The Use of Time Dimension in Recommender Systems for Learning

Eduardo José de Borba

, Isabela Gasparini

and Daniel Lichtnow

Graduate Program in Applied Computing (PPGCA), Department of Computer Science (DCC),

Santa Catarina State University (UDESC), Paulo Malschitzki 200, Joinville, Brazil

Polytechnic School, Federal University of Santa Maria (UFSM), Av. Roraima 1000, Santa Maria, Brazil

Keywords: Recommender System, Context-aware, Time, Learning.

Abstract: When the amount of learning objects is huge, especially in the e-learning context, users could suffer

cognitive overload. That way, users cannot find useful items and might feel lost in the environment.

Recommender systems are tools that suggest items to users that best match their interests and needs.

However, traditional recommender systems are not enough for learning, because this domain needs more

personalization for each user profile and context. For this purpose, this work investigates Time-Aware

Recommender Systems (Context-aware Recommender Systems that uses time dimension) for learning.

Based on a set of categories (defined in previous works) of how time is used in Recommender Systems

regardless of their domain, scenarios were defined that help illustrate and explain how each category could

be applied in learning domain. As a result, a Recommender System for learning is proposed. It combines

Content-Based and Collaborative Filtering approaches in a Hybrid algorithm that considers time in Pre-

Filtering and Post-Filtering phases.

1 INTRODUCTION

Nowadays, there are distinct educational approaches,

e.g. online learning, blended learning, face-to-face

learning, etc. All these approaches can benefit from

the learning management systems, (also called e-

learning systems) for the administration,

documentation, tracking, reporting and delivery of

electronic educational technology. In e-learning

systems when the number of available materials and

learning objects is huge, students may feel lost and

may not find relevant objects to study. Moreover,

there is great probability that some learning

materials never get studied.

For this purpose, researchers have applied

personalization techniques to select the best items

for each student, considering student’s knowledge,

goals, preferences and needs (Brusilovsky, 1998).

Recommender Systems (RS) can help with these

problems, suggesting items to users using

information they have about users and items and

about how item characteristics meet users’ needs.

Context-aware RS (CARS) are an evolution of

traditional RS that apply context information to

improve the quality of recommendations. Among all

dimensions that represent context, time has the

advantage of being easy to capture and has the

potential to improve the quality of recommendation

(Campos et al., 2014).

This work aims to identify how Time-aware RS

(Context-aware RS that uses time context) can be

used in the learning domain. For this purpose, seven

categories of how time can be used in RS are

presented, based on previous works in the area

(Borba et al., 2017). Based on these seven

categories, the application of time in the e-learning

domain is presented through different scenarios.

This work is organized as follows. Section 2

presents Background of this research. Section 3

details the seven categories on the use of time in RS.

Section 4 discusses the Related Works. Section 5

defines scenarios for each of the categories

introduced in section 3. Section 6 presents a

Proposal that applies time context in Recommender

Systems and Section 7 presents Conclusions and

Future Work.

2 BACKGROUND

This section presents the main concepts related to

Recommender Systems for Learning. Firstly, the

definition of Recommender Systems and their

traditional approaches is presented. Followed by

600

Borba, E., Gasparini, I. and Lichtnow, D.

The Use of Time Dimension in Recommender Systems for Learning.

DOI: 10.5220/0006312606000609

In Proceedings of the 19th International Conference on Enterprise Information Systems (ICEIS 2017) - Volume 2, pages 600-609

ISBN: 978-989-758-248-6

Context-aware Recommender Systems and Time-

Aware Recommender Systems.

2.1 Recommender Systems

Recommender Systems are computational tools that

provide personalized suggestions to users (Ricci et

al., 2011). This means that as recommendation each

user receives a different set of items based on his/her

preferences. In recent years, interest in

Recommender Systems applications is growing

strongly (Adomavicius and Tuzhilin, 2005; Beel et

al., 2016). Examples of these applications are

recommendation of Books, CDs, DVDs, etc., in e-

commerce like Amazon or EBAY, recommendation

of movies like MovieLens or Netflix;

recommendation of songs in music websites like

Last.fm or Spotify; friend’s recommendations in

social networks like Facebook.

Recommender Systems emerged as an

independent area in the mid-1990s (Adomavicius

and Tuzhilin, 2005). Others areas are usually

involved, e.g., Information Retrieval, Approximation

Theory, Artificial Intelligence, etc.

Recommender Systems are formally represented

as follows:

:  → 

Where F is the function that predicts the rating

for an unknown item, U represents the users, I

represents the items and R denotes an ordered set of

predicted ratings.

