma-

terials as well as information needed to analyze the

use of Pique. The Learner model is not a comprehen-

sive model of the learner. This component stores two

kinds of information: information about the students

and how they have used Pique to date. Most infor-

mation about students remains constant: their name,

ID, email address, and course. The IDs are auto-

matically generated by the Pique system and serve

to de-identify students as required by our IRB ap-

proval. The ﬁnal component of the student proﬁle is

the one that can change: their interests, which they

select when they start using Pique but are prompted

to change each recommendation ‘cycle’. Each time

the student uses Pique, their data is updated with a

new cycle record, containing timestamps, the papers

they selected, the options they chose from, and their

reﬂections. Their reﬂections comprise responses to

three questions: 1) why the student has selected the

paper, 2) whether the selected paper matches their in-

terests, and 3) what topics the student expects to learn

from the paper when they read it.

3.4 The UX Component of Pique

The User Experience Component of Pique supports

students’ interaction with the following three subcom-

ponents: Selecting interests, Selection of papers, and

Reﬂection.

3.4.1 Selecting Interests

The Selecting Interests subcomponent captures stu-

dents’ interest by prompting them to identify what

they want to know. This prompt assists students to

formulate their learning goals and provide them more

control over their learning choices and enables self-

directed learning. Figure 2 shows the user interface

with the learning options for the students as they were

in the Graduate Teaching Seminar course.

Personalized Curiosity Engine (Pique): A Curiosity Inspiring Cognitive System for Student Directed Learning

Figure 2: UX for selecting interests in Pique.

3.4.2 Selection of Papers

The Selection of Papers subcomponent of Pique en-

ables students’ self-regulated learning, with the inten-

tion of stimulating their intrinsic motivation to learn

and explore. This module presents the papers that

are recommended by the sequence composition mod-

ule of Pique’s AI component. Pique presents the nine

papers in sets of three, based on the Sequence Com-

position subcomponent (see Section 3.2.2). Figure 3

shows an example of papers being recommended in

the Graduate Teaching Seminar course based on the

Orign-Destination sequence composition model: the

top three are closely related to what the student al-

ready knows, the middle three are related to both what

they know and what they are interested in, and the bot-

tom three what they are interested in only. The Selec-

tion of Papers subcomponent informs students about

how novel a particular paper is, and allows them to

manually choose more or less novel papers by select-

ing the drop-down menu labeled ‘show me papers’ in

the top right corner of Figure 3.

Figure 3: The UX for selecting learning content based on

interests and novelty scores.

3.4.3 Reﬂection

The third subcomponent of the Pique UX compo-

nent is Reﬂection. Cognitive studies of students

demonstrate that reﬂection is key to effective learning

[(Sch

on, 2017), (Kolb, 1999), (Cowan, 2006)]. The

Pique system includes two types of reﬂection: one ap-

pears when students select a paper to read (see Figure

4), and one at the end of the semester. The ﬁrst allows

them to reﬂect on their paper selection and to describe

what they expect to learn from it.

The second type of reﬂection asks students to re-

ﬂect on their overall learning experience. Students

are asked to summarize the papers they read and cat-

egorize those papers into groups. Students are asked

to identify the paper they found most interesting and

justify why. This reﬂection allows students to orga-

nize their newly acquired knowledge where the learn-

ing paths are constructed by the students rather than

the instructors. It was also critical for evaluating the

impact of this educational innovation on the student

experience.

Figure 4: Pique nudges students to reﬂect on their paper

selection and learning expectations.

4 THE COURSE EXPERIENCE

OF USING PIQUE

In this section, we present the experiences of students

who have used Pique in the classroom to respond to

the Research Questions presented in Section 1. We

used Pique over four semesters in both undergradu-

ate and graduate courses in Human Centered Design

as well as a PhD course, a Graduate Teaching Sem-

inar. In the Human Centered Design course students

were asked to use Pique for six weeks, and had to sub-

mit weekly and end of semester writing assignments

about the papers they had read. Each week they were

CSEDU 2022 - 14th International Conference on Computer Supported Education

asked to submit a summary of the three papers they

downloaded and read, and identify the most interest-

ing paper among the three. For the end of semester

report, the students were asked to describe their expe-

rience of using Pique, what they learnt, and the most

interesting paper they found (and why). For the Grad-

uate Teaching Seminar course, students were asked

to use the Pique system for the whole semester, but

submitted only a ﬁnal report without any weekly sub-

missions. This was due to the PhD students’ greater

familiarity with reading published articles, as well as

their overall greater autonomy as learners.

