Course Recommendation for University Environments

Boxuan MA

Kyushu University

ma.boxuan.61[email protected]ushu-

u.ac.jp

Yuta Taniguchi

Kyushu University

taniguchi@ait.kyushu-u.ac.jp

Shin’ichi Konomi

Kyushu University

konomi@acm.org

ABSTRACT

Recommending courses to students is a fundamental and also

challenging issue in the traditional university environment. Not

exactly like course recommendation in MOOCs, the selection and

recommendation for higher education is a non-trivial task as it

depends on many factors that students need to consider. Although

many studies on this topic have been proposed, most of them only

focus either on historical course enrollment data or on models of

predicting course outcomes to give recommendation results,

regardless of multiple reasons behind course selection behavior. To

address such a challenge, we first conduct a survey to show the

underlying characteristic of the course selection of university

students. According to the survey results, we propose a hybrid

course recommendation framework based on multiple features. Our

experimental result illustrates that our method outperforms other

approaches. Also, our framework is easier to interpret, scrutinize,

and explain than conventional black-box methods for course

recommendation.

Keywords

Educational Data Mining; Recommender Systems; University

Environments

1. INTRODUCTION

Course selection in university is a crucial and challenging problem

that students have to face. It is difficult to decide which courses

they should take because there are a large number of courses

opened each semester and students have to spend a lot of time

exploring those courses. Moreover, the decisions they make shape

their future in ways they may not be able to conceive in advance.

We collected a dataset during 2015 and 2018 from our university to

gain a better understanding of the elective course enrollment

patterns. Figure 1(a) presents the distribution of the enrolled course

number of students on the left and the distribution of the popularity

for each course of our university on the right. There are hundreds

of elective courses offered by the university while averagely

students only select a few of them to satisfy the requirements for

their degree program. Figure 1(b) shows the distribution of the

enrolled courses for each semester. We can also see that students

may take courses in the first two years mostly

(semester1~semester4), because they may potentially be busy with

an internship or finding jobs in the third and fourth year.

(a) #Course distribution.

(b) #Course distribution for each semester.

Figure 1. Distribution of courses.

From the discussion above, a safe conclusion could be drawn that

due to a large number of available but unfamiliar courses, course

selection is a critical activity for students.

With the increasing amount of available data about undergraduate

students and their enrollment information, data-driven methods

supporting decision making have gained importance to empower

student choices and scale advice to large cohorts [14]. Many

relevant studies on course recommendation focus on online

learning platforms such as MOOCs. Other studies on course

recommendation use datasets collected in physical university

environments, however, they rely on approaches that are similar to

the ones used in recommending MOOC courses without fully

considering the different reasons involved in course selection

process in physically-based university environments.

In fact, course recommendation for higher education can be more

“messy and unorganized” [1] as it depends on many factors that

students need to concern. Intuitively, the reasons behind course

selection are manifold. Likewise, students who enrolled in the same

course may have completely different orientations based on their

own reasons, which serves as different criteria for course selection

[34]. It inspires us to try to find more useful features for the

recommendation.

To make the point clear, a survey is conducted on 81 students in

our university to better understand student perceptions and attitudes

for their course selection process. [10] Figure 2 shows the main

underlying reasons for their course selection.

Boxuan Ma, Yuta Taniguchi and Shinichi Konomi "Course

Recommendation for University Environment" In: Proceedings of

The 13th International Conference on Educational Data Mining

(EDM 2020), Anna N. Rafferty, Jacob Whitehill, Violetta

Cavalli-Sforza, and Cristobal Romero (eds.) 2020, pp. 460 - 466

Proceedings of The 13th International Conference on Educational Data Mining (EDM 2020) 460

Figure 2. The distribution of main reasons for course selection.

(1) Interest

As Figure 2 shows, the overall most important factor was students’

interest and it is often taken as a main contributing factor to the

recommendation. However, students may not choose courses based

purely on their interest in the university environment. It is expected

that students will be more inclined to choose courses that do not

require too much effort or difficulty. For example, some students

would not enroll in a course which contains contents they are

interested in, they just choose the course that allows them to get

credits easily.

(2) High-Grade

Improperly selecting courses would seriously affect the students’

course achievements, which enforces students to drop out [12].

Getting relatively high grades for students is another factor

influences student’s choice especially for successful students.