Traditional approaches of Recommender

Systems are (Adomavicius and Tuzhilin, 2005):

Content-based, Collaborative Filtering, Knowledge-

based, Demographic and Hybrid.

Content-based is the approach where the user

receives a recommendation of items similar to the

ones he had interest in, in the past (Lops et al.,

2011). It usually consists of comparing the

description of the items (a set of keywords) to the

users’ profile (another set of keywords) and

recommending the most suitable item(s). That is

why this approach is related with Information

Filtering techniques, like TF-IDF or Cosine

(Adomavicius and Tuzhilin, 2005). The main

advantages of Content-based approach are (Lops et

al., 2011): (1) no dependence on an active

community of users and (2) no item cold-start. The

main drawbacks of this approach are (Lops et al.,

2011): (1) user cold-start and (2) overspecialization.

Collaborative Filtering approach recommends

items to a user based on what other users - with

similar tastes - have interest in (Jannach et al.,

2011). It is the automatization of “word of mouth”,

where the RS tries to predict item utility to the user

based on the utility of this item to users with similar

tastes to him/her. The main advantage of this

approach is Serendipity (Jannach et al., 2011). The

main drawbacks are (Jannach et al., 2011): (1)

dependence on an active community of users, (2)

User cold-start, (3) Item cold-start and (4) Black

sheep.

Knowledge-based approach recommends items

to users based on the knowledge about how item

features matches user needs and how useful this item

should be (Felfernig et al., 2011). This approach is

usually applied to improve the recommendation

precision or in cases where the other approaches

have problems. This approach should be chosen

where domain allows the representation of

knowledge through structures easy read by

computers, like ontologies (Adomavicius and

Tuzhilin, 2005). The main drawback of this

approach is that it needs the knowledge acquire

(Adomavicius and Tuzhilin, 2005).

Demographic approach recommends based on

the user’s demographic profile, like age, gender,

nationality, etc. This approach uses a

recommendation by demographic classes, in which

users are classified through stereotypes (Burke,

2002). It considers that different recommendations

should be made to different stereotypes. The main

advantage of this approach is to recommend items

according users age, gender, culture, etc. (Burke,

2002). The main drawbacks are (Burke, 2002): (1)

assuming that users with similar demographics have

similar tastes, (2) there are few works in literature

about this.

Hybrid approach combines the mentioned

approaches to recommend items to users. The

objective is to group the advantages of these

approaches to improve the recommendation quality

and with fewer drawbacks of any individual one

(Burke, 2002). Burke (2002) suggests some

combinations of the approaches, for example:

Weighted, Switching, Cascade and Mixed. In

Weighted, the predicted ratings of several

recommendation techniques are combined and each

one has a different weight. In Switching, the system

changes through different recommendation

techniques depending on the current situation. In

Cascade, one recommender refines the

recommendations given by another. In Mixed, all

combined approaches are used and the results are

presented in the same ranking.

The Use of Time Dimension in Recommender Systems for Learning

601

2.2 Context-Aware Recommender

Systems

Traditional RS considers only users and items to

recommend, but it does not consider the context in

which the users are. According to Dey (2001),

context is any information that can be used to

characterize the situation of an entity. In RS, entities

can be the users and the items.

Context-Aware RS are formally represented as:

:  → 

Where F is the function that predicts the rating

for an unknown item, U represents the users, I

represents the items, C represents the context and R

denotes an ordered set of predicted ratings.

Several authors defined different set of

dimensions that could represent context (Schilit et

al., 1994; Chen and Kotz, 2000; Zimmermann et al.,

2007). In this work, we follow Schimidt et al. (1999)

that defines the following dimensions:

 Information on the user, e.g., users’ habits,

users’ emotional state, etc.;

 User’s social environment, e.g., co-location

with others users, social interaction in

social networks, etc.;

 User’s tasks, e.g., general goals, whether it

is a defined task or random activity, etc.;

 Location, e.g., absolute position, whether

the user is at home or office, etc.;

 Physical conditions, e.g., noise, light, etc.;

 Infrastructure, e.g., network bandwidth,

type of device, etc.;

 Time, that could be categorical, e.g., Time

of the day (Morning, Afternoon, Evening),

or continuous, e.g., a timestamp like “June

, 2016 at 17:14:36”.

