PhD defence of Muberra Ozmen – Graph-based strategies for classification with diverse label information

Abstract

This thesis comprises three interconnected projects addressing challenges in multi-label classification and temporal graph learning. In the first project, we tackle the challenge of modelling label dependencies in multi-label classification. We introduce a graph-based dependency module that is capable of modeling multiple types of relations. The module can be incorporated in embedding-based multi-label classification methods, leading to Relation Guided Message Passing (RGMP), a novel multi-label classification approach. We demonstrate via experiments that the proposed method achieves superior or comparable performance to state-of-the-art methods across all studied datasets, without imposing substantial additional model complexity or computational overhead. This emphasizes the importance of capturing diverse label dependencies.

Secondly, we addresses multi-label text classification in annotation-free and scarce-annotation settings. Our method leverages pre-trained language models for natural language inference, constructs a signed label dependency graph, and utilizes message passing along this graph to generate effective label predictions. In the weak supervision setting, where we have access to only a very small set of labelled data, our approach achieves significant performance improvement compared to existing techniques.

Finally, we introduce the task of recent link classification, which is important in industrial settings but has received little attention from the research community. In this task the goal is to predict the label of a recently observed edge between nodes. This problem arises when we observe an interaction between two entities (e.g., a potentially fraudulent transaction in a financial network), but will not have access to the label of the interaction until much later. We outline how this task can act as a benchmark task for evaluating Temporal Graph Learning (TGL) methods. We formalize the task, propose benchmark datasets, and evaluate state-of-the-art methods using robust metrics. We demonstrate how modifications in message aggregation, readout layer, and time encoding strategies can yield substantial performance improvement. Additionally, we present a novel learning architecture (Graph Profiler), capable of encoding previous events’ class information, achieving enhanced performance on most cases of interest.