Some students even prefer to choose what they perceived would be

an easier course for fear that a tougher course might lower their

GPA.

(3) Learning goal and career plan

It is natural to recommend courses that align with student’s learning

goals and career plans as students consider the usefulness of

courses as an important factor in their course selection process.

However, first-year students may lack learning goals and career

planning for the future, and the choice of courses is aimless. Also,

student interest and goal can change as they explore and discover

something meaningful on and off campus.

(4) Social Aspect

Social factor also plays a part in the course selection process. For

example, some students prefer to enroll in a course with their

friends or classmates together. Potts et.al [21] conclude that the risk

of social isolation is a problem in the learning process especially

for first-year students at university, who have difficulty navigating

their new academic and environment. Tinto [22] concludes that

participation in a collaborative learning group encourages student’s

attendance and class participation. Therefore, the classmates or

friends based social links could be important information in course

recommendation.

(5) Popularity

As shown in Figure 1(a), the long-tail distribution of course

popularity indicates that students are more motivated to choose

popular courses as their first choice. However, the popular courses

will be filled up quickly while others will not be selected by

students frequently.

In summary, all these discussions above indicate that there are

complex constraints and contexts that have to be considered

together to balance all those factors above, made more difficult by

the multiple objectives that students want to maximize and risks

they want to hedge against. For example, choosing challenging

courses of value while maintaining a high GPA [16]. This suggests

that recommendations that are aimed only at one or a few factors

are likely not enough to help the students.

To address these challenges which have not been well explored in

the research community, we propose our hybrid course

recommendation framework, which incorporates different criteria

in a modular way. Moreover, in our approach selection criteria can

further be prioritized by the student. We believe that weaving those

criteria could increase the usability of our recommendations

compared to previous work focusing only on one of the two. Also,

our framework is very efficient and easy to interpret.

2. RELATED WORK

2.1 Course selection

Some work has been done on analyzing the college students’ course

selection. Morsy and Karypis [23] investigated how the student’s

academic level when they take different courses, relate to their

graduation GPA and time to degree. This study suggests that

course recommendation approaches could use this information to

better assist students towards academic success, by graduating on-

time with high GPA. Also, understanding students’ reasons for

enrolling in a course provides key information for recommending

courses and improving students’ learning experiences [24-27].

Additionally, there is still a lack of study on the factors that

influence students’ course selection in university and how the

course selection would impact the students’ educational

achievement.

2.2 Personalized Course Recommendation

Various approaches have been used in applications for course

recommendation by learning from historical enrollment data [32,

33].

Content-based filtering approaches recommend a course to a

student by considering the content of the course and clustering

course and student into groups to gain similarity between them [2,3].

Collaborative filtering approaches recommend a course to a student

by investigating student's similarity with the student's historical

data in a system and predict the course that the student would be

interested in [4-6]. Association rules based on frequent patterns are

used to discover interesting relations that describe previous course

selections from students [8,9]. Recently, other methods including

sequence discovery and representation learning have been used in

this domain [11,19,20]. However, those systems often behave like

a “black box", i.e., recommendations are presented to the users, but

the rationale for selecting recommendations is often not explained

to end-users.

2.3 Grade Prediction

While some researchers have focused on between-course

enrollment data, others have focused on models of predicting

grades in future courses [13-6]. Based on what courses they

previously took and how well they performed in them, the predicted

grades give an estimation of how well students are prepared for

future courses, then recommending courses to students that will

help them to get relatively high grades [18,28,29,31].

However, these methods can be prone to recommending relatively

easier courses in which students usually get high grades [17]. In

addition, there are some students who like challenge difficult

courses if they are interested in or think it is helpful for their future

career, for those students, the grade prediction based

recommendations are not enough.

461 Proceedings of The 13th International Conference on Educational Data Mining (EDM 2020)

Despite the significant success of various course recommendations,

constraints on the number of student preferences in the university

environment resulting in inflexibility where a student’s

requirements do not align perfectly with those built into the system.

In contrast to the aforementioned approaches, our model combines

the concerns of performance and interest together. Also, it has the

benefit of allowing for a custom weighting of those components, as

well as the increased explanatory value of the model itself.

3. PROPOSED METHOD

We first give the definition of our recommendation problem in

Section 3.1. Then we propose our hybrid course recommendation

framework with three subsections introducing our Interest-based

Score, Timing-based Score, and Grade-based Score in detail.

