3
$\begingroup$

Background

I am currently working on a project that has me stumped about the difference between imputing a "missing value" and predicting an "unknown value".

So far I would understand imputation as the process of filling in gaps in data that were created due to the measurement process e.g. data that should have been known but was not measured along with anything else. Mostly this meant that the gaps in data - in my mind - didn't split the data set in a "complete training sample" and an "incomplete test sample".

Predictive models conversely, I used to predict unknown values based on data I had measured. These values were unknown because I could not measure them (e.g. because it was future data) but a complete data sample (e.g. with historic data) was available thus splitting data into a "training sample" and a "test sample".

However my current problem is blurring the lines between these two ideas.

Problem

Assuming a sample of products with known attributes (X1,X2,X3, etc.), I want to evaluate a currently unknown attribute Y.

My first step was to let Y rate by some experts for a sample products (this cannot be done for all products due to the sheer quantity) and then predict Y for all the other products using a machine-learning predictive model fitted to the products that have been hand-rated.

However I have been thinking that I could equally understand this as a problem of missing data (i.e. Y is missing for all unrated items) and therefore use mechanisms like MICE to impute Y.

Question

So is my train of thought valid, could I potentially use both predictive models (like RF, SVM, etc.) and imputation methods (like MICE, etc.) to fill in my unknown value Y for all products?

And if both is possible, how would I decide which method is more suited?

$\endgroup$
1
$\begingroup$

Typically imputation will relate to filling in attributes (predictors, features) rather than responses, while prediction is generally only about the response (Y).

Even if imputation is being used to refer to filling in Y's the purpose is different; you're not using it for the primary purpose of getting a prediction for that Y.

$\endgroup$
  • $\begingroup$ Thank you for your answer, additionally I have learned that Imputation always creates some random noise to reflect uncertainty of missing values while a prediction would always output the same value if the criteria is the same. $\endgroup$ – Fnguyen Aug 2 at 8:01
  • $\begingroup$ That's an excellent point, though I'd have said "often" rather than "always" in both cases. Google "mean imputation" and "median imputation" for example, to see that imputation does not always add random noise. Similarly there are many contexts where a prediction is an interval or even a whole distribution, not a single value. $\endgroup$ – Glen_b Aug 2 at 9:59

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.