💻 computer science

SpectralMLP: Message-Passing-Free Dual-Frequency Spectral Filtering on Hypergraphs

SpectralMLP introduces a novel hypergraph neural network framework that replaces traditional aggregation with an efficient, message-passing-free dual-frequency spectral filtering mechanism using learnable Jacobi polynomial bases, achieving state-of-the-art performance across multiple benchmarks while maintaining a parameter count lower than most existing methods.

Original authors: Rong Qian, Yu Cheng, Hongbo Zhao

Published 2026-07-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rong Qian, Yu Cheng, Hongbo Zhao

Original paper licensed under CC BY 4.0 (https://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 massive, chaotic party where people aren't just talking to their neighbors, but to entire groups at once. In the world of computer science, this is called a "hypergraph." Unlike a normal map where lines connect only two dots, a hypergraph uses "super-connections" (called hyperedges) that can link a whole crowd of people together at the same time. This is how we model real life: a group chat, a family tree, or a team of scientists working on a project.

To make sense of these messy groups, computers use something called "neural networks." For a long time, the standard way to teach these computers was "message passing." Imagine a game of telephone where every person whispers what they know to their neighbors, who then whisper it to their own neighbors, and so on. It works, but it's slow, gets messy with big groups, and sometimes the message gets so diluted it stops making sense. A smarter, more theoretical idea is "spectral filtering." Instead of whispering, imagine the whole room suddenly playing a specific song. The computer doesn't listen to neighbors; it just adjusts the volume of different "frequencies" (like bass or treble) in the data to highlight what matters and mute what doesn't. The big question researchers have been asking is: Can we skip the noisy "whispering game" entirely and just use the "music" to solve problems?

Enter SpectralMLP, a new method proposed by Rong Qian, Yu Cheng, and Hongbo Zhao that says, "Yes, we can."

The Problem with the Old Way

Most existing computer models for these group networks rely on that "whispering game" (aggregation). They try to gather information from neighbors to build a picture of the whole. But this has three big headaches:

It gets slow: As the group gets huge, the whispering takes forever.
It gets messy: If the group is too big, the message gets squashed and lost.
It breaks on weird data: Sometimes, the math behind the groups is so strange (a problem the authors call "spectral degeneration") that the whispering game just stops working, leaving the computer confused.

The authors argue that we don't need to whisper at all. Since mathematically, "listening to neighbors" and "adjusting the music" are actually two sides of the same coin, they decided to try the music approach exclusively. They wanted to see if a model could learn just by filtering frequencies, without ever passing a single message.

The Solution: A Dual-Tuner Radio

The team built a framework called SpectralMLP. Think of it as a super-smart radio that doesn't need to talk to anyone to understand the party; it just tunes into the right frequencies.

The core of their invention is a special module called the Adaptive Dual-Frequency Spectral Filter (ADF-SF). Here is how it works in plain English:

The Two Channels: Instead of just one volume knob, this radio has two separate channels. One channel is tuned to "Low Frequencies" (the smooth, bass-heavy sounds that help groups of similar people blend together). The other channel is tuned to "High Frequencies" (the sharp, treble sounds that help distinguish different groups from each other).
The Smart Mixer: The magic happens in a "gate" that acts like a DJ. It looks at the data and decides, "Hey, for this specific group, we need more bass," or "For this other group, we need more treble." It mixes the two channels together perfectly for the job at hand.
No Whispering: Crucially, this whole process doesn't involve passing messages between nodes. It just does some fast math (multiplying sparse matrices) to adjust the frequencies. It skips the slow, messy steps of the old methods.

What They Found

The researchers tested this new "radio" on 10 different datasets (ranging from small groups like a zoo of animals to massive networks like a database of academic papers) and compared it against 9 other top methods.

Here is what the results showed:

It Wins Often: SpectralMLP ranked first or second on 7 out of the 10 datasets.
It Saves the Day on Hard Problems: The biggest win was on a dataset called Congress. In this specific case, the data was so weird that standard methods struggled. The old "whispering" methods dropped in performance, but SpectralMLP soared, beating the baseline by a massive 27.9%. This proves that their dual-frequency approach is a lifesaver when the data is tricky.
It's Efficient: The new method is also lean. It only added about 2.4% more "brain power" (parameters) to the basic model compared to the old, heavy methods. In fact, it uses fewer total parameters than most of the complex message-passing models it beat.
It's Flexible: The "DJ gate" learned to switch between low and high frequencies automatically. On some datasets, it realized it only needed the bass; on others, it needed the treble. This adaptability is why it worked so well across such different types of problems.

The Takeaway

The paper suggests that we might not need the complicated, slow "whispering game" to understand complex groups anymore. By simply tuning into the right frequencies with a smart, dual-channel filter, computers can understand hypergraphs faster, with less memory, and even better accuracy in difficult situations.

While the authors note that their method still relies on a basic "backbone" (a standard neural network) and that there are still some mathematical mysteries about how to apply this directly to the raw group structure without simplifying it first, the results are strong. They showed that for hypergraph learning, sometimes the best way to listen is to stop talking and just start filtering.

