Network analysis method transforms the prediction of potential comorbidity relationships into link prediction problems on complex network. However，most existing similarity measurement methods only consider a certain aspect of network characteristics，which greatly affects the accuracy of prediction. In this paper，three weighted disease networks are established using the real medical datasets from different sources. By comparing the structural differences of different networks，the limitations of existing network similarity indicators are analyzed. On this basis，a new link prediction method based on supervised classification is proposed，which integrates multiple local and global similarity indexes as input feature vectors in order to more accurately evaluate the similarity between nodes. Thus，the effective prediction of potential comorbidity relationships is achieved. The experimental results show that the proposed method can effectively improve the accuracy of link prediction in comorbidity network and has better stability and adaptability in different disease network and classification algorithms.