Aiming at the problem that the prediction method of protein function based on the current Protein-Protein Interaction (PPI) network has low precision and is susceptible to data noise, a new machine learning protein function prediction method named HPMM (HC, PCA and MLP based Method) was proposed, which combined Hierarchical Clustering (HC), Principal Component Analysis (PCA) and Multi-layer Perception (MLP). HPMM took comprehensive consideration from macro and micro perspectives. It combined the information of protein families, domains and important sites into the vertex attributes of PPI networks to alleviate the effect from the data noise of networks. Firstly, the features of function modules and attribute principal components were extracted by using HC and PCA. Secondly, a mapping relationship between multi-feature and multi-function, used to predict protein functions, was constructed by training the MLP model. Three homo sapiens PPI networks, which were annotated by Molecular Functions (MF), Biological Processes (BP), and Cellular Components (CC) respectively, were adopted in the experiments. Comparisons were performed among the HPMM algorithm, the Cosine Iterative Algorithm (CIA) and the Diffusing GO Terms in the Directed PPI Network (GoDIN) Algorithm. The experimental results indicate that HPMM can obtain higher precision and F-measure than algorithms CIA and GoDIN, which are purely PPI network based methods.