Complex Attributed Network Embedding for medical complication prediction

摘要

To assure the development of effective treatment plans, it is crucial for understanding the complication relationships among diseases. In practice, traditional statistical methods are widely used to find the complications of diseases despite the potential errors introduced by the discrepancies in medical records. Recently, with the advances of network embedding techniques, it is promising to predict medical complications in properly constructed biomedical networks. However, due to the variety and sparsity of disease attributes, it is challenging to measure the similarity between attributes of different disease nodes, which seriously interferes the medical complication prediction task. To deal with this problem, in this paper, we propose a novel data-driven Complex Attributed Network Embedding (CANE) method to learn representation for each disease, which can better solve the variety and sparsity. Specifically, we first estimate the initial low-level representations of disease attributes via a matrix factorization technique and then refine the representations via several well-designed attribute modeling modules. Along this line, we introduce aggregation functions to preserve local structure information in the representations of diseases and apply them for complication prediction task. Finally, comprehensive experiments on real-world biomedical data clearly validate the effectiveness of CANE.