Transductive Zero-Shot Recognition via Shared Model Space Learning