None of the classifiers in q2-feature-classifier generate “random” results or give different answers each time you run them, unless if you change the reference sequences… or if there is something wrong with the query sequences, e.g., they are in mixed orientations. Since you say everything else was kept the same (e.g., reference), I suspect your reads must be mixed orientations... the classify-sklearn method assumes that all reads are in the same orientation.
You can use RESCRIPt to orient your reads prior to classification:
Alternatively, you can try the VSEARCH-based classifier, which is not sensitive to read orientation. That would be worth doing as a quick test before installing RESCRIPt and reorienting your sequences.