Technical Summary: SpectralMLP

Problem Statement
Current hypergraph neural networks (HGNNs) predominantly rely on aggregation mechanisms—such as local message passing, global attention, or spectral projection—to enrich node representations. These approaches face significant practical limitations: local message passing suffers from over-squashing and noise sensitivity; global attention incurs $O(N^2)$ complexity, limiting scalability; and exact spectral methods require $O(N^3)$ eigendecomposition, rendering them infeasible for large graphs. Furthermore, existing spectral methods often fail to address "spectral degeneration," a phenomenon where Laplacian eigenvalues cluster tightly (e.g., $\lambda_{max} - \lambda_{min} < 10^{-4}$ ), causing standard filters to collapse into constant responses and lose frequency discrimination. Additionally, a single spectral response cannot simultaneously capture the low-frequency signals beneficial for homophilic graphs and the high-frequency signals required for heterophilic or multi-class tasks.

Methodology
The authors propose SpectralMLP, a message-passing-free framework that replaces aggregation with dual-frequency spectral filtering using learnable Jacobi polynomial bases. The framework consists of three primary components:

MLP Backbone: An MLP transforms input features into hidden representations without any graph operations, serving as a fair baseline for comparison.
Adaptive Dual-Frequency Spectral Filter (ADF-SF): This is the core innovation. Instead of a single filter, ADF-SF decomposes hidden representations into two independently parameterized branches:
- Low-frequency branch: Governed by a learnable Jacobi polynomial basis with parameters $(\alpha_l, \beta_l)$ and coefficients $\theta_{low}$ .
- High-frequency branch: Governed by a separate learnable Jacobi basis with parameters $(\alpha_h, \beta_h)$ and coefficients $\theta_{high}$ .
- Adaptive Gating: A learned gate ( $g = \sigma(W_g[h_{low}; h_{high}] + b_g)$ ) dynamically fuses these branches on a per-node basis, allowing the model to adaptively balance homophilic and heterophilic signals.
- Residual Connection: The filtered output is combined with the original signal via a learnable scale $\beta_{res}$ to ensure stability and signal preservation.
Implementation Details: The method operates directly on the sparse normalized Laplacian (derived via clique expansion) using three-term recurrence relations. This avoids eigendecomposition entirely, reducing per-layer complexity to $O(K \cdot \text{nnz} \cdot H)$ , where $K$ is the polynomial order and $\text{nnz}$ is the number of non-zero elements in the Laplacian.

The training objective combines cross-entropy loss with a structure-aware smoothness regularizer that penalizes large Euclidean distances between nodes within the same hyperedge, computed efficiently via sampling.

Key Contributions

Message-Passing-Free Framework: SpectralMLP demonstrates that spectral filtering alone, without learnable aggregation operators, is sufficient for hypergraph node classification.
Learnable Jacobi Bases: The paper introduces the first application of learnable Jacobi polynomial filtering to hypergraphs. By making the parameters $(\alpha, \beta)$ learnable, the model adapts the polynomial basis to the specific spectral distribution of each dataset, adding only two parameters per filter.
Dual-Frequency Architecture: The framework explicitly addresses the diversity of spectral characteristics in hypergraphs by decoupling low- and high-frequency processing. This design allows the model to handle both homophily and heterophily, and crucially, mitigates spectral degeneration where standard filters fail.
Computational Efficiency: The method eliminates the need for eigendecomposition and neighborhood expansion, relying solely on sparse matrix-vector products.

Experimental Results
The authors evaluated SpectralMLP on 10 hypergraph node classification datasets (ranging from 101 to 41,302 nodes) against 9 baselines, including message-passing HGNNs, transformer-based models, and MLP-based methods.

Performance: SpectralMLP ranked first or second on 7 out of 10 datasets.
- Spectral Degeneration: On the Congress dataset, characterized by severe spectral degeneration ( $\lambda_{max} - \lambda_{min} < 10^{-4}$ ), SpectralMLP achieved 95.88% accuracy, significantly outperforming the MLP baseline (70.47%) and other state-of-the-art methods. This represents a gain of +27.9% over the MLP baseline under uniform settings.
- Multi-class Tasks: On ModelNet40 (40 classes), SpectralMLP achieved 98.08%, surpassing all baselines.
- Robustness: The dual-frequency mode served as a robust "zero-configuration" default, performing within 1% of the best single-branch variant across all datasets.
Parameter Efficiency: SpectralMLP added a median of roughly 2.4% additional parameters over the MLP backbone. Its total parameter count was lower than most message-passing baselines (e.g., AllSet, HyperGT, SheafHyperGCN), confirming that spectral filtering is a parameter-efficient alternative to learnable aggregation.

Significance and Claims
The paper claims that SpectralMLP establishes a new paradigm for hypergraph learning by proving that filtering is theoretically equivalent to aggregation in representational power but superior in practical efficiency. The authors argue that by bypassing aggregation, the framework avoids over-squashing, $O(N^2)$ attention costs, and $O(N^3)$ eigendecomposition. The dual-frequency design is presented as a solution to the "one-size-fits-all" limitation of standard spectral filters, enabling the model to automatically discover the optimal frequency balance for diverse tasks. The results on the Congress dataset specifically highlight the method's ability to recover performance in scenarios where existing spectral methods degenerate.

Limitations and Future Work
The authors note that the current MLP backbone sets an upper bound on performance and suggest that high-capacity residual connections or attention-augmented MLPs could further improve results. Future directions include developing adaptive polynomial orders ( $K$ ) via hypergradient optimization and designing spectral filters that operate directly on the hypergraph incidence structure rather than through clique expansion.

The Problem with the Old Way

The Solution: A Dual-Tuner Radio

What They Found

The Takeaway

Technical Summary: SpectralMLP

More like this