With the development of medical and health services informatization, patient similarity becomes an important task in reuse of Electronic Health Records (EHR). By using the physician feedback on EHR data, patient similarity problem can be transformed to supervised distance metric learning problem, the supervised information usually comes from the tags we make on one patient's EHR data. In the existing work of Patient similarity Computing, the utilization of supervised is pretty circumscribed, the similarity of two different patients is often depended on their EHR data tags' completely equality. But in fact, the patient's tags contains many dimensions, that methods ignores tags' own similarity. In this work, we use the patient's diagnose data as the supervised information and divide the target patient's neighbor area into many margins based on their similarity using metric learning. The supervised information is also more fully used in this algorithm. Finally, in the multi-label KNN classification evaluation experiment, the similarity metric learned from this algorithm performs better than other algorithms in Hamming Loss and a-Accuracy.