📊 statistics

Multiscale Hodge Scattering Networks for Data Analysis

This paper introduces Multiscale Hodge Scattering Networks (MHSNs), a novel framework for analyzing signals on simplicial complexes that leverages multiscale basis dictionaries to extract permutation-invariant, robust features through a CNN-like structure, enabling high-accuracy classification with significantly fewer trainable parameters than modern graph neural networks.

Original authors: Naoki Saito, Stefan C. Schonsheck, Eugene Shvarts

Published 2026-03-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Naoki Saito, Stefan C. Schonsheck, Eugene Shvarts

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand a complex story, but instead of reading words on a page, you are looking at a tangled web of connections. Maybe it's a map of friendships, a chemical molecule, or a network of scientific articles.

For a long time, computers have been great at reading "flat" data (like images or lists of numbers). But when data is shaped like a web or a 3D structure, it gets tricky. This paper introduces a new tool called Multiscale Hodge Scattering Networks (MHSNs). Think of it as a "super-magnifying glass" that can look at these complex webs from every angle, zooming in and out, to find the hidden patterns that matter.

Here is a simple breakdown of how it works, using some everyday analogies:

1. The Problem: Looking at a Web, Not Just a Dot

Most computer programs look at data as individual dots (nodes). Imagine a social network where you only see who is friends with whom.

The Old Way: Traditional tools look at the dots. They might say, "Bob has 5 friends, Alice has 10."
The New Way (MHSN): This paper says, "Wait! The shape of the group matters too." It looks at the triangles, the pyramids, and the clusters. In a social network, it doesn't just see Bob and Alice; it sees the triangle formed by Bob, Alice, and Charlie. It treats these shapes as the basic building blocks of the story.

2. The Tool: The "Swiss Army Knife" of Zooming

To analyze these shapes, the authors created a special set of lenses called Multiscale Basis Dictionaries.

The Analogy: Imagine you have a giant, messy room (your data). You want to clean it up and understand its layout.
- Zooming Out: You look at the whole room from the ceiling (the big picture).
- Zooming In: You look at just the corner with the toys.
- Zooming In More: You look at a single toy.
How MHSN does it: It uses two special "Swiss Army Knives" (called κ-GHWT and κ-HGLET) that can slice the data into these different zoom levels automatically. Unlike other tools that use a fixed set of lenses, these knives are custom-made for the specific shape of your data. They can cut a triangle, a square, or a pyramid perfectly.

3. The Process: The "Scattering" Machine

Once the data is sliced into these zoom levels, the MHSN runs it through a machine that acts like a laundry sorter.

Filter: It passes the data through the "lenses" (the wavelet filters).
Fold: It takes the absolute value (ignoring whether a number is positive or negative, just keeping the "size" of the signal). This is like folding a shirt so the wrinkles don't matter, only the shape does.
Stack: It repeats this process, layer by layer, creating a deep stack of features.
The Magic "Pooling": This is the paper's secret sauce. In traditional AI, you have to decide where to summarize the data. MHSN has a natural way to do this called Pooling.
- Global Pooling: Summarizing the entire room into one sentence. (Good for saying "This is a bedroom.")
- Local Pooling: Summarizing just the "toy corner" or the "bed corner." (Good for saying "There are toys here, but the bed is empty.")
- Why it matters: This allows the computer to understand both the big picture and the small details without getting confused.

4. The Superpower: Translation Invariance

Imagine you take a photo of a cat. If you move the cat one inch to the left, it's still the same cat.

Old AI: Sometimes gets confused if the cat moves.
MHSN: It is invariant. It doesn't care if you shuffle the names of the people in your social network or rotate the molecule. As long as the connections and the shapes stay the same, the MHSN recognizes it as the same object. It's like recognizing a song even if you hum it in a different key.

5. Why This is a Big Deal

The authors tested this on three very different problems:

Classifying News Articles: They turned keywords into shapes. The MHSN could tell the difference between "Physics" and "Medicine" articles better than older methods, even with very little training data.
Classifying Graphs (The "Domain"): They looked at networks of co-authors or chemical molecules. The MHSN was almost as good as the most advanced AI (which takes millions of parameters to learn) but used only 256 parameters.
- Analogy: Imagine a master chef who can cook a 5-star meal using only 3 spices, while the other chefs need a warehouse full of ingredients. MHSN is that efficient chef.
Predicting Molecular Energy: They predicted how much energy a molecule has. Again, it matched the performance of huge, complex AI models but was much faster and lighter.

The Bottom Line

This paper introduces a new way for computers to "see" complex, 3D, web-like data. Instead of trying to force this data into a flat list, it respects the natural shapes (triangles, pyramids) and uses a smart, multi-zoom lens system to extract the most important features.

The result? A tool that is:

Smarter: It understands the shape of the data, not just the numbers.
Sturdier: It doesn't get confused by rearranging the data.
Efficient: It does the work of a giant AI with a tiny fraction of the computing power, making it possible to run on a regular laptop instead of a supercomputer.

1. Problem Statement

The paper addresses the challenge of extracting robust, permutation-invariant features from signals defined on simplicial complexes (arbitrary high-dimensional structures including nodes, edges, triangles, and higher-dimensional analogues).

While traditional Scattering Transforms (Mallat) and Geometric Scattering Networks (Gao et al.) have been successful for images and node-based graph signals, they face limitations when applied to higher-dimensional data:

Dimensionality: Existing methods are largely restricted to node-valued signals ( $\kappa=0$ ).
Basis Limitations: Prior work relies on Morlet wavelets (for images) or Diffusion Wavelets (for graphs), which may not optimally capture the local, piecewise-constant nature of many simplicial signals.
Pooling Difficulty: Defining "local pooling" (aggregating features over local neighborhoods) is non-trivial in non-Euclidean domains without a hierarchical structure, often forcing reliance on global pooling which loses spatial information.

