1
$\begingroup$

I've recently been tasked with developing a prediction model for the scientific computing cluster at my university, with thousands of users.

User behaviour typically looks like this:

enter image description here

Most users are inactive. Few exhibit minor fluctuations, few move up or down and then plateau, but mostly irregular.

I know there's no chance of e.g. correctly predicting when a user might become active after a long period of inactivity, but is there any method I could use for short term (point by point) prediction? For example, if there's currently an upwards trend, predict from historical data (of the same or all users) at which level the user will plateau again?

If so, which method/algorithm (preferably in python/scikit-learn) would you suggest?

Thanks for your help!

Improve this question
$\endgroup$
3
  • 1
    $\begingroup$ I suspect this has some time component, such as holiday or weekend, that would be a consideration. You might find out the likelihood that people start programs running on a Friday and then come back on Monday to check results, or even start them at the end of the day to check in the morning - situations like that would skew the data. $\endgroup$
    – James Phillips
    Commented Nov 21, 2018 at 10:37
  • 2
    $\begingroup$ Starting with the total disk usage by all user will be easier (statistically stable). Then you can try to classify (cluster analysis) the users by grouping the users with similar behavior (inactive, random fluctuations, growing to plateau, bump,....). When that's done, you can average the pattern which is the disk usage of a user in a given group. $\endgroup$
    – AlainD
    Commented Nov 21, 2018 at 12:39
  • 1
    $\begingroup$ Individual user behavior might be difficult to predict, but aggregate time series might display specific patterns. $\endgroup$
    – Skander H.
    Commented Nov 21, 2018 at 18:30

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.