Finally, those different scores are used in our course

recommendation algorithm introduced in Section 3.3.

Figure 3. Overview of the proposed course recommendation.

3.1 Problem Formulation

Like every classic recommendation task, there are two basic

elements user and item in our course recommendation task, where

a user represents a student and an item represents a course. We use

to denote a set of students and

to denote a set of courses. Each

# $!

has enrolled some courses denoted by

& "

and each

'( $

has its enrollment set denoted by

)

& !

. Let

denote a set of all

available semesters, and

to denote a specific semester. Generally,

there are 8 semesters for 4 academic years degree program. Let

denote a set of grades that student could get, and each

- $ ,

denote

a specific grade that student obtained for a course. Let

. /

#2(2- 2+

4#$!2( $ "

2- $ ,2+ $ *5

be the set of all enrollment

relations, which means student

enrolled in course

(

in semester

and got the final grade

Given enough students enrollment data

1!2"2.3

, our goal is

recommending courses to a specific student

which are not in

for next semester.

3.2 Framework

According to the result of the survey shown in Section 1, students

may concern different factors while they choose courses. Inspired

by that, we propose our hybrid course recommendation framework

that considers student interest, the timing of taking the course and

the predicted grade of the student together. Figure 3 shows the

overview of the proposed course recommendation.

For each pair of student and course

1#2(3

, we need to understand

how suitable the course is for the specific student. We use three

different aspects to calculate the

!(6781#2(3

for each pair of

student and course:

(i)

9:+878#+';<#8='!(678'19!3

, which is to measure how

interesting the course is for a specific student. (ii)

*>?>:-';<#8='!(678'1*!3

, which is to measure how suitable

students enroll in the course at a specific time (semester) since

different courses may have different suitable time periods. (iii)

,7<=8';<#8='!(678'1,!3

, which is to predict students’

performance for the course.

We propose our approaches to estimate

'*!

and

respectively. Then, they are fused by a student-specific weight

parameter as the

!(6781#2(2+3

. Once all of the

!(6781#2(2+3

have

been computed, the

courses with the highest score are selected.

3.2.1 Interest-based Score

Let

and

(

be a student and a course, respectively, the goal of

interest score estimation is to calculate

9!1#2(3

In our framework, we extract user interest from student historical

enrollment behaviors. Since each course of university belongs to a

category, let

"A*.

denote the set of all categories,

(<+8

to denote

a specific category, then

"A*. /0(<+8

2(<+8

2D2(<+8

EFGH

We think that there is a strong relationship between student interest

and course categories. For instance, a student frequently enrolls in

courses which belong to “Computer Science” may imply that the

student has an interest in this category or he may have personal

learning goal in this domain. Hence, it is appropriate to recommend

the student the courses such as “Python Programming” and “Data

science”.

For a student

, the idea is to count the number of courses that he

enrolled in and belongs to a category, i.e.,

IJ?1#2(<+83

. Then, all

of the values are normalized as the preference score from 0 to 1,

denoted as

K1#2(<+83

, which is defined as equation (1).

#2(<+8

IJ?

#2(<+8

L<M

)NOPQ$EFGH

IJ?

#2(<+8S

''''''''''''''''1U3

For a student

, the preference vector

, is obtained by the

preference score of each category, which is defined as equation (2).

/ 1K

#2(<+8

2D2K

#2(<+8

EFGH

3''''1X3

We can further use

to calculate the similarity between student

and other students. Let

and

be two students, the similarity

between

and

can be measured by the cosine similarity

measurement as shown below.

#>?

'2#

[

][

[

''''''''''''''''''''''''''''''1b3

For the convenience of computation, we use a matrix form

representation

c' /1[

e'[

'e'ggge'[

to denote the interest

of all students where

/'[

[

means the normalization of

[

. Then the similarity matrix

!>?

can be simply written as:

!>?' /'c

'`'c'''''''''''''''''''''''''''''''''''''''''1k3'

where

!>?

Y2Z

is the result of

#>?

'2#

Based on the similarity, we could estimate the user-based interest

score. For a student

and a course

(

, the Interest Score denoted as

9!1#2(3

, is defined as (5), where

%2l

indicates the set of top-k

similar students of

as neighbors, and

is an indicator function

whose value is 1 when

(' $'"#S

#'2(

`#>?

#'2#S

%Q$p

m2q

#>?

