CE time series challenge preparation
Post date: Sep 15, 2016 7:07:26 AM
Sep 14, 2016
We had a discussion on datasets for the first CE challenge and how we should prepare a follow up challenge on CE pairs in time series. The main points were discussed in a follow up email exchange between Isabelle and Kun:
Isabelle - people are worried about confounders, how will we address this if we
deal with only pairs
Kun - This is a very good point. I think there will be a number of papers in
this direction of research. Some groups, including myself, are working on
related problems.
Isabelle - people think it would be interesting to compared the performance of
detecting causality with the time stamps and throwing away the time stamps
Kun - This is very smart. By throwing away the time stamps, do you mean doing
subsampling or aggregation or just shuffling the data?
Isabelle - people think that we should study a variety of cases and that
seasonality (among other cases) is important
Kun - Sure! There are many issues. We already investigated the effects of
subsampling. Now we are working on the effect of aggregation. Certainly
there are other issues. :-) It is a pity that I couldn't join the
discussion...
Isabelle - people think that stationary systems are going to need to be treated
differently from systems out of equilibrium and that we need to explore
various types of dynamics
Kun - This seems to be well known, and has been discussed in several papers
(see, e.g., Lacerda et al., 2008).
Isabelle - there is the worry that no generic method could work on all cases;
should we create a mixed bag with lots of cases (like for the cause-effect
pair challenge) or keep datasets separate?
Kun - Nice point. To me, it is still hopeful to come up with a rather
"universal" method. We are still at the beginning of causal
discovery--most of the existing methods have clear drawbacks; I personally
believe the state-of-the-art will be very different in 5 years or so...