2. Methodology

The authors propose Multiscale Hodge Scattering Networks (MHSNs), a framework that generalizes scattering transforms to $\kappa$ -dimensional signals using specific multiscale basis dictionaries.

A. Theoretical Foundation: Hodge Laplacians and Multiscale Bases

The method operates on the space of functions defined on $\kappa$ -simplices ( $X_\kappa$ ). It utilizes the Hodge Laplacian ( $L_\kappa$ ) to define the spectral structure of the complex.

$\kappa$ -HGLET (Hierarchical Graph Laplacian Eigen Transform): A generalization of the Hierarchical Block DCT. It constructs a dictionary by recursively bipartitioning the simplicial complex using Fiedler vectors of the Hodge Laplacian. It provides an overcomplete set of orthonormal bases (ONBs) with $O(n^3)$ construction cost.
$\kappa$ -GHWT (Generalized Haar–Walsh Transform): A generalization of the Haar-Walsh wavelet packets. It uses the same hierarchical bipartition tree but constructs basis vectors via local averaging and differencing (Haar) and Walsh operations. It offers $O(n \log n)$ construction cost and shares the same support structure as HGLET.

B. The Scattering Transform Architecture

The MHSN adopts a layered architecture analogous to Convolutional Neural Networks (CNNs) but with fixed, non-learnable filters:

Filtering: The input signal $f$ is convolved with the multiscale basis dictionaries ( $\Phi_j$ ) at various scales $j$ .
Nonlinearity: A pointwise modulus (absolute value) operation is applied to the coefficients.
Pooling:
- Global Pooling: Averages coefficients over the entire domain (permutation invariant).
- Local Pooling: Averages coefficients over disjoint regions defined by the hierarchical partition tree at a specific scale. This is a novel contribution, allowing for local feature aggregation without arbitrary window sizes.
- No Pooling: Retains spatial resolution (useful when node ordering is fixed).
Cascading: The process is repeated for higher orders ( $m=2, 3, \dots$ ), where the output of one layer is fed into the next.
Feature Extraction: The final features are the moments (up to order $Q$ ) of the pooled coefficients.

C. Theoretical Properties

The authors prove that MHSNs possess critical mathematical properties:

Non-expansiveness: The transform is a non-expansive operator ( $\|Sf_1 - Sf_2\| \le \|f_1 - f_2\|$ ), ensuring stability against input perturbations.
Permutation Invariance: The globally pooled features are invariant to the reordering of simplices (node permutation), making them suitable for graph classification.
Stability: The features are stable to deformations of the underlying simplicial complex.

3. Key Contributions

Generalization to Higher Dimensions: The first scattering network framework explicitly designed for signals on $\kappa$ -simplices ( $\kappa > 0$ ), moving beyond node-only signals.
Novel Basis Dictionaries: The integration of $\kappa$ -HGLET and $\kappa$ -GHWT as filter banks, which are better suited for simplicial structures than Diffusion Wavelets.
Natural Local Pooling: The hierarchical bipartition tree induced by these bases allows for a mathematically rigorous definition of local pooling, bridging the gap between global invariance and local feature preservation.
Parameter Efficiency: The method achieves competitive accuracy with orders of magnitude fewer trainable parameters than modern Graph Neural Networks (GNNs) because the feature extractor is deterministic and fixed.

4. Experimental Results

The authors evaluated MHSNs on three distinct problem types:

A. Signal Classification (Science News Dataset)

Task: Classify scientific articles based on keyword graphs.
Finding: Traditional globally-pooled scattering networks failed to differentiate classes due to the sparsity of the signals. However, locally-pooled MHSNs achieved high accuracy (up to ~89.5%), performing on par with non-pooled (non-invariant) methods. The $\kappa$ -GHWT and $\kappa$ -HGLET bases outperformed Diffusion Wavelets.

B. Domain (Graph) Classification

Task: Classify graphs (e.g., social networks, molecular structures) from datasets like COLLAB, IMDB, MUTAG, and PROTEINS.
Finding: MHSNs achieved competitive accuracy with State-of-the-Art (SotA) GNNs (e.g., GCN, GAT, UGT).
Efficiency: While SotA GNNs required hundreds of thousands of parameters (e.g., UGT with ~867k parameters), MHSNs achieved similar results with only 256 parameters. This allows training on consumer-grade hardware.

C. Molecular Dynamics (Regression)

Task: Predict potential energy surfaces for molecules (Aspirin, Paracetamol) from the rMD17 dataset.
Finding: MHSNs (using SVR) achieved Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) comparable to specialized SotA models like SchNet and PaiNN.
Insight: Analyses using edge and triangle-based signals ( $\kappa=1, 2$ ) outperformed node-only analyses, demonstrating the value of higher-dimensional topological features even in molecular graphs.

5. Significance and Impact

Interpretability: Unlike deep GNNs, MHSNs are interpretable; features are deterministic cascades of wavelet transforms, avoiding the "black box" nature of learned filters.
Data Efficiency: The method requires significantly less training data and computational resources, making it ideal for small-data regimes or resource-constrained environments.
Topological Insight: The success of higher-order ( $\kappa > 0$ ) signals suggests that topological features (relationships between edges, triangles, etc.) carry critical information often missed by node-centric GNNs.
Future Directions: The authors propose using these methods to identify important relationships in graphs to narrow the search space for attention mechanisms and graph transformers.

In summary, this paper presents a mathematically rigorous, highly efficient, and robust alternative to deep learning for graph and simplicial complex data, leveraging the structural properties of Hodge theory and multiscale analysis to achieve high performance with minimal learnable parameters.