#'2#S

%Q$p

m2q

'''''''''''''''''''''''1r3

Furthermore, we try to utilize students’ major information together

with their similarity as equation (6).

#>?

#'2#S

/t#>?

#'2#

Uvt

#<?8?<w67

#'2#

''''' 1x3

Proceedings of The 13th International Conference on Educational Data Mining (EDM 2020) 462

Where

#<?8?<w67

#'2#

function equal to 1 if student

and

student

have the same major, otherwise, the function equal to 0.

(limited from 0 to 1) is used to control the weight between

similarity and major information. The underlying rationale is that

each major has its owner preference on courses enrolling, students

have the same major will generally make a similar choice in course

selection. Also, students in the same major are more likely to be

friends or classmates, which brings potential social link

information into the course recommendation. Then equation (5) can

be rewritten as below.

#'2(

`#>?

#'2#S

%Q$p

m2q

#>?

#'2#S

%Q$p

m2q

'''''''''''''''''''''''1y3

3.2.2 Timing and popularity based Score

Different courses may have different suitable time periods

(semesters). For example, in each department, courses can be taken

by students of different grades, e.g., freshman or sophomore.

Previous studies showed that the timing of courses has a strong

correlation with student graduation GPA and time to degree [23].

Based on that, we assume that the timing of courses is also

important for course selection. The suitable timing of courses will

help students for good grades and successful graduation in a timely

manner.

For each course

(

, we define the

*>?>:-';<#8='!(678#'1*!3

denoted as

*!1(2+3

, where

indicates a specific semester. In our

framework,

is considered from two aspects:

(1) Which semester is more suitable for taking this course? For a

specific course, we sum up the number of enrollments for every

semester and normalize all of the values. The result is denoted as

(2+

IJ?

(2+

L<M

IJ?

(2+

''''''''''''''''''''''

where

IJ?

(2+

represents the number of enrollments of course

(

in semester

, and

indicates the set of all time periods, i.e., 8

semesters for 4 academic years degree program.

(2) Which courses are popular now? For a specific semester, we

sum up the number of enrollments for every course and normalize

all of the values. The result is denoted as

{

(2+

{

(2+

IJ?

(2+

L<M

)Q$E

IJ?

(S2+

'''''''''''''''''''''''1|3

where

indicates the set of all courses.

(2+

and

{

(2+

are

then fused by the harmonic mean since we want both of the two

values are relatively high. The final

'*>?>:-';<#8='!(6782

*!1(2+3

can be defined as:

(2+

/ X`

(2+

`'*

{

(2+

{

(2+

'''''''''''''''''''1U}3

Therefore, we can use

(2+

to ensure that the semester

suitable for taking the course

(

and the course

(

is suitable for

taking in the semester

3.2.3 Grade-based Score

Improperly selecting courses would seriously affect the students’

course achievements, which may decrease their GPA even enforce

students to drop out. Accurately predicting students’ grades in

future courses has attracted much attention as it can help identify

at-risk students early [30].

We use the grade prediction method called cross-user-domain

collaborative filtering proposed by Ling et al. [12]. For predicting

the score of each course

'('$'"

for each student

#' $!'

, a small set

of senior students who have already enrolled on course

(

and have

the most similar previous score distribution to student

will be

discovered by means of Pearson correlation coefficient. The

underlying rationale is that students with similar scores in the

previous courses will generally obtain similar scores in the

subsequent courses.

Let

denote the set of senior students who have already enrolled

on course

(

. For any senior student

'$'!

, the following Pearson

correlation coefficient is used to measure the course score

similarity between student

and the senior student

#>?

#'2#

o 1

Y$E

•

o 1

Y$E

€€

•

Y$E

(11)

where

denotes the courses that are enrolled by both students

and

denote the grade of course

by students

and

respectively.

and

denote the average grade of courses

enrolled by students

and

, respectively. Accordingly, the grade

of the course

(

by student

can be predicted as follows.

)

3`#>?

#'2#

m2q

#>?

#'2#

m2q

'''''''''''''''''''1UX3

where

%2l

indicates the set of top-k similar senior students of

. It

should be noticed that students often achieve inconsistent grades in

the various courses they take, and different students may have

varying grades deviations, i.e. the grades deviation compared with

the average grades among all students. Similarly, different courses

may have varying grades deviations, i.e. the score deviation

compared with the average score among all courses. In order to deal

with those variations. We use the grade deviation of student

and

the grade deviation of course

(

to predict student grades.

