A semi-supervised multi-view similarity and discriminative representation joint learning method (SMLSDR) was proposed aiming at the problem that existing semi-supervised multi-view learning methods fail to fully mine complementary information among different views. In the labeled space, a deep view-specific representation learning network was constructed to fully explore and utilize complementary information across views. In the unlabeled space, a local structure-based cross fusion strategy of similarity matrices was adopted to fully excavate and use inter-view complementary information. Spectral clustering was employed to design a label propagation strategy in order to realize semi-supervised learning. Extensive experiments were conducted on five multi-view datasets in the field of machine learning to verify the effectiveness of SMLSDR. The experimental results demonstrated that the proposed SMLSDR outperformed baseline methods. Further ablation experiments prove that both deep discriminative representation learning in labeled space and similarity matrix fusion in unlabeled space contribute positively to semi-supervised classification.