0
$\begingroup$

This is the predictions of a binary classification model. The model is doing predicitons continuously, and these values are the sum of positive labels during a 10 hours period. As you can see, some of the locations are tend to generate positive labels, but really most of them are not true. The x-axis is location and the y-axis is time.

Is there a way, maybe, another model can learn the trends of locations (x-values) and with some kind of a reinforcement learning method, the model can learn that most of the positive predictions of these locations are actually false?

I don't know exactly how much of them are false positives, but I'm definitely sure really most of them are. Can I use this data to train another unsupervised model? Or even a smoothing algorithm maybe would work? Thank you.

$\endgroup$
5
  • $\begingroup$ maybe something like a model combination is what you need, e.g. boosting, random forest etc. otherwise, intuitively it seems you shouldn't be able to accomplish what you want. your model ate up all raw data and spat out noisy interpretation of it. now you want another model to improve it without access to raw data. that shouldnt work $\endgroup$ – Aksakal Jan 7 at 17:18
  • $\begingroup$ I don't actually want to improve the model. I want to cut out some false positives from locations that are tend to produce positive classes, but I don't know which method can do this. $\endgroup$ – emremrah Jan 7 at 17:44
  • $\begingroup$ What I mean is the model can still suck at these locations, it will still predict badly. I will filter/use the model's predictions. So it's not improving the model as you mentioned in the first comment. $\endgroup$ – emremrah Jan 7 at 18:30
  • $\begingroup$ I suppose something is missing in the context for me, because it's an unusual setting of a problem. In algorithmic trading often "signals" are used, such as sentiments on a given stock. The sentiment signal is an extremely noisy input, so nobody in right mind would trade on it alone. However, if there's a tiny bit of information in the noise, it can be useful along with other signals and data. So when you plug the noisy input, its slope/beta will tend to be very low among other variables, which is effectively like smoothing of sorts. $\endgroup$ – Aksakal Jan 7 at 18:35
  • $\begingroup$ In your case it appears that you're not planning to use raw data or other sources of data, and that your decision is solely based on this model which is very noisy or maybe even biased. So you want to somehow smooth it, and it's not obvious to me why smoothing should help. If the output in last time period was biased, noisy and wrong overall, then why would not letting the output change this period help? if you could explain this aspect it would help others answer your question. $\endgroup$ – Aksakal Jan 7 at 18:38

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.