Learning User Interface Semantics from Heterogeneous Networks with Multimodal and Positional Attributes