Graph Based Relational Features for Collective Classification
This work addresses the need for simpler and more flexible relational learning methods for domains with sparse labeled data, though it is incremental as it builds on existing feature-based ML approaches.
The paper tackled the problem of improving classification accuracy by integrating relational information without complex inference procedures, showing that their relational features achieve results comparable to collective inference methods on three benchmark datasets and outperform them with additional information.
Statistical Relational Learning (SRL) methods have shown that classification accuracy can be improved by integrating relations between samples. Techniques such as iterative classification or relaxation labeling achieve this by propagating information between related samples during the inference process. When only a few samples are labeled and connections between samples are sparse, collective inference methods have shown large improvements over standard feature-based ML methods. However, in contrast to feature based ML, collective inference methods require complex inference procedures and often depend on the strong assumption of label consistency among related samples. In this paper, we introduce new relational features for standard ML methods by extracting information from direct and indirect relations. We show empirically on three standard benchmark datasets that our relational features yield results comparable to collective inference methods. Finally we show that our proposal outperforms these methods when additional information is available.