Accordingly, equation (12) could be rewritten as below.

)

3`#>?

#'2#

m2q

#>?

#'2#

m2q

''''''1Ub3

where

;

'/'•u;

'u;

)

denotes the baseline estimate for

with

•

being the overall mean grade of all courses enrolled by all

students,

;

'/'-

'v'•

being the grade deviation of student

and

;

)

'/'-

)

'v'•

being the grade deviation of course

(

, where

the overall mean grade of student

and

)

is the overall mean

grade of course

(

Finally, we could use the grades that students are expected to obtain

in future courses to boost the performance of our recommendation.

The final

',7<=8';<#8='!(678#2

,!1 #2(3

can be defined as

normalized values of grades.

#2(

L<M

)

%)Q

''''''''''''''''''''''''''1Uk3

The total score of student and course pair

'#(678'1#2(3'

can be

written as:

!(678 /‚`9!

#2(

uƒ`*!

#2+

u„`,!

#2(

'''''''1Ur3

Where

‚

'ƒ

„

are parameters to control the proportion of weights

from different sources. By taking those scores into account

simultaneously, a course that the student interested in, and suitable

for him to take to get a high grade could be ranked higher than other

courses. Also, student could control the weighting of those

components to have a better understanding of the data and decision-

making.

463 Proceedings of The 13th International Conference on Educational Data Mining (EDM 2020)

3.3 Course Recommendation Algorithm

The whole framework can be written as Algorithm 1. Student set

course set

, enrollment set

.1#2(2-2+3

are input and output is a

list of recommendations

…#

for student’s next semester that

includes up to

recommendations per student.

Algorithm 1: Generating a list of course recommendations for

student

Input ：

Student set !, course set ", enrollment set .1#2(2-2+3;

Output ：

Recommendation results for each student …#.

Calculate student interest'†

for each #' $'! by equation 1;

Calculate student interest similarity by equation 4;

Calculate student grade similarity by equation 11;

Calculate student deviation of each student;

Calculate course grade deviation of each course;

foreach #' $'! do

Calculate user interest-based score'9!1#2(3 by equation 7;

Calculate timing-based score'*!1(2+3 by equation. 10;

Calculate user grade-based score',!1#2(3 by equation 13

and equation 14;

Calculate final score '!(6 78'1#2(2+3 by equation 15;

Let …# be the sorted list of " ordered by its !(678'1#2(2+3 in

descending order.

endfor

4. EVALUATION

In this section, we conducted a series of experiments to evaluate the

effectiveness of our proposed method. We first describe the dataset

and experimental settings. Next, the evaluation methodology and

metrics are introduced in detail. Finally, the results are shown in

Section 4.4.

4.1 Dataset

This work focuses on undergraduate students in a traditional

educational institution. We used a dataset from our university that

spans for 5 years. The dataset consisted of per-semester course

enrollment information of 2,366 students from 12 departments,

with a total of 38,968 pseudonymized enrollment records from

2014 through 2018. Each row of the course enrollment data

contained semester and department information, an anonymous

student ID and course information included course name, instructor

and course category.

4.2 Experiment Settings

4.2.1 Data selection

The most natural approach to evaluate the model is to split the data

by semesters. As shown in Figure 1(b), most of the undergraduate

students may take courses in the first two years. Therefore, for

students who enrolled in 2015, the semester of Spring 2015 was

used for training, the subsequent semesters of Fall 2015, Spring

2016 and Fall 2016 are regarded as the testing semesters, each of

which is tested separately. The results are evaluated by comparing

the predicted courses and the ground-truth courses he/she has

enrolled in.

4.2.2 Comparison

We name our methods as Hybrid Course Recommendation (HCR).

We compare our method with two group popularity approaches [14]

and Random recommendation (Random). The two group popularity

approaches including the department level (Grp-Pop-1), which

recommend the most popular courses in the major, and the

academic level (Grp-Pop-2), which recommend the most popular

courses on the major and the academic level of the student

(“freshmen”, “sophomores”, “juniors”, and “seniors”).

4.3 Evaluation Metrics

Like previous work [11,14,15,20], we used Recall@ns and

Coverage as the evaluation metric for the performance.