Adomacivius and Tuzhilin (2011) define three

paradigms of context in the recommendation

process:

 Contextual Pre-Filtering, where the context

filters the data that represents the user and

then a traditional RS approach is applied;

 Contextual Post-Filtering, where a

traditional RS approach is applied and then

the result is filtered according to the

context;

 Contextual Modelling, in which the context

is applied directly in the recommendation

algorithm.

Verbert et al. (2012) say that, in e-learning, RS

traditional approaches are not enough to recommend

properly items to students, because this domain

offers some specific characteristics that are not

covered by these approaches. For example, it is

much more dangerous recommend a bad material to

a student, which could demotivate him/her to study,

than recommend a bad product in an e-commerce

system. According Verbert et al. (2012) this

application domain requires a major level of

personalization. Using some context dimensions is

an alternative to improve the personalization of e-

learning environments, recommending properly to

actual student situation, e.g., Learning History,

Environment, Timing and Accessible Resources

(Verbert et al., 2012).

The next section presents a specific kind of

Context-Aware Recommender Systems that uses

time context to recommend. This kind of RS could

also be used with others context dimensions.

2.3 Time-Aware Recommender

Systems

Among all context dimensions, time has an

advantage to be easy to capture, considering that

almost every device has a clock that could capture

the timestamp when an interaction occurs. Besides

that, works in this area showed that the context of

time has potential to improve recommendation

quality (Campos et al., 2014). This kind of RS is

called Time-Aware Recommender Systems (TARS).

TARS are formally represented as:

:  → 

Where F is the function that predicts the rating

for an unknown item, U represents the users, I

represents the items, T represents time context and R

denotes an ordered set of predicted ratings.

According to Merriam-Webster dictionary,

time is “a non-spatial continuum that is measured in

terms of events that succeed one after another from

past through present to future” (Merriam-Webster,

2016). This enables to establish an order to time

events.

As seen in section 2.2, time may be a continuous

or a categorical variable. Continuous variables are

those that represents the exact time at which items

are rated/consumed (Campos et al., 2014).

Categorical variables are calculated regarding time

periods of interest in the recommendation (Campos

et al., 2014). Also, it can be represented in several

time units, e.g., seconds, minutes, hours, days,

months, years, etc. Time units are hierarchical, e.g.,

1 day has 24 hours, 1 hour has 60 minutes, 1 minute

has 60 seconds.

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

602

3 USE OF TIME CATEGORIES

A Systematic Mapping was conducted in previous

work (Borba et al., 2017), using Peterson et al.

(2008) methodology, aiming to explore Time-Aware

Recommender Systems. The main research question

defined was: How the time is used in Context-aware

Recommender Systems? To answer the main

research question, three secondary research

questions were defined: (1) How recommender

algorithms use time? (2) What are the differences

about the use of time in different application

domains? (3) What others dimensions of context are

used to be applied together with the time dimension?

It is important to emphasize that this previous

work did not consider only e-learning recommender

systems. After all the process of selection of papers,

88 papers were considered to answer the research

question. Between another analysis, like

recommendation approach (Content-based,

Collaborative Filtering, Hybrid, etc.) or

representation of time (categorical or continuous), in

this previous work, it were observed seven

categories of how the time is used in Recommender

Systems. Here, an overview of these seven

categories, following (Borba et al., 2017), is

presented. Section 5 details each one in depth, and

explains how it could be applied in RS for e-

learning.

 Restriction: the time is used to restrain

which items are recommended. Thus, the

RS matches the user’s available time with

time required to use the item. Examples:

recommend only restaurants that are open

when user’s going to have lunch;

recommend a movie with length less or

equal to user’s available time.

 Micro-profile: the user has distinct profiles

for each time. Here, time is usually

categorical, so the user has a profile for

weekdays and another profile for weekends,

or the user has a profile for morning, a

profile for afternoon and another for

evening. Example: recommend a mobile

app to the user at Sunday morning based

only in apps used by this user in past

Sunday mornings.

 Bias: time is the third dimension of a User

x Item matrix. So, collaborative filtering

has more information to compare users,

find k-neighbours and predict user’s rating

to a non-viewed item. Example: Koren

(2009) proposes a Tensor Factorization

strategy using a User x Item x Time tensor.

 Decay: the time is used as a decay factor, in

which old interactions are less important

than new ones. Example: take into account

the use of RS techniques in E-commerce,

consider items the user searched recently

more important for producing

recommendation.

 Time Rating: time is considered by the RS

to infer user’s preferences, i.e., the more the

user stays at the item, more he likes it. It

means that time gives feedback of a user to

an item implicitly, i.e., without need of user

rate the item. Example: at a shopping mall,

the more a user stays at some store, more

he likes it, and the RS could recommend

products of this store when it has sales.