In response to our ﬁrst research question concern-

ing how the use of Pique helped enable self-directed

exploration we investigated how the student cohort

differed in the resources they explored, as a measure

of how self-directed their experiences were. In Table

1 we summarize our results. Though students’ op-

tions for selecting interests remain constant (39 and

55 interests in Human Centered Design and Teach-

ing Seminar courses, respectively), we found that stu-

dents were presented with very diverse sequences of

learning resources. Pique recommended a total of

621 unique papers for one semester of the Gradu-

ate Teaching Seminar course, even though that course

only included ﬁve students. 55% of those papers were

recommended to at least two students, due to overlaps

in topics of interest. Those ﬁve students selected a

total of 66 papers to read, with 86% of the selected

papers being selected by only one student. Across all

four courses we saw 72% of recommended papers be-

ing recommended to at least another student, but the

selections made by students were highly diverse, with

70% of the selected papers being unique to that indi-

vidual student.

Our second research question asked how using

Pique assisted students in expanding their learning in-

terests. In response to this we investigated the change

in students’ interests over time, illustrated in Figure

5. The top two sub-ﬁgures are for HCD courses

(Spring and Fall) and the bottom two sub-ﬁgures are

for Graduate Teaching Seminar courses (Spring and

Fall), where X-axis represent the number of Pique

cycles and Y-axis represents the average cumulative

growth of interest selections. We calculated the num-

ber of interests selected by each student during each

cycle. We aggregated this across students within a co-

hort to give the average number of interests selected

by the students in that cycle. The cumulative num-

ber of interests in Figure 5 shows the expansion of

stated interests over the semester. When starting to

use Pique, total students in the HCD courses selected

an average of just four interests. The searching of

learning interests increased as students used the Pique

system, and at the end of the semester all students had

explored an average of 67 interests. Similarly, total

students in the Graduate Teaching Seminar on average

started with only two interests, and over the semester

their average number of searching learning interests

increased to 42.

Figure 5: The growth of learning interests over the courses.

We noticed a difference between the students in

the HCD courses and those in the Graduate Teaching

Seminar, illustrated in Figure 6. The top two sub-

ﬁgures are for HCD courses (Spring and Fall) and

the bottom two sub-ﬁgures are for Graduate Teach-

ing Seminar courses (Spring and Fall), where X-axis

represent the number of Pique cycles and Y-axis rep-

resents the percentage of students searched for new

interests that they had not selected in earlier Pique

cycles. The students in the HCD courses were un-

dergraduate and graduate students who initially ex-

panded their learning interests and over time they re-

duced the number of new interests. In contrast, the

students in the Graduate Teaching Seminar courses

were PhD students who kept exploring new interests.

For instance, all the PhD students in the Fall semester

of Graduate Teaching Seminar continued to add new

interests until the end of the semester. 71% of Grad-

uate Teaching Seminar PhD students in the Spring

semester included new interest in their 8th Pique cy-

cle, but only 18% of the undergraduate students in the

HCI course continued exploring in the 8th Pique cy-

cle. This result indicates that students use the Pique

system differently to expend their learning selections.

After the students had ﬁnished using Pique, stu-

dents were asked to reﬂect on which paper from the

system they found most interesting and why. Two re-

searchers performed a thematic analysis on students’

written responses to identify meaningful patterns in

the data (Braun and Clarke, 2006). The use of mul-

tiple coders provided investigator triangulation to our

analysis (Patton, 1999). To establish a broad consen-

Personalized Curiosity Engine (Pique): A Curiosity Inspiring Cognitive System for Student Directed Learning

Table 1: Distribution of learning materials to personalize learning.