Coverage is measured based on the percentage of courses that have

been recommended at least once to students, which describes the

ability of a recommendation system to explore the long-tail item.

Recall@ns is the percentage of actually enrolled courses of

semester

that were contained in the recommendation list, where

ns is the number of courses that the student took in the target

semester. The reported metrics are averaged out across all students.

Since our proposed course recommendation method considers both

student interest and the grade he/she may obtain, we cannot only

use the Recall metric, and instead, we use a variation of it. For the

list of the courses

…

that recommended to a student

, Let

is the

set of courses in the test set of

is the set of courses which

student is expected to get the grade equal to or higher than his/her

average previous grade. We use the ratio of

4…

‡*

‡A

and

‡A

to measures the fraction of the actual well performed

courses that are retrieved.

4.4 Results

4.4.1 Interest-based Score

In collaborative filtering strategy, taking how many similar

students as neighbors is an important problem which is sensitive to

the quality of the result. We investigate the performance of our

interest model with different neighbor numbers. As shown in

Figure 4, the performance of the model increases with the increase

of neighbor number at first then decreases. According to the

observation above, we pick a practical value 40 as the value of the

neighbor number parameter in our follow-up experiments.

We also investigate the performance of our interest model with

different weights between similarity and major information. As

shown in Figure 5, we can observe that the performance of the

model increases with the decrease of

in terms of Recall. The

reason is that the model considers not only the similarity but also

the major information. That is, each major has its own preference

for courses enrolling, major information will improve the

performance of the algorithm.

However, 100% recall could be bad because the system just

recommends what students do anyway. We noticed that the

Coverage also decreases with the decrease of

. The model seems

benefit from the major information while scarifying the diversity of

results. Recommendations for courses at other departments

sometimes are useful to mine more long-tail student interest while

students usually ignored that these courses existed or that their

content matched their interests. To achieve the best performance

of recommendation, we need to make a trade-off. According to the

observation above, we set

as 0.2 in our follow-up experiments.

Proceedings of The 13th International Conference on Educational Data Mining (EDM 2020) 464

Figure 4. Performance of different neighbor numbers.

Figure 5. Evaluation of major information.

4.4.2 Influence of different factors

To illustrate the influence of different factors, we set each

parameter

‚2ƒ2„

from 0 to 1 with a step size 0.05 to find the

optimal combination.

Table 2. Performance of different parameters

Model

1‚2ƒ2„3=

(1,0,0)

1‚2ƒ2„3

= (0,1,0)

1‚2ƒ2„3

= (0,0,1)

1‚2ƒ2„3*

Recall@ns

0.380

0.412

0.341

0.434

Recall(a)@ns

0.311

0.315

0.332

0.322

Coverage

0.534

0.212

0.356

0.516

As shown in Table 2, the interest score and timing score has a good

explanatory value for the recommendation than others. Also, the

suitable timing of taking a course will help students to get a good

grade. A recommendation only based on the grade has a good

performance for recommending high grade courses. However, the

results cannot help all the students. We reached the best Recall@ns

with (

‚

=0.4,

=0.45 and

„

=0.15)*.

The results indicate that recommendations that are aimed only at

one factor are likely not to be satisfied by every student. As we

discussed before, different students may have completely different

orientations based on their own reasons, which serves as different

criteria such as their preferences, interests, needs, performance,

etc. Such a hybrid system could provide explanations and user

controls for different categories of target students to support the

interpretation of the data and decision-making.

Table 3. Evaluation of course recommendation

Semester

Model

Recall@ns

Recall(a)@ns

Coverage

Fall 2015

Random

0.048

0.036

Fall 2015

Grp-Pop-1

0.374

0.306

0.272

Fall 2015

Grp-Pop-2

0.452

0.342

Fall 2015

HCR

0.472

0.393

0.578

Spring 2016

Random

0.025

0.020

Spring 2016

Grp-Pop-1

0.325

0.201

0.305

Spring 2016

Grp-Pop-2

0.423

0.372

0.237

Spring 2016

HCR

0.431

0.402

0.342

Fall 2016

Random

0.002

Fall 2016

Grp-Pop-1

0.326

0.243

0.213

Fall 2016

Grp-Pop-2

0.441

0.387

0.250

Fall 2016

HCR

0.463

0.392

0.559

4.4.3 Comparison result

We analyze the performance of different algorithms. The results in

Table 3 show that our framework performs well when compared

with other methods.

