1
$\begingroup$

How can I check if a machine learning model is feasible on a given dataset? What techniques like EDA, correlation etc. can be used to judge if a model is possible i.e. data and predictor variables will give reasonably accurate forecasts or in other words there is good enough signal in predictors?

$\endgroup$
2
$\begingroup$

When in doubt simply build a model and test how good it is (accuracy, MSE, whatever).

If those KPIs are not up to par, think about improving the parameters, feature engineering, etc.

I find this approach to be more valuable than to spend a lot of time analyzing correlations and other classical statistical analysis.

If simple models (Regression, GLM, RandomForest, XGB) do not produce results with any sensible accuracy and you see no path to successful feature engineering then you have your answer as well.

The only downside to this approach might be a perfectly good model on the first try ;)

| improve this answer | |
$\endgroup$
  • $\begingroup$ Really helpful suggestion, agree corr analysis and EDA are time consuming. Someone suggested doing a PCA analysis which helps you quickly gauge whether model is feasible or not, what do you think of this approach? $\endgroup$ – Vikrant Arora Sep 19 '19 at 16:49
  • 1
    $\begingroup$ PCA is helpful in feature engineering and reduction. It will reduce the amount of predictors while usually improving quality. That said it doesn't tell you a lot about feasibility of the model. Cross-validating a first draft model is better for this than any other analysis. EDA will tell you whether prediction might be possible, validating a model will show whether it is. $\endgroup$ – Fnguyen Sep 19 '19 at 17:41

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.