Improving retention: predicting at-risk students by analysing clicking behaviour in a virtual learning environment

Annika Wolff and Zdenek Zdrahal10th December 2013

Student retention

• Struggling students don’t always ask for help – drop-out of module or fail and then don’t progress further

• When timely help is offered, this can make the difference between success and failure.

• It can be hard to know who’s in trouble and where to direct resources

Open University context

students

tutors

Distance learning:• Content through VLE• Contact mediated

through VLE – how to tell if students are struggling?

Solution: develop predictive models from student data

Data sources and data sets

VLE Assessment Demographic

Learning contentForumsQuizzes….

Ongoing assessmentsFinal exam

AgeGenderPrevious study…..

Typical VLE clicks

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 470

500

1000

1500

2000

2500

3000

Students Tutors

VLE activity (prior TMA1)• No VLE activity … 317 students• 1-20 clicks ……….. 609 students• 21-80 clicks ……… 943 students• 81-150 clicks ……. 621 students• 151-300 clicks …. 803 students• 301-600 clicks …. 516 students• > 600 clicks ……… 355

students

Problem specification

• Given:– Demographic data at the Start (may include information about

student’s previous modules studied at the OU and his/her objectives)– Assessments (TMAs) as they are available during the module– VLE activities between TMAs– Conditions student must satisfy to pass the module

• Goal: – Identify students at risk of failing the module as early as possible so

that OU intervention is meaningful.

Comments on problem specification

• OU intervention is meaningful if the cost of the intervention is lower than the expected gain from retaining the student.

• Modelling the problem:

We are here

Comments on problem specification

• OU intervention is meaningful if the cost of intervention is lower than the expected gain from retaining the student.

• Modelling the problem:

We are here

History we know

Comments on problem specification

• OU intervention is meaningful if the cost of intervention is lower than the expected gain from retaining the student.

• Modelling the problem:

We are here

History we know Future we can estimate

Comments on problem specification

• OU intervention is meaningful if the cost of intervention is lower than the expected gain from retaining the student.

• Modelling the problem:

We are here

History we know Future we can estimate

… and we can influence!

Comments on problem specification

• OU intervention is meaningful if the cost of intervention is lower than the expected gain from retaining the student.

• Modelling the problem:

We are here

History we know Future we can estimate

How can we estimate the future? … Based on student’s history and properties of upcoming parts of the module known from previous presentations.

Prediction at TMA1

– Why? TMA1 is a good predictor of success or failure

– It is enough time to intervene

We are hereHistory we know Future we can affect

Building a classifier

Training instances

New instances

FAIL

PASS

PassFail

Pass

Fail

FailPass

Assessment 1 score?

>40% <40%

Decision Tree – first results (no demographics)

Performance drop (VLE+TMA)

Final outcome

Naïve Bayes network

Sex

Education

N/C

VLE

TMA1

• Education:– No formal qualif.– Lower than A level– A level– HE qualif.– Postgraduate qualif.

• VLE:– No engagement– 1-20 clicks– 21-100 clicks– 101 – 800 clicks

• N/C:– New student– Continuing student

• Sex:– Female– Male

Goal:Calculate probability of failing at TMA1 • either by not submitting TMA1,• or by submitting with score < 40.

Predicting final result from TMA1

TMA1 Final resultTMA7TMA2

Pass/Distinction

Fail

TMA1 >=40

TMA1 <40

Prior probabilities: P(Success) = 0.807, P(Fail) = 0.193

Posteriori probabilities: P(Success|TMA1) = 0.858, P(Fail|TMA1) = 0.142P(Success|~TMA1) = 0.093, P(Fail|~TMA1) = 0.907

Bayes minimum error classifierIf student fails in TMA1 he/she is likely to fail the final result

VLE

P(Fail|TMA1-score), P(Pass/Dist|TMA1-score)

0-39 40-59 60-69 70-79 80-1000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

FailPass/Dist

TMA1

Predicting final result from TMA1

Sex

Education

N/C

VLE

TMA1 Final resultTMA7TMA2

Pass/Distinction

Fail

TMA1 >=40

TMA1 <40

Prior probabilities: P(Success) = 0.807, P(Fail) = 0.193

Posteriori probabilities: P(Success|TMA1) = 0.858, P(Fail|TMA1) = 0.142P(Success|~TMA1) = 0.093, P(Fail|~TMA1) = 0.907

Bayes minimum error classifierIf student fails in TMA1 he/she is likely to fail the final result

VLE

Demo Case 1• Demographic data

– Student fits certain demographic profile of gender, educational background etc.

Sex

Education

N/CTMA1

Without VLE:Probability of failing at TMA1 = 18.5%

Sex

Education

N/C

VLE

TMA1

Clicks Probability Nr of students0 64% 4

1-20 44% 3

21-100 26% 5

101-800 6.3% 14

With VLE:

Demo Case 2• Demographic data

– Different demographic profile to previous slide

Sex

Education

N/CTMA1

Without VLE:Probability of failing at TMA1 = 7.7%

Sex

Education

N/C

VLE

TMA1

Clicks Probability Nr of students0 39% 35

1-20 22% 74

21-100 11.2% 178

101-800 2.4% 461

With VLE:

TMA1? … it might be too late!

Can we predict TMA1 from VLE activities 1 week before the TMA1 deadline? How about 2, 3, … weeks?

We are here

History Future we can affect

predicted to fail

has not engaged with VLE

average score < 40

Dashboard and Chart

at least one TMA below 40

Has not submitted TMA5

has not engaged with VLEaverage score = 81.71 !!!

However

Dashboard – new design

Conclusions

• In a distance learning context, the VLE data provides a valuable source of data for prediction

• Prediction improves as a module progresses, but this is too late!

• We need to optimise methods for early prediction