As the results show, both of the Recall and Recall(a) of Random

recommendation strategies are very low since there are a large

number of courses, but each student only averagely chooses a few

courses per semester. Hence, it is difficult to recommend the right

course. Popularity approaches are having considerably satisfactory

performance in Recall since popular courses which are taken by

students frequently usually attract most of students. However, Grp-

Pop-1 and Grp-Pop-2 do not consider student preference, it is also

difficult to mine more long-tail student interest as the Coverage is

low. In addition, Grp-Pop-1 and Grp-Pop-2 are not good in Recall(a)

since they only consider the popular courses, ignore the

performance the student is expected to get in the recommended

courses.

5. CONCLUSION

This research aims to recommend suitable courses for learners and

study how to design a personalized course recommendation in the

university environments. In this paper, we propose a hybrid course

recommendation framework that considers student interest, the

timing and popularity of courses, and predicted performance of

students, simultaneously. Experiments are conducted to confirm

the effectiveness of the proposed approach. The results show that

the proposed hybrid course recommendation approach performed

well compared to other methods. Also, the model itself is flexible

in the sense that one can easily adjust or extend it by changing the

recommendation formula and incorporate more information.

6. REFERENCES

[1] E. Babad and A. Tayeb. 2003. Experimental analysis of students’

course selection. British Journal of Educational Psychology,

73(3):373–393.

[2] Piao G, Breslin JG. 2016. Analyzing MOOC Entries of

Professionals on LinkedIn for User Modeling and Personalized

465 Proceedings of The 13th International Conference on Educational Data Mining (EDM 2020)

MOOC Recommendations. In Proceedings of the 2016 Conference

on User Modeling Adaptation and Personalization. ACM, 291–

292.

[3] Apaza RG, Cervantes EV, Quispe LC, Luna JO. 2014. Online

Courses Recommendation based on LDA. In SIMBig. 42–48.

[4] BYDŽOVSKÁ, Hana. 2016. Course Enrollment Recommender

System. Proceedings of the 9th International Conference on

Educational Data Mining. Raleigh, NC, USA: International

Educational Data Mining Society. 312-317.

[5] Khorasani ES, Zhenge Z, Champaign J. 2016. A Markov Chain

Collaborative Filtering Model for Course Enrollment

Recommendations: 2016 IEEE International Conference on Big

Data (Big Data). 3484 – 3490.

[6] Jing, X., Tang,J. 2017. Guess you like: course recommendation in

Moocs. In: Proceedings of the International Conference on Web

Intelligence, ACM, 783–789.

[7] Bhumichitr K, Channarukul S, Saejiem N, Jiamthapthaksin R,

Nongpong K. 2017. Recommender Systems for university elective

course recommendation. 14th International Joint Conference on

Computer Science and Software Engineering (JCSSE) (2017): 1-5.

[8] Aher SB, Lobo LM. 2013. Combination of machine learning

algorithms for recommendation of courses in E-Learning system

based on historical data. Knowledge-Based Systems (2013), 1–14.

[9] Bendakir N, Aïmeur E. 2006. Using association rules for course

recommendation. Proceedings of the AAAI Workshop on

Educational Data Mining. Vo l. 3 .

[10] Ma, B.X., Lu, M., Taniguchi, Y. and Konomi, S. 2020. Exploring

the Design Space for Explainable Course Recommendation

Systems in University Environments. In Companion Proceedings

of the 10th International Conference on Learning Analytics &

Knowledge.

[11] Polyzou A, Nikolakopoulos AN, Karypis G. 2019. Scholars Wa lk :

A Markov Chain Framework for Course Recommendation.

International Educational Data Mining Society. (2019 Jul).

[12] Huang L, Wang CD, Chao HY, Lai JH, Philip SY. A score

prediction approach for optional course recommendation via cross-

user-domain collaborative filtering. IEEE Access. (2019 Feb

7);7:19550-63.

[13] Sweeney, M., Lester, J., Rangwala, H., and Johri, A. 2016. Next-

term student performance prediction: A recommender systems

approach. Journal of Educational Data Mining 8, 1, 22–51.

[14] Elbadrawy, A. and Karypis, G. 2016. Domain-aware grade

prediction and top-n course recommendation. In Proceedings of the

10th ACM Conference on Recommender Systems. ACM, 183–190.