 Novelty: only new items could be

recommended. Thus, the RS has a threshold

and items older than that are not

recommended. Example: in news website,

it’s more precise to recommend news of, at

least, one day ago.

 Sequence: the RS observe items that are

usually consumed one after the other. So, if

the first one of the sequence is consumed,

the second one should probably be

consumed too. Example: in music

recommendation, songs of the same album

are most likely to be heard together, so if

the user selects one of them, the next one

should be recommended.

Figure 1 presents recommendation process in

CARS. Extraction and Recommendation

phases

are common to every RS, while Pre-Filtering and

Post-Filtering are more related to CARS.

In this model, Time Rating category appears in

the first phase, as information extraction in an

implicit approach. Decay and Micro-profiles appears

in Pre-Filtering phase, and in this phase, we could

include many other Pre-Filtering strategies using

other context dimensions. Bias category appears in

Recommendation phase, always applied with

Collaborative Filtering, or Hybrid RS that uses

Collaborative Filtering. Finally, Novelty, Restriction

and Sequence categories are classified in Post-

Filtering phase, and in this phase, we could also

include other Post-Filtering strategies using other

context dimensions.

It is worth saying that is not necessary to use all

these categories of the use of time. It is possible to

implement a TARS that uses just one category, or a

combination of two or more strategies.

The Use of Time Dimension in Recommender Systems for Learning

603

Figure 1: Recommendation model with use of time

categories.

4 RELATED WORK

In this section are presented works that apply time

context in the recommendation process. These works

were chosen to represent some of the seven

categories described in section 3, but not all

categories have works in e-learning domain.

The first work is described in Gallego et al.

(2012) that define a proactive Context-Aware RS

that recommends learning objects to teachers and

scientists that will produce learning resources the

students will consume. The recommendation process

is divided into three phases: (1) Generation of social

context information related to the users in the

environment; (2) Current situation is analysed,

considering User’s social environment, Location,

Time and User’s task; (3) Suitability of learning

materials to be recommended is analysed. Time in

this work is used like Restriction; the RS tries to find

learning materials that matches actual user’s time.

Chen et al. (2012) propose a hybrid RS to

recommend learning items in users’ learning

processes. The proposed method consists of two

steps: (1) discovering content-related items using

Collaborative Filtering approach; and (2) applying

the item sets to sequential pattern mining algorithm

to filter items according to common learning

sequences. Time in this work is used as Sequence,

RS compares items in common learning sequences

in order to find items that are usually consumed

together and decide which one recommend.

Luo et al. (2009) propose a Context-Aware

resource recommendation model and relevant

recommendation algorithm to pervasive learning

environments. The calculation of relevant items to

be recommended can be divided in two: (1) Content-

based and Collaborative Filtering are combined

together, meanwhile learners’ historical sequential

patterns of resource accessing are also considered to

further improve the accuracy of recommendation;

(2) Connection type and Time satisfaction degree are

calculated, considering other relevant contexts. The

two parts are combined and results in Top-N items

recommended. Time in this work is used as

Sequence, considering learning materials order to

improve recommendation, and Decay, giving less

weight to older accesses.

Benlamri and Zhang (2014) propose a

knowledge-driven recommender for mobile learning

on the Semantic Web. This work uses an approach

for context integration and aggregation using an

upper ontology space and a unified reasoning

mechanism. Time is used as Restriction, where

learning resources have expected learning time and

the approach considers this to recommend more

properly to each user.

In Related Works, it is possible to observe that,

in learning, most common uses of time is related to

Restriction, Sequence and Decay categories. We do

not find other categories of use of time in the

learning domain.

5 USE OF TIME IN LEARNING

RECOMMENDER SYSTEMS

In this section, we describe in depth each category

described in section 3 and see how each of them

could be used in RS for e-learning. For this purpose,

we use scenarios to represent possible interactions of

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

604

learners with an e-learning system. Each category is

demonstrated in a distinct scenario.

We consider that recommender system can help

in different educational approaches and situations.

For example, in online learning, there are courses

that are open and do not have schedule to start or to

finish (i.e. classes that student can study whenever

he/she wants). But also there are courses where there

is a schedule to start and to finish. Both cases can

benefit of recommender systems. In blended

learning of face-to-face learning, the e-learning

system can be used as a support of the classes, e.g.

in the classroom – with technological resources, or

out of the class – students homework, additional

study, etc.

We consider this distinction of the different

learning situations important, to present some

scenarios in the next sections.

