Accepted paper at WI-IAT 2020

Authors

Evgeniy Faerman, Felix Borutta, Julian Busch, Matthias Schubert

The 2020 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT'20)
14–17 December 2020, Virtual

Abstract

In many applications, data is represented as a network connecting nodes of various types. While types might be known for some nodes in the network, the type of a newly added node is typically unknown. In this paper, we focus on predicting the types of these new nodes based on their connectivity to the already labeled nodes. To tackle this problem, we propose Adaptive Node Similarity Using Multi-Scale Local Label Distributions (Ada-LLD) which learns the dependency of a node’s class label from the distribution of class labels in this (a) node’s local neighborhood. In contrast to previous approaches, (b) our approach is able to learn how class labels correlate with labels in variously sized neighborhoods. We propose a neural network architecture that combines information from differently sized neighborhoods allowing for the detection of correlations on multiple scales. Our evaluations demonstrate that our method significantly improves prediction quality on real world data sets.