I have a data-set which has n number of features containing both categorical and continuous values. its a binary classification problem. The data set has many missing values. I am interested in imputing the values. For features such as amount/price I could use regression, Query : I could predict the values using other features(which have missing values), is this the right way?is the predicted value good/correct enough? Can I calculate the correlation between the variables for the whole data-set with missing values?
There is a multitude of ways on how to treat missing values.
The most straight-forward but brutal way is to drop all data points where values are missing. However, this is not necesserily the best way, especially if the percentage of data with missing values is large. Other than that it depends on the feature itself and you should consider what makes sense according to the feature's meaning. Other simple options include:
- Set a constant for the missing values.
- Set the missing value to a statistic of the feature (mean, median, mode)
Edit: As mentioned in the comments these methods might be too simple and lead to inferior results. A more sophisticated and well-performing approach is multiple-imputing.