Multiscale Graph Construction Using Non-local Cluster Features

Reina Kaneko,Hayate Kojima,Kenta Yanagiya,Junya Hara,Hiroshi Higashi,Yuichi Tanaka

DOI: https://doi.org/10.1109/MLSP58920.2024.10734767

2024-11-13

Abstract:This paper presents a multiscale graph construction method using both graph and signal features. Multiscale graph is a hierarchical representation of the graph, where a node at each level indicates a cluster in a finer resolution. To obtain the hierarchical clusters, existing methods often use graph clustering; however, they may ignore signal variations. As a result, these methods could fail to detect the clusters having similar features on nodes. In this paper, we consider graph and node-wise features simultaneously for multiscale clustering of a graph. With given clusters of the graph, the clusters are merged hierarchically in three steps: 1) Feature vectors in the clusters are extracted. 2) Similarities among cluster features are calculated using optimal transport. 3) A variable $k$-nearest neighbor graph (V$k$NNG) is constructed and graph spectral clustering is applied to the V$k$NNG to obtain clusters at a coarser scale. Additionally, the multiscale graph in this paper has \textit{non-local} characteristics: Nodes with similar features are merged even if they are spatially separated. In experiments on multiscale image and point cloud segmentation, we demonstrate the effectiveness of the proposed method.

Computer Vision and Pattern Recognition,Signal Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that when constructing multi - scale graphs, existing Graph Signal Processing (GSP) methods usually only consider spatial relationships and ignore signal changes of nodes or clusters. This has led to the following problems:

1. **Failure to detect cluster boundaries related to signal changes**: Existing methods may not be able to correctly identify clusters with similar features but different signal changes.
2. **Inability to connect non - locally similar clusters**: Since only spatial relationships are considered, existing methods have difficulty in merging spatially separated but similarly - featured nodes into the same cluster.

To overcome these problems, the author proposes a new multi - scale graph construction method that considers both the graph structure and node features and is implemented through the following steps:
1. **Feature extraction**: Extract feature vectors from each cluster.
2. **Similarity calculation**: Calculate the similarity between cluster features using Optimal Transport.
3. **Graph construction and spectral clustering**: Construct a Variable k - Nearest Neighbor Graph (VkNNG) based on similarity and apply spectral clustering to obtain coarser - grained clusters.

Specifically, the main contributions of the paper include:
- Proposing a multi - scale graph clustering method that combines graph structure and node features.
- Using optimal transport to evaluate the similarity between different clusters.
- Constructing VkNNG to connect non - locally similar nodes.
- Verifying the effectiveness of this method in multi - scale image and point cloud segmentation tasks through experiments.

Through these improvements, this method can effectively connect non - local nodes with similar features while maintaining the original cluster boundaries.