Regularized Simple Graph Convolution (SGC) for improved interpretability of large datasets

• Published in: Journal of big data 2020-10, Vol.7 (1), p.1-17, Article 91
• Main Authors: Pho, Phuong, Mantzaris, Alexander V.
• Format: Article
• Language: English
• Subjects: Big Data; Classification; Communications Engineering; Computational Science and Engineering; Computer Science; Convolution; Data Mining and Knowledge Discovery; Data points; Database Management; Datasets; Dimensionality reduction; Feature extraction; Graph Convolutional Network; Graph neural networks; Information Storage and Retrieval; Interpretability; Machine learning; Mathematical Applications in Computer Science; Network topologies; Networks; Neural networks; Parameters; Predictive analytics; Recommender systems; Scientific papers; Simple Graph Convolution; Social networks
• ISSN: 2196-1115; 2196-1115
• DOI: 10.1186/s40537-020-00366-x

Classification of data points which correspond to complex entities such as people or journal articles is a ongoing research task. Notable applications are recommendation systems for customer behaviors based upon their features or past purchases and in academia labeling relevant research papers in order to reduce the reading time required. The features that can be extracted are many and result in large datasets which are a challenge to process with complex machine learning methodologies. There is also an issue on how this is presented and how to interpret the parameterizations beyond the classification accuracies. This work shows how the network information contained in an adjacency matrix allows improved classification of entities through their associations and how the framework of the SGC provide an expressive and fast approach. The proposed regularized SGC incorporates shrinkage upon three different aspects of the projection vectors to reduce the number of parameters, the size of the parameters and the directions between the vectors to produce more meaningful interpretations.