2
$\begingroup$

I want to know if the following is a valid approach to create labels, if I have measurements under some conditions, and the conditions are similar but never exactly the same.

This doesn't correspond exactly to my real problem but for convenience lets say I have two WiFi Networks A and B. I want to know under which conditions A or B performs better.

My first step is to transmit Data over A and over B. I measure the network conditions and the time it takes to transmit the Data. The problem is that the captured conditions are never exactly the same. Hence I can't directly assign a label (e.g. "under this conditions A or B is better").

So I would perform a k-means clustering of the conditions and group similar conditions together. For each point in a cluster I lookup weather the transmission was performed with A or B and compare e.g. the medians of the transmission times.

Now I have a label (A better or B better) for each cluster center and can train a supervised model to generalize.

Is this a valid or common approach in those situations?

$\endgroup$

1 Answer 1

2
$\begingroup$

I'd suggest an alternative approach: train a regression model for each of the two networks A and B, which takes the conditions as input features and predicts the performance of the network under these conditions. Based on these two models, it is possible to directly find out which one is better under any conditions, by applying the two models and comparing their predicted performance.

I think that this approach is more direct in representing how the information is likely to impact the results. The clustering approach might work but it would lose some information in the process, because the clustering will introduce errors and the impact of the conditions on the performance wouldn't be directly represented in the model.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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