I'm increasingly running across the case in machine learning problems where I have a double-type, continuous variable column with some NAs which should not be imputed by kNN or median type methods (or, in fact, at all).
For example, if I have a garage/no garage column in the standard real estate value problem, and also an 'age of garage' column, one cannot impute the age column. There is no garage. One cannot set it to zero, because that would be a fresh garage. What is the standard method of dealing with this?
Can you please give more detail on your problem. Is it a clustering, classification, or a regression problem?
In a regression case, I would add a random noise which is not correlated with the output and I would also consider the interaction effect.
I would hope that in this case "having a garage" comes out significant, "age of garage" insignificant, but "having a garage x age of the garage" significant.
In the classification case, you need to think of a metric of similarity.
I guess it is safe to say the distance between a house with a garage of 10y old and a house with a garage of 15y is 5.
But what is the distance/dissimilarity between a house with a garage of 10y old and a house without a garage. Maybe you can assume a house without a garage is like a house with a very old garage that no one would use; i.e., assign a random number higher that observed maximum.
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running