| Title | : | Higher Order Propagation Framework for Deep Collective Classification |
| Speaker | : | Priyesh V (IITM) |
| Details | : | Tue, 17 Apr, 2018 2:00 PM @ Ada Lovelace Confere |
| Abstract: | : | In many real-life applications, entities in an environment are not independent but rather influenced by each other through their interactions. Such relational datasets have been unanimously modelled as graphs where the entities make up the node and the edges represent an interaction. Collective classification (CC) is a semi-supervised graph-based machine learning paradigm where the task is to classify unlabeled nodes in a partially labelled graph. We proposed a modular framework for CC called the Higher Order Propagation Framework (HOPF) comprising of a differentiable graph kernel and an iterative inference mechanism. We were able to better analyze the relative merits and demerits of numerous existing models as special instantiations of our highly modular framework. Through these analyses, we were able to point out the incapacities of existing models to effectively capture information from multiple hops in light of scalability and two previously unreported issues of Node Information Morphing (NIM) and Recursive Weight Dependencies (RWD) among multiple hops. To address these issues, we first proposed a specific instantiation of HOPF, called the Node Information Preserving (NIP) models, which preserves the node information at every propagation step. Then, we combined it with a linear fusion component to mitigate the RWD issue. And finally, we coupled them with an iterative inference component that allows us to scale to multiple hops by incorporating distant hop information concisely as summaries of the inferred labels. An extensive comparison with existing CC approaches on 11 datasets from different domains shows that unlike existing algorithms, the proposed models are robust across all the datasets and can handle full neighbourhood information in a scalable manner. |
