Abstract: This paper proposes the Modified Similarity Network Fusion (MSNF) algorithm, which realizes a direct fusion of non-square matrices by introducing self-built kernel matrices. At the same time, the kernel matrices are fused in advance to avoid the extension of negative influences caused by the kernel matrices constructed based on poorly behaved features. Experimental results on the three datasets demonstrate that the MSNF algorithm achieves higher recognition accuracy than the Similarity Network Fusion algorithm in cover song identification tasks, and greatly reduces the time complexity at the same time.