0
$\begingroup$

following this question, I'm making a data analysis again because I tried to use machine learning algorithms like Random forest to predict a value from certain features but it didn't work for me. I also tried Neural Networks and the results were worse than Random Forest.

So I thought maybe is something wrong with my data as I also read that data can have bad quality somehow(I'm not sure what does this mean). Since I'm not a data scientist and don't have background in Data Science in general (I'm an Electrical Engineer), I want to ask here and maybe someone can help me.

I read that correlation between features and target that I want to predict is important, now this is not my case, my features doesn't have a strong correlation with my features, does this mean that my data is bad and I can't predict that target from those features ?

I ll add here an image of the heatmap of my data: enter image description here

my target that I want to predict is xdistance and ydistance. It have negative correlation with one another but I think it is not important since I'm predicting every target on its own, but what is interesting here is that there is no feature that correlate strongly with one of the two targets. So I fear that this is an impossible task to try to predict those targets, am I right?

$\endgroup$
1
$\begingroup$

You may have heard

Correlation does not imply causation

but there is one more thats even more important

Without correlation there CAN BE causation. One example

SO just because they do not correlate does not mean that they are uninformative (think also about feature interactions)

I think you are looking for some test/theory that can give you some statement before hand. Without some HARD asssumptions thats impossible. Only way is to exhaustively try pretty different approaches and to reach desired metric scores.

|improve this answer|||||
$\endgroup$
  • $\begingroup$ Sure without correlation there can be causation. But in the example given, the interaction is causing a near zero correlation which is what the model will pick up. It doesn't matter if on a straight surface, speed increases by stepping harder on the pedal. If you feed to a model only data about hills, it won't pick up the natural speed/pedal relationship so it clearly depends on the data not on 'commong knowledge' and that's why you need to look at correlation matrices when doing EDA and not counting on what we know as humans. $\endgroup$ – Blenz Dec 19 '19 at 11:10
  • $\begingroup$ He needs to look deeper for places where there is indeed correlation, because on average as it appears there is none from which the model can predict the target. $\endgroup$ – Blenz Dec 19 '19 at 11:11

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.