Course name

Graduate

Teaching

Seminar

Graduate

Teaching

Seminar

Human

Centered

Design

Human

Centered

Design

Semester Spring 2020 Fall 2020 Spring 2020 Fall 2020

Number

of students

24 5 12 12

Number of unique

learning sequences

generated by students

24 5 12 12

Total papers selected

by students over

the Pique cycles

221 66 76 77

% of selected papers

uniquely picked

by individuals

50%

(111 papers)

86%

(57 papers)

71%

(54 papers)

74%

(57 papers)

Total papers recommended

by Pique

1987 612 729 774

% of papers recommended

to at least one other

84%

(1669 papers)

54%

(333 papers)

75%

(548 papers)

72%

(558 papers)

% of papers recommended

to only one student

16%

(318 papers)

46%

(279 papers)

25%

(181 papers)

28%

(216 papers)

Figure 6: Percentage of students searching for new learning

interests over the pique cycles.

sus, the two researchers conducted a parallel coding

workshop on the ﬁrst 10% of the written responses.

After identifying this initial set of themes, they each

coded the rest of the data separately and then con-

verged on a set of collaboratively authored themes

through follow-up workshops.

Through this analysis we identiﬁed three major

themes underlying why most students found papers

interesting: novelty, personal relevance, and curios-

ity. The ﬁrst theme captured how students found pa-

pers interesting because of the novelty and innova-

tion of the idea presented in it. The second theme

captured how students found papers particularly in-

teresting when they could connect its contributions or

implications to their current work or personal experi-

ence. For example, one student found a paper describ-

ing a VR gaming application named ‘Spider Hero’ in-

teresting because he was a huge Spiderman fan. The

third theme captured how the recommended papers

made students curious about the research ﬁeld as a

whole (HCI or computing education) and helped to

grow their interest in the ﬁeld. For example, one stu-

dent expressed that they learned something new from

each of the recommended papers, and became so curi-

ous that they did extensive further personal research to

learn more about those speciﬁc topics. We observed

that this curiosity theme was related to the idea of stu-

dents connecting their class lessons with the recom-

mended papers. For example, one student learned the

concept of a ‘Wizard-of-Oz’ study through the HCI

class lectures and got excited when he found the same

concept in a research paper. Taken overall, these writ-

ten responses show that the recommended papers mo-

tivated students to explore and learn more in the do-

main.

5 CONCLUSION AND FUTURE

WORK

We present a cognitively inspired system architecture,

Pique, that presents students with personalized se-

quences of novel learning resources. We have shown

that these sequences encourage curiosity and may be

CSEDU 2022 - 14th International Conference on Computer Supported Education

helpful to support self-directed learning. Pique uses

computational models of novelty to identify docu-

ments from a corpus of learning materials that are

both relevant to the student’s interest and novel with

respect to the corpus. Pique is effectively an educa-

tional recommender system with a goal of inspiring

individuals’ curiosity to learn rather than shepherding

them through a speciﬁc curriculum.

Through our four-semester evidence-based explo-

ration of how to inspire students’ curiosity, we devel-

oped two separate computational models of novelty:

one based on keyword co-occurrence and one based

on the co-occurrence of topics from a topic model-

ing algorithm. These computational models rely on

the same underlying information theoretic approach

to novelty or surprise as features that negatively cor-

relate, but differ in the way we generated a keyword

or topic representation of the documents. Both had

their strengths and their weaknesses, and each was

able to identify some surprising-seeming papers that

the other missed. In future work we aim to explore

a variety of other approaches to representing our cor-

pus, including NLP and machine learning techniques.

These new representations would extend our current

computational models of learning resource novelty.

Throughout the Pique project we also developed

three recommendation models: one based only on

the student’s stated interests (their ‘destination’), one

based on taking them on a journey from what they

already knew (their ‘origin’) to their interests (their

‘destination’ again), and one based on blending that

origin-destination effect with similarity to the things

they’ve recently explored. Each of these recommen-

dation models combined student preferences with our

computational models of novelty to encourage curios-

ity in the learning process. We did not compare our

sequence recommendation models directly, although

we do believe, from the evidence of using them in the

classroom, that the latter two models both offer ad-

vantages over their predecessors.

This paper presents a proof of concept and deploy-

ment of Pique, a personalized curiosity engine for ed-

ucation. A limitation of our study is that we did not

collect data on how much time the students spent on

reading the papers before reﬂecting on them. From

evaluating the experiences of students who used Pique

extensively as part of their courses, we identiﬁed three

aspects that made recommended learning materials

interesting: how novel they were, how personally rel-

evant they were, and the curiosity and further self-

directed learning that they evoked. These ﬁndings are

evidence of how curiosity can be elicited from stu-

dents as part of a course experience, at least when

self-directed and open-ended engagement with learn-

ing resources is desirable. While we cannot claim

that student curiosity was entirely due to Pique, we

conclude that the approach of encouraging curiosity

Pique shows promise for future research on computa-

tional novelty in open-ended learning environments.