5.1 Restriction

The Restriction category uses the time to restrain

which items are recommended. To understand how

this category could be applied in learning domain,

we consider a student who is going to study. The

environment asks the student about how much time

he intend to study. The student indicates that he is

going to study for 3 hours (it is possible to thing in a

system that uses information about user to infer this

without ask the user about available time for

studying).

The RS knows about the items that the student

already accessed, so it should suppose that these are

no longer need by the user. The RS also knows that

the learning style of this user is visual, so it tries to

recommend only videos to him/her. If no video is

available, then the RS recommends other types of

items. After applying a traditional RS approach to

select the videos that best matches the user profile,

the RS filters the list of recommendation, removing

all videos that goes over 3 hours.

Thus in scenario 1, the student watches a video

that is 1h45min longer. Then, the student has

1h15min left. The next time the student asks for

recommendations, the RS filters videos that go up to

1h15min. This process goes on and on until

student’s available time is over.

5.2 Micro-profile

Micro-profile category uses time to create different

profiles of a user. Then, the user must have two or

more profiles, depending on time, and the RS selects

which profile are going to be used, based on some

criteria.

Thus, in scenario 2 there is a RS in an learning

management system that uses Content-based

approach. This RS represents students’ profile as a

set of keywords and each of the items another set of

keywords. Student’ keywords come from the items

he/she liked (rated positively). However, item’s

keywords are the words that most appeared in the

material and it is discovered through an algorithm

called TF-IDF (Term Frequency – Inverse

Document Frequency).

In this scenario, a teacher uses the learning

management system described above to support

his/her in-person classes (face-to-face learning).

He/she provides papers, presentations, links, games,

etc., that may help student while studying.

The RS using Micro-profile strategy could split

student’s profile in three. One for the time (period)

of the classes, other for weekdays (regarding the

time out of the class), and other for the weekends.

The RS knows what items the student access in each

of these time periods. Then, it will recommend items

during the face-to-face classes based on items the

student accessed during classes, will recommend to

him in weekdays out of class based on items

accessed in this period and recommend in weekends

based on weekend’s access, using Content-based

approach.

The RS might found out that, for example, one of

the students likes to see complementary materials,

like presentations while in the classroom to go along

with teacher presentation. However, he likes more

complete and complex materials to study in depth

the subject while on weekdays. Moreover, the

student wants short videos in the weekends where

he/she will not spend much time studying. These

preferences are reflect in each of students’ profiles,

so the recommender system is going to understand

them and improve its recommendations.

5.3 Bias

In Bias category, time is the third dimension of the

User x Item matrix, and it is only applied in

Collaborative Filtering approach. Time improves the

process of finding the k-users more similar to the

one that will receive recommendation and the

prediction of ratings to non-viewed items.

Thus, in scenario 3 there is an online course,

available for six months, and totally non-presential.

Despite of there is a schedule to end the course, the

system allows students entry at any time. Two users

started the course in different times: John started two

The Use of Time Dimension in Recommender Systems for Learning

605

months ago and Stuart started five months ago. Also,

both use to study 2 hour by day. Items that Stuart

accesses now are probably different of the items

John accesses now, because Stuart is forward in the

course.

In this case, the RS using time as Bias category

compares John’s profile today with Stuart’s profile

of three months ago (when he was also on the

second month of the course). When comparing these

two profiles, the RS finds out that John and Stuart

are similar users, the RS can use Stuart ratings of

three months ago to recommend items that John

might like. Using this strategy, the RS does not

recommend items to John based on what he is

studying now, that could be too advanced to him.

Instead, recommends items that are probably on the

same topic of what John is studying, based in the

assumption that these two users are considered

similar.

5.4 Decay

Decay category uses time as decay factor to user’s

interaction, i.e., the older the interaction, less

important it is.

Thus, in scenario 4, student Frank is enrolled in

a discipline of Data Structure, that lasts one semester

and that has four main topics (stack, queue, list and

tree). In this discipline, there are four tests, one for

each topic. Also, suppose that the RS in this

discipline uses Content-based approach, i.e.,

recommends items similar to the ones the student

accessed.

Before the first test, Frank only studied items

about stacks, so he might receive only

recommendations about stack. After the first test,

Frank starts studying the second topic: queues. If the

RS keeps recommending only stacks, the user will

probably not like the recommendations he receives.

The RS, using Decay category, gives less weight

to the old items that Frank accessed about stacks and

gives more weight to the new items about queues.

