| Summary: | 碩士 === 國立成功大學 === 統計學系 === 107 === The main academic contribution of this thesis effectively introduce global information into network embedding learning and acquire a significant improvement in the experiment of node classification and link prediction.
Specifically, the original network embedding learning which generates a node sequence through random walk only considers its local information (immediate neighbors).
In addition, global information such as node communities and social circles is being used as compressing nodes and forms multi-level super graphs, so that network embedding learning can be also learned from the global information.
The Compression-based Feature Representation Learning is a general-purpose architecture for global information in various of applications network embedding learning models such as deepwalk, node2vec, LINE, and struc2vec. The experimental result shows the accuracy of deepwalk and node2vec can be significantly improved from 0.3 to 0.6.%by about 15% - 20%.
|
|---|
