Abstract:Graph models provide efficient tools to capture the underlying structure of data defined over networks. Many real-world network topologies are subject to change over time. Learning to model the dynamic interactions between entities in such networks is known as time-varying graph learning. Current methodology for learning such models often lacks robustness to outliers in the data and fails to handle heavy-tailed distributions, a common feature in many real-world datasets (e.g., financial data). This paper addresses the problem of learning time-varying graph models capable of efficiently representing heavy-tailed data. Unlike traditional approaches, we incorporate graph structures with specific spectral properties to enhance data clustering in our model. Our proposed method, which can also deal with noise and missing values in the data, is based on a stochastic approach, where a non-negative vector auto-regressive (VAR) model captures the variations in the graph and a Student-t distribution models the signal originating from this underlying time-varying graph. We propose an iterative method to learn time-varying graph topologies within a semi-online framework where only a mini-batch of data is used to update the graph. Simulations with both synthetic and real datasets demonstrate the efficacy of our model in analyzing heavy-tailed data, particularly those found in financial markets.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to learn time - varying graph models that can effectively represent heavy - tailed distribution data. Specifically, existing methods often lack robustness to external outliers when dealing with data with heavy - tailed distributions (such as financial data), and are unable to effectively handle noise and missing values. In addition, existing methods fail to fully capture graph topologies with specific structures and spectral properties. ### Main problem summary: 1. **Lack of robustness to external outliers**: Existing methods perform poorly when dealing with data with heavy - tailed distributions. 2. **Unable to effectively handle noise and missing values**: Many real - world datasets contain noise or missing values, and existing methods have not been able to meet this challenge well. 3. **Failure to capture graph topologies with specific structures and spectral properties**: Existing methods do not fully consider the structure and spectral properties of graphs, such as k - connected component graphs, which are very important in tasks such as data clustering. ### Method proposed in the paper: To solve the above problems, the author proposes a new framework based on a stochastic model, which can: - Use the non - negative vector autoregressive (VAR) model to capture the time - varying of graph weights. - Use the Student - t distribution to model signals from time - varying graphs to better handle heavy - tailed data. - Update graphs using mini - batch data in a semi - online framework, thereby improving efficiency and reducing computational costs. - Introduce spectral and structural constraints to ensure that the learned graph has specific properties, such as k - connected component graphs. ### Formula representation: The formulas involved in the paper include but are not limited to: - **Probability density function of the Student - t distribution**: \[ p(x_t|w_n)\propto(\det^*(L_{w_n}))^{1/2}\left(1 + \frac{x_t^{\top}L_{w_n}x_t}{\nu}\right)^{-(\nu + p)/2} \] where \(\nu> 2\), \(t\in F_n\), and \(L_{w_n}\) is the Laplacian matrix of the time - varying graph. - **Minimization objective function**: \[ \min_{w_n\geq0,a\geq0,X_n}\frac{1}{T_n\sigma_n^2}\|Y_n - M_n\odot X_n\|_F^2-\log\det^*(L_{w_n})+\frac{\nu + p}{T_n}\sum_{t\in F_n}\log\left(1 + \frac{x_t^{\top}L_{w_n}x_t}{\nu}\right)+\alpha\|w_n - a\odot\hat{w}_{n - 1}\|_1+\beta\|w_n\|_0+\gamma a^{\top}1 \] Through these improvements, this method can not only handle data with heavy - tailed distributions more robustly, but also effectively deal with noise and missing values, and learn graph topologies with specific structures and spectral properties.

Time-Varying Graph Learning for Data with Heavy-Tailed Distribution

Online Graph Topology Learning from Matrix-valued Time Series

Online Graph Learning for Time‐varying Graphs

Learning Spatiotemporal Graphical Models From Incomplete Observations

Time-varying graph learning from smooth and stationary graph signals with hidden nodes

Modeling and Analysis of Time-Varying Graphs

Efficient Variational Bayes Learning of Graphical Models With Smooth Structural Changes

A Unifying Model for Representing Time-Varying Graphs

An Efficient Vertex-Driven Temporal Graph Model and Subgraph Clustering Method

Time-varying Signals Recovery via Graph Neural Networks

Time-varying Graph Signal Estimation via Dynamic Multi-hop Topologies

Time-varying Graph Learning Under Structured Temporal Priors

Dynamic Chain Graph Models for Ordinal Time Series Data

Time-Dependent Graphs: Definitions, Applications, and Algorithms

Temporal Graph Neural Networks for Irregular Data

Bending the Future: Autoregressive Modeling of Temporal Knowledge Graphs in Curvature-Variable Hyperbolic Spaces

Random graph models for dynamic networks

On Sparse High-Dimensional Graphical Model Learning For Dependent Time Series

Sparse-GEV: Sparse Latent Space Model for Multivariate Extreme Value Time Serie Modeling

Robust Time-Varying Graph Signal Recovery for Dynamic Physical Sensor Network Data

Remaining Useful Life Estimation Under Uncertainty with Causal GraphNets