🤖 machine learning

CondPSE: A Polynomial-Filtered Structural Encoder with Conditional Modulation for Graphs

CondPSE is a learned polynomial-filtered structural encoder that significantly outperforms existing methods in distinguishing non-isomorphic graphs on synthetic benchmarks through FiLM-style modulation, though its frozen application yields only comparable, rather than superior, results on real-world molecular property prediction tasks.

Original authors: Woohyun Lee, Hogun Park

Published 2026-07-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Woohyun Lee, Hogun Park

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 teach a robot to understand the shape of things. In the world of computer science, this robot is called a Graph Neural Network (GNN). Think of a "graph" not as a chart on a spreadsheet, but as a map of connections: dots (nodes) linked by lines (edges). These maps can represent anything from molecules in a chemistry lab to friends on a social network.

The problem is that these robots have a blind spot. They are great at looking at a single dot and its immediate neighbors, but they struggle to see the "big picture" of how the whole map is shaped. It's like trying to understand the layout of a city by only looking at one street corner; you might know the coffee shop next door, but you won't realize you're in a cul-de-sac or a grid system. Scientists call this limit the "1-WL test." To fix this, researchers give the robot "positional encodings"—special badges or ID cards that tell the robot where a dot sits in the grand scheme of things. Some of these badges are pre-made (like a map drawn by a human), while others are learned by the robot itself through practice. The big question is: if we train a robot to become a master at recognizing shapes in a video game, will that skill actually help it when it goes to work in the real world?

This paper introduces a new, super-smart robot badge-maker called CondPSE. The researchers wanted to see if they could build a better badge system that helps the robot distinguish between tricky, look-alike shapes. They created CondPSE by combining two powerful ideas: a "polynomial filter bank" and "conditional modulation."

Here is how it works, using a playful analogy. Imagine the robot is handed a bag of random, static-filled radio signals (these are the "Gaussian node probes"). In the old way, the robot would just listen to the static and try to guess the shape. CondPSE, however, puts these signals through a special set of polynomial filters. Think of these filters as a series of different-sized sieves or lenses. One lens blurs the signal to show the big, global shape of the graph; another lens sharpens it to show the tiny, local details. This creates a bunch of different "structural response branches"—like having a team of artists, each sketching the same object from a different perspective.

But CondPSE doesn't just stop there. It uses a technique called conditional modulation (specifically FiLM-style) to act like a conductor for this orchestra of artists. The conductor looks at the whole room:

Cross-filter: It compares what the "big picture" artist is seeing with what the "close-up" artist is seeing.
Local: It listens to the immediate neighborhood of the node.
Global: It checks the statistics of the entire graph.

Based on this mix of information, the conductor tells each artist, "You, focus more on the edges!" or "You, ignore the noise!" This refines the sketches into a perfect, unique ID card for every single node.

The researchers put this new badge-maker to the test in two very different arenas. First, they tested it on synthetic puzzles (specifically the CSL and EXP benchmarks), which are like logic games designed to trick robots. In these games, the goal is simply to tell if two shapes are identical or different. Here, CondPSE was a superstar. It boosted the accuracy of a standard robot from 42.9% to 97.3% on the CSL puzzle and from 68.3% to 99.9% on the EXP puzzle. The paper suggests that the "polynomial filter bank" (the different lenses) was the main hero, accounting for most of this massive improvement.

However, the story takes a twist when the robot moves from the logic game to the real world. The researchers tested CondPSE on molecular property prediction, which involves predicting how real chemical molecules behave. This is the "downstream" task where the robot actually has to do useful work. Surprisingly, CondPSE did not beat the previous best system (called GPSE). In fact, the results were a mixed bag: CondPSE was slightly better on some molecular datasets and slightly worse on others.

The paper explicitly argues against the idea that being a champion at shape-recognition puzzles automatically makes you a champion at real-world tasks. The authors suggest that while CondPSE is incredibly good at spotting structural differences (like telling two identical-looking rings apart), this super-power doesn't always translate to predicting chemical properties. They hypothesize that the "badges" CondPSE creates might be too detailed or focused on the wrong kind of shape features for the specific job of predicting molecular behavior. It's like having a robot that is perfect at identifying the exact geometry of a car, but that skill doesn't necessarily help it predict how fast the car will go.

In the end, the paper concludes that strong performance on structural discrimination benchmarks does not automatically guarantee a better outcome in downstream applications. The "CondPSE" encoder is a powerful diagnostic tool that proves we can build better structural encoders, but it also reveals that the path from "smart shape recognition" to "useful real-world prediction" is not a straight line. The success of these encodings depends heavily on how they are plugged into the final robot's brain (the downstream architecture), and sometimes, being too good at spotting tiny details might even get in the way of seeing the forest for the trees.

Technical Summary: CondPSE

Problem Statement

Message-passing Graph Neural Networks (MPNNs) are fundamentally limited by the 1-Weisfeiler-Lehman (1-WL) test, rendering them unable to distinguish certain non-isomorphic graphs or capture specific topological structures. To address this, Positional and Structural Encodings (PSE) inject topology-derived signals into node features. While hand-crafted encodings (e.g., Laplacian eigenvectors, random-walk statistics) exist, learned PSE encoders like GPSE offer a transferable alternative by pretraining a single encoder to reconstruct multiple structural targets from random node probes. However, a critical gap remains: it is unclear whether strong structural discrimination capabilities at the benchmark level (e.g., distinguishing graph isomorphism classes) automatically translate into improved performance on downstream graph-level prediction tasks, particularly in molecular property prediction.

