Learning Time-Varying Graphs for Heavy-Tailed Data Clustering

Time-varying graph models serve as powerful tools for capturing the dynamic structure of data defined over networks, where the interactions between entities vary with time. Current methods for learning time-varying graphs are not robust to outliers and cannot effectively model data with heavy-tailed distributions, e.g., financial data. In this work, we investigate the problem of learning time-varying models with graph structure for efficient representation of heavy-tailed data. Unlike traditional approaches, we assume specific graph structures that further allow our model to be applied for data clustering. Our approach is based on a semi -online framework for estimation of a time-varying graph topology from data under a stochastic model. Numerical results depict the effectiveness of our model for clustering of heavy-tailed data, particularly financial data.