There still possibilities that Frank receives

recommendations about stacks, but the RS are going

to prioritize the new items about queues to

recommend materials.

5.5 Time Rating

Time rating category is related to the fact that RS

uses time to understand user’s preferences. For

example, if the user stays a long time in an item, this

means to the RS that the user likes it or has

interesting in it. But if the user stays small time in

the item, it means that the user doesn’t like it or does

not have interesting in it.

Thus, in scenario 5, there is a RS in an e-

learning environment that uses Collaborative

Filtering approach. This approach requires an active

community of students and requires feedback of the

Students to the items. The feedback is usually made

through explicit rating, e.g., from 1 to 5 stars.

Suppose that this environment have an active

community, but the students rarely rate the items

that they access. In this scenario, time is useful to

receive implicit feedback about how the student

liked this item.

To exemplify, considering two students Anna

and Bruce and that all items were created with the

same length and therefore the students will spend the

same time to use an item. A problem that must be

considered by the RS is that some students usually

stay more in some items, while others stays less in

the same items. Anna has an average time of 30

minutes per item. Bruce has an average time of 5

minutes per item. If Anna stays 15 minutes in the

item I, this probably means she did not likes much

this item. If Bruce stays 15 minutes in the item I, this

probably means he liked this item.

To treat this problem the RS could be use the

following equation to calculate how much the

student u like item i, similar to the bias strategy used

by Koren (2009):



,





,







Where 

,

is the rating calculated of user u to

item i, 

,

is the time (in minutes) that the user u

stayed in the item i, and 





is the average time of

user u in the items of the system. This equation

express that if the user u stays in the item i more

than its average, the rating calculated is more than 1,

so this means the user like it. But, the user u stays in

item i less than its average, the rating is between 0

and 1, meaning the user didn’t like it.

5.6 Novelty

Novelty category uses time to filter only new items

to be recommended. The RS knows when the items

were created and compares them with actual time to

decide if the item should or not be recommended.

Thus, in scenario 6 the student Fernando signs-

up to a course about new technologies, like HTML

5, CSS 3, Ruby on Rails, Angular, etc. This subject

is in constant changing, because these actual

technologies are been updated and upgraded very

frequently. This course is updated every time one of

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

606

these technologies changes and the course

administrator tells the system in metadata when this

item was created.

The RS, using Novelty category, would

recommend only items that are new to the system

and up-to-date with the technologies. This kind of

RS in e-learning does not define a threshold like

news recommendation, because does not make sense

to ignore items just from it timestamps. But if two

items are similar to each other, the newer will be

recommended. Also, if some item is too old

comparing with the others, it would never be

recommended.

5.7 Sequence

Sequence category observes items that are used in a

sequential pattern to improve recommendation

process. If the RS observes a set of items that is

accessed in a specific order, the system should

recommend them in this order to the user.

This, in scenario 7 there is a short course of

Algorithms that occurs one time for semester and

lasts a month. This short course has 30 materials

numbered from 1 to 30, within papers, links, images,

videos, etc.

In the first semester of 2016, 80 students enrolled

the short course and RS observed that the learning

path most frequent was:

1→9→7→15→23→12→25

In the second semester of 2016, the RS, using

Sequence category, starts the short course

recommending the item 1, followed by item 9, and

so on. This means RS uses the learning path learnt

from the last semester to recommend material to

new students.

6 PROPOSAL SYSTEM

Taking into account the scenarios presented in

section 5, it is possible to define a recommender

system architecture that consider time in its

recommendation process. In this sense, the system

can use time in distinct ways aiming to improve the

quality of recommended items.

6.1 Architecture Overview

The components of system architecture consists of a

module to maintain learning objects (include,

remove, update and assign metadata) and a

recommender module.

6.1.1 Maintain Learning Objects Module

This module allows inserting and maintaining

learning objects. The learning objects are described

using LOM – Learning Object Metadata (IEEE,

2002). Taking into account LOM metadata allows

improving the recommendation in some scenarios

presented in section 5.

Thus, metadata 2.3.3 of LOM (Date, that stores

when a learning object were created and indicate

how old a learning object is), is very important to

Novelty that considers more important to

recommend newer items. Besides metadata 4.7

(Duration Time, continuous time a learning object

takes when played at intended speed) and 5.9

(Typical Learning Time, approximate or typical time

it takes to work with or through this learning object

for the typical intended target audience) are very

important in Restriction.

6.1.2 The Recommender Module

The Recommender Module consists of a module that

uses a Hybrid Approach to generate the

recommendation. This hybrid approach combines

Collaborative Filtering and Content-Based approach.

