Graph Filter Transfer via Probability Density Ratio Weighting

TL;DR

This paper introduces a transfer learning approach for graph filters that adapts to topology changes by using probability density ratios, improving signal reconstruction accuracy.

Contribution

It proposes a novel graph filter transfer method that accounts for topology changes using density ratio weighting, enhancing estimator robustness.

Findings

Outperforms existing methods on synthetic data

Effectively adapts to topology changes

Reduces reconstruction error in dynamic graphs

Abstract

The problem of recovering graph signals is one of the main topics in graph signal processing. A representative approach to this problem is the graph Wiener filter, which utilizes the statistical information of the target signal computed from historical data to construct an effective estimator. However, we often encounter situations where the current graph differs from that of historical data due to topology changes, leading to performance degradation of the estimator. This paper proposes a graph filter transfer method, which learns an effective estimator from historical data under topology changes. The proposed method leverages the probability density ratio of the current and historical observations and constructs an estimator that minimizes the reconstruction error in the current graph domain. The experiment on synthetic data demonstrates that the proposed method outperforms other methods.