MOOC Drop-out Prediction

Problem Statement

Students' high dropout rate on MOOC platforms has been heavily criticized, and predicting their likelihood of dropout would be useful for maintaining and encouraging students' learning activities

Code Structure

preprocessing_eventcount.py

- Pre-processes  1st Feature Vector : Events Count 
- Includes load data and produces Feature Table in CSV.

preprocessing_daycount.py

- Pre-processes  2nd Feature Vector : Day Count 
- Includes load data and produces Feature Table in CSV.

preprocessing_combined.py

- Pre-processes & Combines both Feature Vectors 
- Reads directly from CSV’s produced in a. and b.
- Produces final Feature Table in CSV.

Rf.py

- Classifier #1 -> Outputs a .txt file and ROC curves

mlp.py

- Classifier #2 -> Outputs a .txt file and ROC curves

xgb.py

- Classifier #3 -> Outputs a .txt file and ROC curves.

hard_voting.py

- Calls Ensemble method

dataset (folder)

- Contains all the data provided by XueTang.
- Contains a "feature" folder that stores the CSV's produced in a, b, and c.

outputs (folder)

- Contains all .txt files produced in d, e, and f.
- Contains the ROC curves that are to be manually saved as PNG's.		      	

Understanding the Data

• Date.csv : Gives us more information about the timespan of each course
• Object.csv : Gives us more information about each module in a course
• Enrollment_(train/test).csv : User enrollment records
• Log_(train/test).csv : Behavior Records
• True_train.csv : Ground Truth about a dropout

Data-Based Hypotheses

Any student who completes a course would have more counts for each event i.e. watching more course videos, solving problems, participating in discussions etc.

Besides events, there is also the information of whether the student accessed the course in a server or a browser. This information was useful in terms of how a browser is more frequently used to watch videos and other interactive events compared to a server. Thus, I hypothesized that a student who spends more time in a browser would have a lower chance of dropping out of a course as well.

Finally, the greater the number of active days (days where the student had events on a particular enrollment), the smaller the chance of a dropout, and vice versa.

Pre-Processing & Feature Engineering

Feature Vector 1: Event Counts

• In order to build this feature set, the data was grouped by the enrollmend_id. And then for each enrollment_id, each of the event-category counts were added.

Index	Variable
1	# of problems solved
2	# of videos watched
3	# of access
4	# of wiki
5	# of discussions
6	# of navigations
7	# of page close
8	# of server connections
9	# of browser connections

Feature Vector 2: Time Series & Day Counts

Handling Corrupt Data

• Before engineering the next feature vector, I noticed that the Enrollment Log showed that students accessed some modules before they were even posted i.e. enrollment_log time stamp was before the start-date of a course module in object.csv.
• In order to solve this, I deleted those log instances so as to ensure that it did not affect the calculation of active days. It turned out that roughly 2% of the Enrollment ID’s (1570 out of 72395) had such corrupt log instances.

Building the Vector

Index	Variable
1	last log time - first log time
2	# of effective study days
3	# events in start time-period of course
4	# events in middle time-period of course
5	# events in end time-period of course

• The dataset was grouped by enrollment_id, then converted the time element into Dates after which it was easy to find the first and last date of the event for a specific user.
• The courses' duration was divided into 3 equal time-spans: beginning, middle, and end. This information was useful to determine key user behaviors (eg. users that had some log record during the end period of a course were more likely to complete the course without a dropout, etc.)

Both feature vectors proved to give useful insight into the users' behaviors of dropout. As a result, the two vectors were combined to form a final feature vector with 14 Variables.

Classifers

Results

Ensemble: Hard Voting

Conclusion