Most Graph Neural Networks (GNNs) operate on a simple intuition: "birds of a feather flock together." This principle, known as homophily, assumes that connected nodes likely share the same label. Mathematically, standard GNNs act as low-pass filters, smoothing features across neighbors to reduce noise.
However, real-world networks often exhibit heterophily, where "opposites attract." In a protein-protein interaction network or a transaction graph for fraud detection, connected nodes are often intentionally different. In these cases, the standard low-pass smoothing of GNNs becomes a liability, leading to "over-smoothing" where distinct node features are washed away. To solve this, a growing body of research focuses on adaptive spectral filtering—mechanisms that allow GNNs to dynamically tune their frequency response to suit the underlying graph structure.
The Spectral View of Graph Signals
To understand why adaptive filtering works, we must view a graph through its spectrum. The graph Laplacian’s eigenvalues represent "frequencies." Low frequencies correspond to smooth signals where connected nodes have similar values, while high frequencies correspond to sharp signals where connected nodes differ significantly.
Traditional GNNs are often rigid low-pass filters. As Sitao Luan points out, this rigidity makes them struggle with heterophilic datasets where the most discriminative information resides in the high-frequency components. The research direction identified here moves away from fixed smoothing toward models that can "see" and "weight" different parts of the spectrum. For instance, EigenNetwork models, introduced by Vijay Lingam and colleagues, perform eigendecomposition to learn non-linear functions that modulate these eigenvalues, effectively selecting the most useful structural signals regardless of whether they are low or high frequency.
Learning the Filter Shape
If a graph is heterophilic, we might need a high-pass filter to emphasize differences, or a band-pass filter to capture specific patterns. Rather than manually designing these filters, researchers have developed learnable polynomial filters.
The choice of polynomial basis determines how flexibly a model can approximate the ideal frequency response. Xiyuan Wang and Muhan Zhang demonstrated that using Jacobi polynomials provides a highly flexible basis that adapts to varying signal densities. Similarly, CayleyNets, proposed by Ron Levie and Michael M. Bronstein, use complex rational filters with a learnable dilation parameter to "zoom in" on specific frequency ranges.
These adaptive bases allow the model to transition between smoothing and sharpening. For example, the $p$GNN architecture by Guoji Fu and Yatao Bian functions as a flexible "all-terrain" filter, capable of acting as either a low-pass or high-pass operator depending on the task.
From Global Filters to Local Adaptivity
A major insight in this field is that a single graph is rarely uniform. One region might be homophilic (a dense community), while another is heterophilic (a boundary between groups). A global spectral filter applied to the whole graph is often insufficient.
To address this, researchers like Yimeng Min and Guy Wolf introduced Geometric Scattering Attention Networks (GSAN). These models use node-wise attention to weight different frequency channels for every individual node. This allows the network to smooth features in uniform clusters while preserving high-frequency details at the boundaries.
Similarly, Jingwei Guo and Kaizhu Huang proposed Diverse Spectral Filtering (DSF), which generates unique filter weights for nodes based on their local structural context. This local adaptivity is crucial for tasks like anomaly detection. As Ziwei Chai and Jiarong Xu argue in their work on AMNet, outliers are essentially high-frequency signals. By using node-level attention to prioritize high-pass signals for potential anomalies and low-pass signals for normal nodes, the model can isolate outliers that standard GNNs would otherwise hide.
Navigating the Spectral Trade-offs
Adaptive spectral filtering also provides a unified framework to solve two of the most persistent problems in GNNs: over-smoothing and over-squashing.
Dai Shi and Junbin Gao showed that these issues are linked by a spectral trade-off. Their physics-informed approach treats feature propagation like a heat equation, using multi-scale spectral filtering to "reverse time" and sharpen features when they begin to blur. By adaptively mixing low-pass and high-pass views—a technique also explored in PolyGCL by Jingyu Chen and Zhewei Wei—models can maintain distinct representations even in deep architectures or bottlenecked structures.
Ultimately, the shift toward adaptive spectral filtering represents a move toward more "structural intelligence." By learning to navigate the graph's spectrum, these models stop assuming how nodes should relate and start learning how they actually relate.