To understand learnable spectral filters, we must first view graph convolutions through the lens of signal processing. A graph signal—such as a set of feature vectors assigned to nodes—can be decomposed into different "frequencies" using the eigenvectors of the graph Laplacian. Low frequencies represent smooth variations where connected nodes share similar values, while high frequencies capture sharp differences between neighbors.
Traditional Graph Neural Networks (GNNs) act as rigid, hand-crafted sieves. For instance, standard graph convolutional layers operate primarily as fixed low-pass filters, repeatedly smoothing features across neighborhoods. While this works well for homophilic networks (where similar nodes connect), it fails on heterophilic networks (where dissimilar nodes connect) and leads to the notorious problem of over-smoothing in deep architectures. As noted by Nian Liu et al. in A General Graph Spectral Wavelet Convolution via Chebyshev Order Decomposition, relying on fixed, hand-crafted kernel shapes severely limits a model's capacity to adapt to diverse graph topologies and feature distributions.
The core breakthrough of learnable spectral filtering is replacing static mathematical heuristics with differentiable, data-driven transformations. Instead of applying a predetermined polynomial to the graph Laplacian, models parameterize the filter function itself, allowing backpropagation to shape the frequency response dynamically.
Parameterizing the Frequency Response
There are several distinct strategies for introducing flexibility into the graph spectrum:
- Learnable Polynomial Coefficients: Instead of fixing the weights of multi-hop neighborhood expansions, models can optimize them directly. For example, Eli Chien et al. in Adaptive Universal Generalized PageRank Graph Neural Network treat the coefficients of a Generalized PageRank expansion as learnable parameters, allowing the network to automatically balance local and global information. Similarly, Bingheng Li in PC-Conv: Unifying Homophily and Heterophily with Two-Fold Filtering utilizes learnable coefficients within a polynomial expansion to approximate arbitrary real-valued filter functions, allowing the network to transition smoothly between low-pass and high-pass behavior.
- Neural Network Generators: Rather than relying on explicit polynomials, some architectures pass the graph's eigenvalues or structural properties through a neural network to generate the filter. Renjie Liao et al. in LanczosNet: Multi-Scale Deep Graph Convolutional Networks use multi-layer perceptrons (MLPs) operating directly on approximate eigenvalues to learn data-driven frequency responses. In a more continuous framing, Matthias Minder et al. in FiGLearn: Filter and Graph Learning using Optimal Transport model graph filters as neural networks to approximate complex signal-generating processes directly from observed distributions.
- Learnable Interpolation and Basis Functions: Filters can also be parameterized via geometric constraints. Enyan Dai et al. in Learning Graph Filters for Spectral GNNs via Newton Interpolation introduce a set of learnable interpolation points to flexibly approximate arbitrary filter shapes in the frequency domain. To capture highly non-smooth or complex responses, GuoguoAi et al. in GrokFormer: Graph Fourier Kolmogorov-Arnold Transformers employ learnable Fourier series as activation functions, allowing the model to construct arbitrary filter shapes across the graph spectrum.
Why It Works: Dynamic Spectral Tuning
By making the filter parameters differentiable, the network can dynamically adjust its focus across different bands of the graph spectrum based on the task and data properties.
Consider a dataset that exhibits high heterophily. A standard GNN would smooth out the crucial differences between neighboring nodes, destroying the signal. A learnable spectral filter, however, can adapt its coefficients to suppress low frequencies and amplify high frequencies, effectively acting as an edge detector or sharpener. This capability is demonstrated by Moshe Eliasof et al. in pathGCN: Learning General Graph Spatial Operators from Paths, where learning weights along paths allows the network to transition from simple smoothing to complex spatial operations like sharpening.
Furthermore, the network can learn to "zoom in" on narrow, critical bands of the spectrum. Ron Levie et al. in CayleyNets: Graph Convolutional Neural Networks With Complex Rational Spectral Filters introduce a learnable scaling parameter applied to the graph Laplacian, enabling the model to dynamically adjust its spectral resolution and isolate localized features like community structures. Similarly, Mingxuan Ju et al. in Adaptive Kernel Graph Neural Network treat the maximal eigenvalue of the graph Laplacian as a learnable parameter for each layer, allowing the model to automatically optimize its frequency filter and switch between low-pass smoothing and all-pass signal preservation.
Fixed Low-Pass Filter (Standard GCN)
Eigenvalues: [ Low Freq (Amplify) | High Freq (Block) ]
Result: Smooths out all localized variations.
Learnable Spectral Filter (Adaptive)
Eigenvalues: [ Low Freq (Attenuate) | Band-Pass (Amplify) | High Freq (Amplify) ]
Result: Preserves both global structure and sharp neighbor differences.
Granular and Multi-Scale Adaptivity
Beyond global adjustments, learnable spectral filtering allows for fine-grained control across different feature channels and structural scales. Instead of applying a uniform filter to the entire graph signal, different dimensions of the data can be filtered independently.
Brian Jalaian et al. in AdaGNN: Graph Neural Networks with Adaptive Frequency Response Filter implement an architecture where each feature channel learns its own filter parameters. This per-channel adaptivity prevents over-smoothing in deep networks by allowing some channels to preserve high-frequency details while others capture global trends.
This principle also extends to non-parametric kernel methods and wavelets. Luciano Vinas et al. in Simplifying GNN Performance with Low Rank Kernel Models demonstrate that a non-parametric learner within a Reproducing Kernel Hilbert Space (RKHS) can learn an optimal spectral transformation directly from the data distribution, matching the performance of far more complex architectures. When applied to generative tasks and anomaly detection, Jianbo Zheng et al. in Dynamic Spectral Graph Anomaly Detection utilize trainable mixtures of spectral wavelets to adaptively learn the specific frequency patterns associated with anomalies, proving far more robust than static, hand-designed alternatives.
Key Foundational Work
Semi-Supervised Learning with Adaptive Spectral Transform (2016) — contributed 1 concept elements to this direction.
Collaborative filtering via graph signal processing (2017) — contributed 1 concept elements to this direction.