REFERENCES

Adamopoulos, P. and Tuzhilin, A. (2014). On unexpected-

ness in recommender systems: Or how to better ex-

pect the unexpected. ACM Transactions on Intelligent

Systems and Technology (TIST), 5(4):1–32.

Ai, F., Chen, Y., Guo, Y., Zhao, Y., Wang, Z., Fu, G.,

and Wang, G. (2019). Concept-aware deep knowl-

edge tracing and exercise recommendation in an on-

line learning system. International Educational Data

Mining Society.

An, S., Bates, R., Hammock, J., Rugaber, S., Weigel,

E., and Goel, A. (2020). Scientiﬁc modeling using

large scale knowledge. In International Conference

on Artiﬁcial Intelligence in Education, pages 20–24.

Springer.

Baker, T., Smith, L., and Anissa, N. (2019). Educ-ai-tion

rebooted? exploring the future of artiﬁcial intelligence

in schools and colleges. Retrieved May, 12:2020.

Barria-Pineda, J., Akhuseyinoglu, K., and Brusilovsky,

P. (2019). Explaining need-based educational rec-

ommendations using interactive open learner mod-

els. In Adjunct Publication of the 27th Conference

on User Modeling, Adaptation and Personalization,

pages 273–277.

Barria Pineda, J. and Brusilovsky, P. (2019). Making edu-

cational recommendations transparent through a ﬁne-

grained open learner model. In Proceedings of Work-

shop on Intelligent User Interfaces for Algorithmic

Transparency in Emerging Technologies at the 24th

ACM Conference on Intelligent User Interfaces, IUI

2019, Los Angeles, USA, March 20, 2019, volume

2327.

Berlyne, D. E. (1966). Curiosity and exploration. Science,

153(3731):25–33.

Blei, D. M. and Lafferty, J. D. (2007). A correlated topic

model of science. The annals of applied statistics,

1(1):17–35.

Bouma, G. (2009). Normalized (pointwise) mutual in-

formation in collocation extraction. Proceedings of

GSCL, 30:31–40.

Boyle, G. J. (1983). Critical review of state-trait curiosity

test development. Motivation and Emotion, 7(4):377–

397.

Braun, V. and Clarke, V. (2006). Using thematic analysis

in psychology. Qualitative Research in Psychology,

3(2):77–101.

Carter, A. S. and Hundhausen, C. D. (2011). A review of

studio-based learning in computer science. Journal of

Computing Sciences in Colleges, 27(1):105–111.

Cobos, C., Rodriguez, O., Rivera, J., Betancourt, J., Men-

doza, M., Le

on, E., and Herrera-Viedma, E. (2013).

Personalized Curiosity Engine (Pique): A Curiosity Inspiring Cognitive System for Student Directed Learning

A hybrid system of pedagogical pattern recommenda-

tions based on singular value decomposition and vari-

able data attributes. Information Processing & Man-

agement, 49(3):607–625.

Cowan, J. (2006). On becoming an innovative university

teacher: Reﬂection in action: Reﬂection in action.

McGraw-Hill Education (UK).

Del Solato, T. and Du Boulay, B. (1995). Implementation

of motivational tactics in tutoring systems. Journal of

Artiﬁcial Intelligence in Education, 6:337–378.

Grace, K. and Maher, M. L. (2015a). Speciﬁc curiosity as a

cause and consequence of transformational creativity.

In ICCC, pages 260–267.

Grace, K. and Maher, M. L. (2015b). Surprise and reformu-

lation as meta-cognitive processes in creative design.

In Proceedings of the third annual conference on ad-

vances in cognitive systems ACS, page 8.

Grace, K., Maher, M. L., Davis, N., and Eltayeby, O.

(2018). Surprise walks: Encouraging users towards

novel concepts with sequential suggestions. In Pro-

ceedings of the 9th International Conference on Com-

putational Creativity (ICCC 2018). Association of

Computational Creativity.

Grace, K., Maher, M. L., Mohseni, M., and PEREZ

Y PEREZ, R. (2017). Encouraging p-creative be-

haviour with computational curiosity. ICCC.

