What I normally do is I have a report that has the following:
- "Occurrences" metric
- my Classification dimension, but only for the "Unspecified" value.
- Breakdown that "Unspecified" value by the original dimension's values.
With that, I can see the values that are not being classified properly. Hopefully, you can identify a pattern for what's causing the missing classifications.
I had a case where non-English characters were not being classified properly when using the UTF-8 encoding. Updating the Classification Importer CSV file to indicate "v3" in the 2nd row solved it for me (after working it out with Client Care).