I've got sets of time series data collected from Weibo. It contains the number of posts under certain topics over a year. It can be found that from time to time there're bursts of discussion on certain topic. What's interesting is some burst die down quickly while others sustain on the agenda for quite a long time and die down gradually. How to predict and model the decays given a time series data. Also, I've got data of some exogenous data including time series data of links, pictures, and retweeting and original data in the community. Can I predict the duration of decay from these variables?
1 Answer
$\begingroup$
$\endgroup$
1
I guess, a nice place to start is seminal work of John Kleinberg: https://www.cs.cornell.edu/home/kleinber/bhs.pdf Yes, this paper is outdated, but anyway it's still very influential and simple. Also, digging articles, citing this paper, could be very helpful. By the way, these guys solve very similar task ( and they even use Weibo logs as well!): http://www.csce.uark.edu/~xintaowu/publ/pakdd15.pdf
-
$\begingroup$ Thanks a lot for your sharing. I've gone through these two pieces and like the first one very much. However, it aims at a different problem as what I'm expecting. I might not make that point clearly in my questions. I'm actually working on time series data with fixed time gap for each action. Therefore the variable that consists the key part of the model is invariant in my case. The second piece seems relevant to mine. It mentioned the Cantelli’s inequality that serves as burst detection theorem. $\endgroup$– zhiziCommented Nov 3, 2015 at 8:29