Methodology

The authors propose CondPSE, a learned structural encoder designed to generate robust positional and structural signals through a two-stage process: polynomial filtering followed by conditional modulation.

1. Input and Extraction Phase

CondPSE begins with standard Gaussian node probes ( $Z$ ) sampled i.i.d. from a standard normal distribution. These probes serve as random inputs to query graph operators.

Polynomial Graph Filter Bank: Instead of using stacked message-passing layers (as in GPSE), CondPSE applies a bank of $K$ learnable polynomial graph filters of degree $d$ .
The filters compute mixtures of propagation-depth responses using powers of the adjacency matrix ( $A^i Z$ ).
The output of each filter is normalized and projected to the model dimension $D$ , forming $K$ initial structural-response branches. This stage captures graph-structural information through adjacency propagation rather than initial probe values.

2. Conditioning and Modulation Phase

The $K$ structural-response branches are refined through $B$ conditioning blocks. Each block employs FiLM-style modulation (Feature-wise Linear Modulation) to adaptively scale and shift the branches based on three distinct conditioning signals:

Cross-filter Conditioning: Computes pairwise comparison features between different filter branches to facilitate information exchange among the structural responses.
Local Conditioning: Applies a shared ResGatedGCN over each filter branch independently to aggregate edge-neighborhood information.
Global Conditioning: Computes graph-wide statistics (mean, standard deviation, maximum) for each branch and broadcasts them to nodes.
These signals are concatenated and mapped via an MLP to generate scale ( $\gamma$ ) and shift ( $\beta$ ) vectors, which modulate the hidden states of the filter branches.

3. Training and Integration

Pretraining: The encoder is pretrained to reconstruct node-level positional/structural targets (e.g., ElectrostaticPE, LapPE, RWSE) and graph-level invariants (e.g., Eigenvalues, Cycle counts). Crucially, this pretraining uses no downstream property labels.
Freezing: Once pretrained, the CondPSE encoder is frozen.
Downstream Usage: The resulting node encoding matrix is passed through a lightweight adapter (linear or MLP projection) and concatenated with raw node features before being fed into downstream GNN backbones (e.g., GIN, GraphGPS, Transformers).

Key Contributions

Architecture: Introduction of CondPSE, a structural encoder that utilizes a polynomial-filtered extraction stage followed by conditional modulation (cross-filter, local, and global) to refine structural responses.
Ablation Insights: Demonstration that the polynomial filter bank is the primary driver of performance gains in structural discrimination, while the conditioning modules provide additional, albeit smaller, refinements.
Empirical Gap Analysis: Evidence showing that superior benchmark-level structural discrimination (on CSL and EXP datasets) does not automatically guarantee superior downstream transfer performance in molecular property prediction.

Experimental Results

Synthetic Structural Discrimination

CondPSE was evaluated on the Circular Skip Link (CSL) and EXP graph isomorphism benchmarks using a GIN backbone.

CSL: Accuracy improved from 42.9% (GPSE) to 97.3% (CondPSE).
EXP: Accuracy improved from 68.3% (GPSE) to 99.9% (CondPSE).
Ablations: Removing the polynomial filter bank caused the most significant performance drop, confirming its role as the core extraction mechanism. Removing conditioning modules resulted in smaller but visible declines.

Molecular Property Prediction

CondPSE was evaluated on molecular datasets (ZINC, MolHIV, MolPCBA, PCQM4Mv2) using various backbones.

GraphGPS Backbone: CondPSE performed comparably to GPSE across all datasets, with slight variations (e.g., slightly better on MolPCBA, slightly worse on ZINC and MolHIV).
ZINC Backbone Sweep: When tested across different backbones (GCN, GatedGCN, GIN, GINE, Transformer), the relative performance of CondPSE and GPSE varied:
- CondPSE outperformed GPSE on GCN, GatedGCN, and Transformer backbones.
- GPSE outperformed CondPSE on GIN and GINE backbones.
Conclusion: There is no consistent ordering between the two encoders; the utility of a specific structural encoder depends heavily on the downstream architecture.

Significance and Claims

The paper positions CondPSE primarily as a diagnostic tool to study the alignment between benchmark-level structural discrimination and downstream utility.

Decoupling Discrimination from Transfer: The authors claim that strong structural discrimination capabilities (separating non-isomorphic graphs) do not automatically translate into better downstream graph-level predictions.
Role of Integration: The results suggest that the "downstream integration" mechanism—how a specific backbone combines raw features, message passing, and structural encodings—is a critical factor. Different architectures may favor different structural representations.
Target Mismatch: The authors hypothesize that a mismatch between pretraining targets (reconstructing generic structural statistics) and downstream labels (molecular properties) may limit transfer, even when the encoder is highly expressive.
Modest Scope: The paper does not claim CondPSE is a universal superior encoder. Instead, it highlights that the choice of structural encoder should be viewed as an architectural hyperparameter dependent on the downstream model, rather than a one-size-fits-all solution driven solely by isomorphism benchmark scores.