Hall Jr, O. P. and Ko, K. (2008). Customized content deliv-

ery for graduate management education: Application

to business statistics. Journal of Statistics Education,

16(3).

Hill, J. R. and Land, S. M. (1998). Open-ended learning

environments: A theoretical framework and model for

design.

Jones, A. and Issroff, K. (2005). Learning technologies:

Affective and social issues in computer-supported

collaborative learning. Computers & Education,

44(4):395–408.

Kang, M. J., Hsu, M., Krajbich, I. M., Loewenstein,

G., McClure, S. M., Wang, J. T.-y., and Camerer,

C. F. (2009). The wick in the candle of learning:

Epistemic curiosity activates reward circuitry and en-

hances memory. Psychological science, 20(8):963–

973.

Kashdan, T. B. and Fincham, F. D. (2004). Facilitating cu-

riosity: A social and self-regulatory perspective for

scientiﬁcally based interventions.

Keller, J. M. (1987). Strategies for stimulating the motiva-

tion to learn. Performance and instruction, 26(8):1–7.

Keuning, H., Jeuring, J., and Heeren, B. (2018). A system-

atic literature review of automated feedback genera-

tion for programming exercises. ACM Transactions

on Computing Education (TOCE), 19(1):1–43.

Kolb, D. A. (1999). Learning style inventory. McBer and

Company Boston, MA.

Kose, U. and Arslan, A. (2016). Intelligent e-learning sys-

tem for improving students’ academic achievements

in computer programming courses. The International

journal of engineering education, 32(1):185–198.

Krajcik, J. and Blumenfeld, P. C. (2006). The cambridge

handbook of the learning sciences. r. keith sawyer.

Kuo, R., Krahn, T., and Chang, M. (2021). Behaviour

analytics-a moodle plug-in to visualize students’

learning patterns. In International Conference on In-

telligent Tutoring Systems, pages 232–238. Springer.

Loewenstein, G. (1994). The psychology of curiosity: A

review and reinterpretation. Psychological bulletin,

116(1):75.

Merrick, K. E. and Maher, M. L. (2009). Motivated rein-

forcement learning: curious characters for multiuser

games. Springer Science & Business Media, .

Niu, X., Abbas, F., Maher, M. L., and Grace, K. (2018).

Surprise me if you can: Serendipity in health informa-

tion. In Proceedings of the 2018 CHI Conference on

Human Factors in Computing Systems, pages 1–12.

Papamitsiou, Z. and Economides, A. A. (2014). Learn-

ing analytics and educational data mining in prac-

tice: A systematic literature review of empirical ev-

idence. Journal of Educational Technology & Society,

17(4):49–64.

Patton, M. (1999). Enhancing the quality and credibil-

ity of qualitative analysis. Health services research,

34:1189–1209.

Pozdniakov, S., Posov, I., and Anton, C. (2021). Interaction

of human cognitive mechanisms and “computational

intelligence” in systems that support teaching math-

ematics. In International Conference on Intelligent

Tutoring Systems, pages 259–266. Springer.

Roberts, M. E., Stewart, B. M., and Tingley, D. (2019). Stm:

An r package for structural topic models. Journal of

Statistical Software, 91(1):1–40.

Sampson, D. and Karagiannidis, C. (2002). Personalised

learning: educational, technological and standarisa-

tion perspective. Digital Education Review, (4):24–

39.

Schmidhuber, J. (2010). Formal theory of creativity, fun,

and intrinsic motivation (1990–2010). IEEE Transac-

tions on Autonomous Mental Development, 2(3):230–

247.

Sch

on, D. A. (2017). The reﬂective practitioner: How pro-

fessionals think in action. Routledge.

Vygotsky, L. (1978). Interaction between learning and de-

velopment. Readings on the development of children,

23(3):34–41.

Wood, D. F. (2003). Problem based learning. Bmj,

326(7384):328–330.

Zawacki-Richter, O., Mar

ın, V. I., Bond, M., and Gou-

verneur, F. (2019). Systematic review of re-

search on artiﬁcial intelligence applications in higher

education–where are the educators? International

Journal of Educational Technology in Higher Educa-

tion, 16(1):1–27.

CSEDU 2022 - 14th International Conference on Computer Supported Education