Hi @hraanan,
I wonder if your R1 and R2 reads are flipped / oriented incorrectly. There are other forum threads you can search related to this issue. Can you try running qiime rescript orient-seqs... command on your rep-seqs.qza file, then re-classify? Try with and without the --i-reference-sequences flag. If this is not provided then all sequences will be reverse complimented. Otherwise each read will be oriented based on how well it matches the provided reference alignment, i.e. SILVA*. These files are located here.
The issue is that the sklearn classifier requires that the sequence data match the orientation of the reference sequences. You can also use qiime feature-classifier classify-consensus-vsearch ... which does not care about orientation. However, if your reads are actually in mixed orientation... that is some are sequenced correctly (R1 is really the forward read, and R2 is really the reverse read), and others are not (R2 is really the forward read and R1 is really the reverse read), then any phylogenies, comparisons, etc, will be incorrect.