This is necessary because we consider distinct

use: (1) in short courses with few users/students and

(2) long duration courses with many students where

it is possible to build a user community. In (1)

content-based approach seems more suitable,

because it is more difficult to have enough users to

generate the recommendation. While in (2)

Collaborative Filtering can be applied and may

improve recommendation quality as time goes by.

Taking into account Burke (2002), the

hybridization method chosen is mixed, where

recommendation from several different

recommenders are presented at the same time. Burke

(2002) emphasizes that “mixed hybrid avoids the

“new item” start-up problem: the content-based

component can be relied on to recommend new

shows on the basis of their descriptions even if they

have not been rated by anyone.” It does not get

around the “new user” start-up problem, since both

the content and collaborative methods need some

data about user preferences to get off the ground, but

if such a system is integrated into other source of

information (e.g. social network, digital television,

etc.) it could track user’s behaviour and build his

profile accordingly.

The Proposal System applies Collaborative

Filtering and Content-Based approaches separately,

each of them applying Pre-Filtering and Post-

The Use of Time Dimension in Recommender Systems for Learning

607

Filtering with time dimension. For Pre-Filtering, the

RS uses Decay category and, for Post-Filtering, uses

Restriction and Novelty. The last one (Novelty) is

only used if specified by the course manager,

because it is specific to some subject.

In Decay Pre-Filtering, each user interaction

(consumption or rating) are evaluated in terms of

how old it is. The older the interaction, less weight

the system gives to it. In this way, it considers more

interactions that happened recently and items the

user is studying in the last few days. In Restriction,

the items will be evaluated in terms of Duration

Time and Typical Learning Time metadata of LOM,

matching it with user’s available time. Moreover, in

Novelty the items recommended will be evaluated in

terms of age, if this is indicated in the system

configuration, taking in account Date metadata of

LOM and current data.

6.2 A Scenario of Use

In this scenario, we take into account the Proposal

System, applied to a short online course about

Introduction to Algorithms. It’s a course that, in

average, is two months long, but students can join in

every time of the year and each one goes in his/her

own speed. There is no tests to evaluate students, but

there is plenty exercises to each user evaluate

himself/herself. In this course, there are more than

300 active users every day. Also, there are 150

learning objects available and characterized with

LOM metadata. It’s possible to tell that both

Collaborative Filtering and Content-Based

approaches can be applied in this scenario.

Suppose a user of this system is already in the

middle of the course. This user access twice a week

the course and spends, in average, one hour and a

half each time.

The Proposal System will show

recommendations to this user that do not exceed one

hour and a half, using time as Restriction to the list

of recommendations. In Content-based list, we have

items similar to the ones the user last accessed,

because of using the time as Decay. While in

Collaborative Filtering, even giving more weight to

last items, it’s possible to have surprising

recommendations, because it is calculated based on

items that other users with similar tastes liked. In

this scenario, there is no need to apply Novelty

category.

Figure 2 shows a prototype of an interface of the

Proposal System, where recommendations are

divided into Based on contents you accessed

(Content-Based approach) and Based on users

similar to you (Collaborative Filtering approach).

Note that, although Hybrid approach is used,

Content based and Collaborative Filtering are

calculated apart and shown apart in interface. Also,

it’s important to explain to the user where these

recommendations came from, so he/she might have

more trust on the items recommended.

Figure 2: Prototype of Recommendations.

7 CONCLUSION

In this paper is described the use of time context in

Recommendation Systems (RS) for learning. Time

context has demonstrated its impact in RS,

improving recommendation quality, but in learning

situations, few works were found that uses this

dimension.

In the present work, we take into account seven

categories of how time can be used in Recommender

System algorithms, based in our previous works.

The proposed scenarios illustrate the use of time in

learning situations and help to better explain each

category of time.

Based in this work, a Recommender System

architecture is proposed, that combines Content-

Based and Collaborative Filtering approaches in a

Hybrid algorithm. The system proposed uses time in

three different ways (Decay, Restriction and

Novelty) and takes advantage of some information

of LOM, a set of metadata used to represent

Learning Objects, for example, Date, Duration and

Typical Learning Time.

As future work, the proposed system should be

detailed, implemented and tested using a real

environment with active users, as well as others

combinations between Recommendations approach

and uses of time.

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

608

REFERENCES

Adomavicius, G.; Tuzhilin, A. 2005. Toward the next

generation of recommender systems: a survey of the

state-of-the-art and possible extensions. IEEE

transactions on knowledge and data engineering, vol.