[15] Morsy, S. and Karypis, G. 2017. Cumulative knowledge-based

regression models for next- term grade prediction. In Proceedings

of the 2017 SIAM International Conference on Data Mining.

SIAM, 552–560.

[16] Jiang W, Pardos ZA and Wei Q. 2019. Goal-based course

recommendation. In Proceedings of the 9th International

Conference on Learning Analytics & Knowledge. ACM, 36–45.

[17] Morsy S, Karypis G. 2019. Will this Course Increase or Decrease

Your GPA? Towards Grade-aware Course Recommendation[J].

arXiv preprint arXiv:1904.11798.

[18] Okubo F, Yamashita T, Shimada A, Ogata H. 2017. A neural

network approach for students’ performance prediction. In

Proceedings of the Seventh International Learning Analytics &

Knowledge Conference (2017 Mar 13). 598-599.

[19] Pardos ZA, Jiang W. 2019. Combating the Filter Bubble:

Designing for Serendipity in a University Course Recommendation

System. arXiv preprint arXiv:1907.01591 .2019.

[20] Pardos ZA, Fan Z, Jiang W. 2019. Connectionist recommendation

in the wild: on the utility and scrutability of neural networks for

personalized course guidance. User Modeling and User-Adapted

Interaction. (2019 Apr 1). 487-525.

[21] Potts BA, Khosravi H, Reidsema C, Bakharia A, Belonogoff M,

Fleming M. 2018. Reciprocal peer recommendation for learning

purposes. In Proceedings of the 8th International Conference on

Learning Analytics and Knowledge (2018 Mar 7). 226-235.

[22] Tinto V. 1997. Classrooms as communities: Exploring the

educational character of student persistence. The Journal of higher

education. (1997 Nov 1).68(6):599-623.

[23] Morsy, S. and Karypis, G. 2019. A study on curriculum planning

and its relationship with graduation gpa and time to degree. In

Proceedings of the 9th International Conference on Learning

Analytics & Knowledge. ACM, 26–35.

[24] Kinnunen P, Malmi L. 2006. Why students drop out CS1 course?

In Proceedings of the second international workshop on

Computing education research (2006 Sep 9). 97-108.

[25] Crues R, Bosch N, Anderson CJ, Perry M, Bhat S, Shaik N. 2018.

Who They Are and What They Want: Understanding the Reasons

for MOOC Enrollment. International Educational Data Mining

Society. (2018 Jul 16).

[26] Kardan, H. Sadeghi, S. S. Ghidary, and M. R. F. Sani. 2013.

Prediction of student course selection in online higher education

institutes using neural network. Computers & Education. vol. 65.

1–11.

[27] Feng W, Tang J, Liu T X. 2019. Understanding dropouts in

MOOCs. Proceedings of the AAAI Conference on Artificial

Intelligence. 2019. 33: 517-524.

[28] Elbadrawy, A., Studham, R. S., and Karypis, G. 2015.

Collaborative multi-regression models for predicting students’

performance in course activities. In Proceedings of the 5th

International Learning Analytics and Knowledge Conference.

[29]Hu, Q. and Rangwala, H. 2018. Course-specific markovian models

for grade prediction. In Pacific-Asia Conference on Knowledge

Discovery and Data Mining. Springer, 29–41.

[30] Hu, Qian and Huzefa Rangwala. 2019. Reliable Deep Grade

Prediction with Uncertainty Estimation. In Proceedings of the 9th

International Conference on Learning Analytics & Knowledge.

ACM,76–85.

[31] Backenköhler M, Scherzinger F, Singla A, Wolf V. 2018. Data-

Driven Approach towards a Personalized Curriculum.

International Educational Data Mining Society.

[32] Morsomme R, Alferez SV. 2019. Content-based course

recommender system for liberal arts education. In Proceedings of

the 12th International Conference on Educational Data Mining

2019. Vol . 7 4 8 , p . 7 5 3 .

[33] Bydžovská H. 2016. Course Enrollment Recommender System.

International Educational Data Mining Society.

[34] Esteban A, Zafra A, Romero C. 2018. A Hybrid Multi-Criteria

Approach Using a Genetic Algorithm for Recommending Courses

to University Students. International Educational Data Mining

Society. (2018 Jul).

Proceedings of The 13th International Conference on Educational Data Mining (EDM 2020) 466