4

I am trying to solve a decision making problem. In it, information evolves and increases with time for each event observed, and the history of the event may be useful.

The problem is as follows: in an electric network at time $t_0$ there is signal that automatically triggers the possible need of a technical intervention in a network to correct the problem. The system at $t_0$ has a large set of general indicators describing the state of the network based on which I'd like to determine whether to proceed with an intervention at $t_{0}$.

It's possible to carry specific measurements at the point where the problem was detected (this takes some time $\Delta t$). At this stage we can again ask whether the intervention is necessary or not? We can iterate this procedure a few times. Then, my question is, which is the machine learning method best suited for this classification problem? Can it be formulated as a Survival problem?

ARB
  • 41
  • 4
  • Are you saying that at all times `ti` you have a set of features `xi`, and then if you perform the specific measurements, you'll have `xi` plus some new features, `x2i`? If the dimensionality changes if you perform these measurements, you'll need to formulate the problem carefully. – jamesmf Nov 24 '15 at 21:37
  • This looks like multi armed bandit (https://en.wikipedia.org/wiki/Multi-armed_bandit) problem to me. At different time moments you would like to make a max reward decision given some context, policy. It is implemented in Vowpal Wabbit which scales very well with large datasets. – Vladislavs Dovgalecs Jan 22 '16 at 21:48

1 Answers1

1

From the description of the problem, you can just choose whatever classification algorithm you like to classify the need of intervention at $t_{0}$.

The general rule for choosing a classifier for this is to start simple--e.g., by using nearest neighbor and iterate to more powerful classifiers until you get enough accuracy.

If additional measurements at $t_{0}$ can be made a classifier can be used to determine the necessity of intervention without this additional data, then based on the probability output of this classifier and a threshold the additional measurements can be triggered. Based on this measurements another probability can be output which hopefully uses more informative features.

If you care about $\Delta t$, you should think about wether there is a real time dependency. Does the need of intervention depend upon the series of classifications? Is the order of determining interventions as necessary important for you?

In my opinion, you do not have to care about ordering for this application. Rather, I would suggest pooling classifications for one instance according to the most frequent label or something like that.

Kyle.
  • 1,483
  • 1
  • 16
  • 32
Tobias Würfl
  • 101
  • 3
  • Can you provide some references for your "general rule for choosing a classifier" statement? I agree that parsimony should be a guiding principle in choosing an ML approach, but I disagree that that means starting with nearest neighbors. For example, linear SVM has a single parameter to optimize ($c$), whereas knn has two (choice of $k$ and the distance metric to use). – Kyle. Oct 24 '15 at 13:07
  • I was thinking along the lines of simplicity of the method. Everyone understands NN and is able to implement it if necessary. Linear SVM requires more thought. Additionally NN without the k and by just using euclidean distance has no parameters at all. – Tobias Würfl Nov 03 '15 at 23:20