newbie here. I'm starting to work on a custom model for a very specific task, so I found no pre-trained models for this task so far.

After checking (un)supervised learning approaches I believe that regression will work the best, but could you please take a look and say if there's something better around?

The data is like that:

input1(int) input2(float) input3(int) => output(int)

The set of [inp1, inp2, inp3] is unique and can generate only one output. There are always two and only two such sets that generate the same output (like [2, 5.0, 3] => 0 and [12, -3.0, 1] =>0 too). It's always three inputs in a set (no less, no more), I see no reason to consider there's anything like a label, output varies from +100million to -100 million or even more in the future.

There will be no texts, the data is not time-correlated. The full dataset size is a few hundred million of such sets, if it matters.

So, is regression the best choice here? Any advice on this task for beginner?

Thank you.

  • 1
    $\begingroup$ What is the meaning of the output ? Is there some ordinal order ? Like is is better to predict $2$ or $58$ when you expect $3$ ? If it is classification both are usually as bad (like predicting horse or dog instead of pandas, both errors are as bad) while regression is sensible to numerics. $\endgroup$
    – Lelouch
    Commented Jul 11, 2023 at 12:06
  • 1
    $\begingroup$ Classification models predict a label. In your case that may be a whole number (e.g. 0, 1, 2, ...). Regression models predict a value. An important question is: How many possible outputs do you expect to have? If it is limited and the values are whole numbers you could treat the problem as a classification problem. $\endgroup$ Commented Jul 11, 2023 at 12:55
  • $\begingroup$ @Lelouch, Brian, it is hardly classification because output varies from +100million to -100 million or even more in the future. And yes, getting 58 is much worse than 2 if we expect 3. $\endgroup$
    – Putnik
    Commented Jul 11, 2023 at 16:51
  • 1
    $\begingroup$ So its probably regression then. Try the simplest model and increase the complexity if needed. $\endgroup$
    – Lelouch
    Commented Jul 11, 2023 at 19:16
  • $\begingroup$ @BrianO'Donnell What you said makes no sense, classification predicts discrete values, regression predicts continuous values, value and label have no interesting meaning in your comment. $\endgroup$
    – Dr. Snoopy
    Commented Jul 12, 2023 at 0:44


You must log in to answer this question.

Browse other questions tagged .