Algorithm Unrolling-based Denoising of Multimodal Graph Signals

Hayate Kojima, Keigo Takanami, Junya Hara, Yukihiro Bandoh, Seishi Takamura · 2025

Read Algorithm Unrolling-based Denoising of Multimodal Graph Signals by Hayate Kojima, Keigo Takanami, Junya Hara, Yukihiro Bandoh, Seishi Takamura (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-22175.txt.

Introduction

Hayate Kojima and co-authors clean noisy signals that live on graphs with several data kinds per node, in a 2025 paper. Sensor webs yield such layered data, tied both across space and across kinds. Often the tie graphs themselves are unknown and must be learned while cleaning.

Their method alternates between restoring the signal and learning the two graphs, with closed form updates for the signal. The loop's settings are learned from training data by unrolling the steps into a trainable stack. On made up and real data the method beat both model based and deep rivals. The work serves sensor and network data where noise and unknown ties coincide.

Catalog record

CREATORHayate Kojima, Keigo Takanami, Junya Hara, Yukihiro Bandoh, Seishi Takamura
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-22175

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Algorithm Unrolling-based Denoising of Multimodal Graph Signals Thanks: H. Kojima and K. Takanami contributed equally for this work. The preliminary version of this paper is presented in [ 1 ] . H. Kojima, J. Hara, H. Higashi and Y. Tanaka are with the Graduate School of Engineering, The University of Osaka, Osaka, Japan. K. Takanami is with the Graduate School of Engineering, Tokyo University of Agriculture and Technology, Tokyo, Japan. Y. Bandoh is with Shimonoseki City University, Yamaguchi, Japan. S. Takamura is with Graduate School of Computer and Information Sciences, Hosei University, Tokyo, Japan, and NTT Corporation, Kanagawa, Japan. Corresponding author: Hayate Kojima (email: [email protected]). Hayate Kojima, Keigo Takanami, Junya Hara, Yukihiro Bandoh, Seishi Takamura Abstract We propose a denoising method for multimodal graph signals by an alternating minimization scheme that sequentially solves signal restoration and graph learning problems. Many complex-structured data, i.e., those on sensor networks, can capture multiple modalities at each measurement point, referred to as modalities . They are also assumed to have an underlying structure or correlations in modality as well as space. Such multimodal data are regarded as graph signals on a twofold graph and they are often corrupted by noise. Furthermore, their spatial/modality relationships are not always given a priori: We need to estimate twofold graphs during a denoising algorithm. In this paper, we consider a signal denoising method on twofold graphs, where graphs are learned simultaneously. Specifically, the graph learning subproblems are solved using the primal-dual splitting (PDS) algorithm, while the signal update has a closed-form solution. Parameters in this iterative algorithm are learned from training data by unrolling the iteration with deep algorithm unrolling. Experimental results on synthetic and real-world data demonstrate that the proposed method outperforms existing model- and deep learning-based graph signal denoising methods. I INTRODUCTION Sensor networks have been extensively utilized in various applications including environmental monitoring, Internet of Things (IoT), and home automation systems [ 2 , 3 , 4 , 5 ] . Observed signals on sensor networks often corrupted by noise which affect to performances for the following applications. Therefore, denoising is a fundamental task for sensor network-based signal processing systems [ 6 , 7 ] . Networks are mathematically represented by graphs and signals on a network are called graph signals . Graph signal processing (GSP) includes theory and algorithms to analyze graph signals [ 8 , 9 , 10 ] . GSP mainly considers spatially distributed signal values whose domain is represented as nodes of a spatial graph . In sensor networks, however, we also encounter relationships among modalities since sensors often collect multiple features like temperature and humidity. Some modalities are strongly correlated and the others are not. This can be represented as a modality graph . It is visualized in Fig. 1 . As a result, we can assume many sensor network data can be viewed as signals lying on a twofold graph [ 11 ] . There are three main types of approaches to graph signal restoration: 1) model-based approach, 2) data-driven approach, and 3) hybrid approach. The model-based approach designs a regularized objective function based on the prior for the observed data and finds its minimum solution. The most widely used assumption is signal smoothness [ 12 , 13 ] . It is often enforced via graph low-pass filters [ 14 , 15 ] , Sobolev norms [ 16 ] , or graph wavelets/filter banks [ 17 ] . Piecewise constant signals have also been studied. In this case, total variation (TV) is often considered [ 18 ] . Furthermore, low-rank approximations of the graph signal matrix are utilized to recover the underlying data structure, often in combination with smoothness constraints [ 19 , 20 ] . More recently, generative approaches such as diffusion models have also been introduced to characterize complex signal distributions [ 21 ] . These model-based methods are typically solved using iterative algorithms. Note that the model-based method usually needs to tune hyperparameters, such as the regularization strengths. They are fixed throughout the iterative process, and in many cases, they are manually tuned. The data-driven approach learns (hyper)parameters from given training data with machine learning techniques. The current majority is based on graph neural networks (GNNs) [ 22 , 23 ] . The representative example is graph convolutional networks (GCNs) [ 24 , 25 , 26 , 27 , 28 ] . However, GNNs do not always outperform model-based methods for signal restoration [ 29 ] . This may stem from the fact that, in contrast to uniformly sampled signals like image and audio signals, deep layers for GNNs may not result in good performance [ 30 , 31 ] . We also encounter a challenge for training data: We need to collect enough number of data for training, but this is not often the case for the graph setting. The hybrid approach has been proposed to overcome the aforementioned challenges in the model- and data-driven methods. This often integrates (sometimes black-box) denoisers in an iterative algorithm [ 32 ] . In terms of the hyperparameter tuning, deep algorithm unrolling (DAU) [ 33 , 34 ] has attracted the signal processing community. DAU organizes each iteration of the model-based approach as a layer of neural network, and learns the hyperparameters in each layer from training data. DAU has the following benefits. First, it automatically learns hyperparameters in the iterative algorithm. While the convexity of the whole minimization problem is not guaranteed in general, it practically works well due to the different tuned parameters in different layers. Second, DAU usually requires a smaller number of parameters than the full neural networks: This contributes to reducing the training data. Third, since it is based on iterative minimization algorithms, the algorithm maintains some sense of interpretablity. For graph signal restoration, several DAU-based methods for single modal graph signals have been proposed [ 31 , 35 ] . An extension for multimodal graph signals has also been proposed [ 11 , 36 ] , leveraging two graphs representing spatial and modality relationships. However, all of them assume the graphs are given a priori. In this paper, we consider a denoising problem of multimodal signals on a twofold graph under the condition that the spatial/modality graphs are not given a priori: We need to learn/estimate twofold graphs for that case. The proposed method simultaneously learns a twofold graph from observed data during denoising. Specifically, an iterative algorithm is performed alternately for signal denoising and graph learning. In the proposed method, first, we design an optimization problem of denoising on twofold graphs and derive an iterative algorithm that leverages the smoothness of signals on the twofold graph, i.e., assuming the signals are smooth both on a spatial graph and a modality graph. Each iteration consists of two building blocks: Denoising in the spatial domain and that in the modality domain. Each building block is further divided into two phases: Graph learning and signal smoothing. In the graph learning phase, we utilize an existing technique [ 37 , 38 ] which assumes signal smoothness on the graph in the corresponding domain. In the signal smoothing phase, we apply the well-known Tikhonov regularization [ 12 ] to the multimodal signals. Finally, we unroll the

Read the complete text