wzfhaha / dropout_prediction Goto Github PK

Hi @wzfhaha , as titled, KDDCup2015 ended, wondering may the gt for testset be released so the KDDCup2015 dataset is fully accessible for the public. (Need it as 1. it's smaller->easier to start with 2. used in the paper, so others may need the test gt when trying to report KDDCup2015 dataset results for fair comparison), would be best if the dataset could also be archived on moocdata.cn website.

Regards,
Zan

Hi @wzfhaha, thx for sharing the code, I'm trying to re-produce the result and do some experiments to compare with https://github.com/srijankr/jodie for dropout prediction.

Could u pls tell if I hope to work on KDDCup2015 dataset which also got reported in the paper, the raw procedure for feature-extraction and pre-processing would be similar to current implementation which only works for AAAI2019 version dataset?

Hi @wzfhaha , to make things clear, may you pls elaborate on the details of definition of dropout for KDDCup2015 and AAAI2019 dataset separately:

1. For KDDCup2015 dataset, it's claimed if user do not have records for 10 days after the last day of class, he/she is considered as dropout, it's bit wried as if someone finished the course, he/she may not visit it in the next 10 days but may still get the certificate(taken as completion rather than dropout), and one-month seems shorter than the whole semester, we u explain it(sorry I started taking courses on XuetangX since Sep 2014, hard to understand what happened before that)?

2. For AAAI2019 dataset, students do have have activity record during the prediction period are taken as dropout from the course, that is reasonable, but how do you decide which period to be taken as prediction preriod, especilly for SPM courses?

3. Are these non-dropouts all taken as completion whether that get the certificate or not? Or dropout is sometimes defined by whether the student get the certificate or not?

4. The dropout rate is below 95% according to the statistics listed in the paper, are these users randomly selected from all users of XuetangX weighted by user activity frequency(so more active users got selected)?

Thx for making the dataset,paper and code publically avaiable!

Regards,
Zan

Hello @wzfhaha ,
I am trying to reproduce the result you got. But I could not find the code for generating the clusters.
Can you kindly provide the code for preparing label_5_10time.npy and user_dict?