17, no. 6.

Adomavicius, G; Tuzhilin, A. 2011. Context-aware

Recommender Systems. In: Ricci, F. (Org.); Rokach,

L. (Org.); Shapira, B. (Org.); Kantor, P. B. (Org.).

Recommender System: Handbook. Springer.

Beel, J.; Breitinger, C.; Langer, S.; Lommatzsch, A.; Gipp,

B. 2016. Towards reproducibility in recommender-

systems research. User Model User-Adap Inter, 26:69–

101.

Borba, E. J.; Gasparini, I.; Lichtnow, D. 2017. Time-

Aware Recommender Systems: A Systematic

Mapping. HCI International, Springer. (to appear).

Brusilovsky, P. Methods and Techniques of Adaptive

Hypermedia. 1998. Adaptive Hypertext and

Hypermedia, Kluwer Academic, Publishers, p. 1-43.

Burke, R. D. 2002. Hybrid Recommender Systems: survey

and experiments. UserModel UserAdapt Interact., v

.12, n.4, p. 331–370.

Campos, P. G.; Díez, F.; Cantador, I. 2014. Time-aware

recommender systems: a comprehensive survey and

analysis of existing evaluation protocols. User Model

User-Adapt Inter, vol. 24, pp. 67-119.

Chen, G.; Kotz, D. 2000. A Survey of Context-Aware

Mobile Computing Research. Technical report.

Dey, A.K. 2001. Understanding and using context.

Ubiquitous Comput, vol. 5, no. 1, pp. 4–7.

IEEE 2002. Draft Standard for Learning Object Metadata,

Available at:

http://grouper.ieee.org/groups/ltsc/wg12/20020612-

Final-LOM-Draft.html..

Ferfernig, A.; Friederich, G.; Jannach, D.; Zanker, M.

2011. Developing Constraint-based Recommenders.

In: Ricci, F. (Org.); Rokach, L. (Org.); Shapira, B.

(Org.); Kantor, P. B. (Org.). Recommender System:

Handbook. Springer.

Jannach, D.; Zanker, M.; Felfernig, A.; Friedrich, G. 2011.

Recommender Systems: An Introduction. New York,

USA: Cambridge University Press.

Koren, Y. 2009. The bellkor solution to the netflix grand

prize. Netflix prize documentation.

Koren, Y; Bell, R. 2011. Advances in Collaborative

Filtering. In: Ricci, F. (Org.); Rokach, L. (Org.);

Shapira, B. (Org.); Kantor, P. B. (Org.). Recommender

System: Handbook. Springer.

Lops, P.; Gemmis, M. de; Semeraro, G. 2011. Content-

Based Recommender System: State of the Art and

Trends. In: Ricci, F. (Org.); Rokach, L. (Org.);

Shapira, B. (Org.); Kantor, P. B. (Org.). Recommender

System: Handbook. Springer.

Merriam-Webster. Last access: 2016. Time. Available at:

http://www.merriam-webster.com/dictionary/time,

Petersen, K.; Feldt, R.; Mujtaba, S.; Mattsson, M. 2008.

Systematic Mapping Studies in Software Engineering.

12th International Conference on Evaluation and

Assessment in Software Engineering, vol. 17, no. 1.

Ricci, F.; Rokach, L.; Shapira, B. 2011. Introduction to

Recommender System Handbook. In: Ricci, F.;

Rokach, L.; Shapira, B.; Kantor, P. B. Recommender

Systems: Handbook. Springer.

Schmidt, A.; Beigl, M.; Gellersen, G. H. 1999. There is

More to Context than Location. Computers and

Graphics, vol. 23, no. 6, pp. 893-901.

Schilit, B.; Adams, N.; Want, R. 1994. Context-Aware

Computing Applications. Proc. First Workshop

Mobile Computing Systems and Applications

(WMCSA ’94), pp. 85-90.

Verbert, K.; Manouselis, N.; Ochoa, X.; Wolpers, M.;

Drachsler, H.; Bosnic, I.; Duval, E. 2012. Context-

Aware Recommender Systems for Learning: A Survey

and Future Challenges. IEEE Transactions on

Learning Technologies, vol. 5, n. 4.

Zimmermann, A.; Lorenz, A.; Oppermann, R. 2007. An

Operational Definition of Context. Proc. Sixth Int’l

and Interdisciplinary Conf. Modeling and Using

Context (CONTEXT ’07), pp. 558-571.

The Use of Time Dimension in Recommender Systems for Learning

609