Top contenders in the 2015 KDD cup include the team from DataRobot comprising Owen Zhang, #1 Ranked Kaggler and top Kagglers Xavier Contort and Sergey Yurgenson. Get an in-depth look as Xavier describes their approach. DataRobot allowed the team to focus on feature engineering by automating model training, hyperparameter tuning, and model blending - thus giving the team a firm advantage.
2. ● XuetangX, a Chinese MOOC learning platform initiated
by Tsinghua University,
● launched online on Oct 10th, 2013.
● more than 100 Chinese courses and over 260
international courses
● high dropout rate
The competition host
3. ● challenge: predict whether a user will drop a course
within next 10 days based on his or her prior activities.
● data:
○ enrollment_train (120K rows) / enrollment_test (80K rows):
■ Columns: enrollment_id, username, course_id
○ log_train / log_test
■ Columns: enrollment_id, time, source, event, object
○ object
■ Columns: course_id, module_id, category, children, start
○ truth_train
■ Columns: enrollment_id, dropped_out
Problem to solve
6. How we worked as a Team
● worked separately on feature engineering. 90% of
our time was spent here.
● delegated Modeling part to DataRobot to:
○ find best algorithm (with XGboost as a winner!)
○ model text features
○ tune hyperparameters
○ experiment different feature sets and blend 8 XGBoost
using different sets
○ communicate results
7. Feature engineering techniques used
● counts
● time statistics (min, mean, max, diff)
● entropy
● sequences treated as text on which we ran
○ SVD on 3grams
○ DataRobot text mining solution
● 20 first components of SVD on user x object
NB: removed duplicated log info and used training + test
sets to build most features
9. Key course features
● course_id
● first log time
● enrollment counts
● unique log counts
● mean time interval
10. Key enrollment count features
● log counts
● unique log counts
● ratio between unique log counts over log counts
● unique log counts by event (nagivate, access,
problem, video, page_close, discussion, wiki)
● unique log counts before end of course (5 days, 10
days and 30 days before)
● sequence number of enrollment in that course
11. Key enrollment time stats
● log time stats (min, mean, max)
● gap between first and last log of enrollment
● gap between enrollment first log and course first log
● gap between enrollment last log and course last logs
● difference between mean log time and mid point
between first and last log
● log interval stats (mean, 90, 99 and 100 quantiles)
12. Enrollment entropy features
enrollment entropy over
● days
● weekdays
● fraction (4) of weekdays
● hours of the day
● hours of the day for the last 1/3/7 days before last
logs
● object (when event == problem)
● chapter ids
13. Example of entropy feature
- log(weekday_log_count / enrollment_log_count) *
weekday_log_count / enrollment_log_count
Sum => weekday_entropy[enrollment_id==1]
1.589988
14. Enrollment sequence features
● for each enrollment_id, built sequences of
○ weekdays
○ objects
■ all objects / 'problem' and 'video' objects only
○ events
● treated sequences as 4 text variables. Ran for each
○ svd on 3 grams => first 10 components
○ DataRobot stacked predictions from logistic regr.
& Nystroem SVM on (tuned) n-grams
18. Key user count features and time
stats
● enrollment count
● binary indicator whether user signed up for each of
the 38 courses
● unique log count
● mean log time interval
● sequence number of enrollment for that user
20. User sequence features
● for each user, built sequences of
○ weekdays
○ chapter_ids
○ events
● treated them as 3 text variables. Ran
○ SVD on 3 grams => first 10 components
○ DataRobot stacked predictions from logistic regr.
+ Nystroem SVM on (tuned) n-grams
21. How we got to the TOP3
● entropy features mentioned before
● exploited info in
○ log count in the 5 / 10 / 20 days after end of course
○ log counts by event, sign_up counts and day entropy in the next
10 days after end of course
○ time to sign up for new course
○ time until the next log for same user
added ~0.001 to AUC (vs
less powerful features)
added ~0.002 to AUC