SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.01920 · 2026

PIMPC-GNN: Physics-Informed Multi-Phase Consensus Learning for Enhancing Imbalanced Node Classification in Graph Neural Networks

Abdul Fofanah, Lian Wen, David Chen

arXiv · PDF · Open in the Atlas

Code that ran

6 of 7 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (5 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 6 have no contract verdict; 0 failures of our instrument; 1 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2602.01920.json · badge: /api/ran/2602.01920.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
afofanah/PIMPC-GNN community
Papers with Code archive
6 ran · 1 no recorded run
FunctionStatusWhere it lives
AdaptiveThreshold Ran
ran on a synthesized input · constructed an object; no result was computed
afofanah/PIMPC-GNN/models/model.py
code served (permissive licence) · get_code("281ef5f9a29511e0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, AdaptiveThreshold: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Set feature_dim to 8 (must be divisible by num_heads=8)
feature_dim = 8

# Create the model instance
model = nn.Module()

# Create input tensors
batch_size = 2
num_nodes = 3

# combined_features: (batch_size, num_nodes, feature_dim)
combined_features = torch.randn(batch_size, num_nodes, feature_dim, dtype=torch.float32)

# node_features: (batch_size, num_nodes, feature_dim)
node_features = torch.randn(batch_size, num_nodes, feature_dim, dtype=torch.float32)

# Build the AdaptiveThreshold model
adaptive_threshold_model = nn.Module()

# Create ARGS dict with all parameters for AdaptiveThreshold.__init__
ARGS = {
    'feature_dim': feature_dim,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type AdaptiveThreshold — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a AdaptiveThreshold output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
ContrastiveLoss Ran
ran on a synthesized input
afofanah/PIMPC-GNN/models/model.py
code served (permissive licence) · get_code("a6c95628ced93c31")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 0, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

INIT_ARGS = {
    'feature_dim': 4,
    'temperature': 0.1,
}

ARGS = {
    'features': torch.randn(2, 3, 4),
    'labels': torch.tensor([[0, 1, 0], [1, 0, 1]]),
    'class_counts': torch.tensor([2, 2]),
}

if __name__ == '__main__':
    from torch import nn
    import torch.nn.functional as F
    from typing import Dict, List, Tuple, Optional, Any

    class ContrastiveLoss(nn.Module):

        def __init__(self, feature_dim: int, temperature: float=0.1):
            super(ContrastiveLoss, self).__init__()
            self.temperature = temperature
            self.projection_head = nn.Sequential(nn.Linear(feature_dim, feature_dim, dtype=torch.float32), nn.LayerNorm(feature_dim), nn.GELU(), nn.Dropout(0.1), nn.Linear(feature_dim, feature_dim // 2, dtype=torch.float32), nn.LayerNorm(feature_dim // 2), nn.GELU(), nn.Linear(feature_dim // 2, feature_dim // 4, dtype=torch.float32))

        def forward(self, features: torch.Tensor, labels: torch.Tensor, class_counts: Optional[torch.Tensor]=None) -> torch.Tensor:
            batch_size, num_nodes, feature_dim = features.shape
            projected_features = self.projection_head(features)
            projected_features = F.normalize(projected_features, dim=-1)
            features_flat = projected_features.view(-1, projected_features.size(-1))
            labels_flat = labels.view(-1)
            if class_counts is not None:
                valid_mask = (labels_flat >= 0) & (labels_flat < len(class_counts))
                if not valid_mask.all():
                    features_flat = features_flat[valid_mask]
                    labels_flat = labels_flat[valid_mask]
                    if len(labels_flat) == 0:
                        return torch.tensor(0.0, device=features.device)
            pos_mask = labels_flat.unsqueeze(0) == labels_flat.unsqueeze(1)
            neg_mask = labels_flat.unsqueeze(0) != labels_flat.unsqueeze(1)
            sim_matrix = torch.mm(features_flat, features_flat.T) / self.temperature
            eye_mask = torch.eye(sim_matrix.size(0), device=sim_matrix.device).bool()
            sim_matrix.masked_fill_(eye_mask, -float('inf'))
            if class_counts is not None:
                class_weights = 1.0 / (class_counts.float() + 1e-08)
                class_weights = class_weights / class_weights.sum()
                max_label = labels_flat.max().item()
                if max_label >= len(class_weights):
                    pad_size = max_label + 1 - len(class_weights)
                    class_weights = F.pad(class_weights, (0, pad_size), value=1e-08)
                weight_matrix = class_weights[labels_flat].unsqueeze(0) * class_weights[labels_flat].unsqueeze(1)
                pos_mask = pos_mask.float() * weight_matrix
            exp_sim = torch.exp(sim_matrix)
            neg_exp = exp_sim * neg_mask.float()
            hard_neg_sum = torch.sum(neg_exp, dim=1, keepdim=True)
            pos_exp = exp_sim * pos_mask.float()
            pos_sum = torch.sum(pos_exp, dim=1, keepdim=True)
            pos_sum = torch.clamp(pos_sum, min=1e-08)
            hard_neg_sum = torch.clamp(hard_neg_sum, min=1e-08)
            contrastive_loss = -torch.log(pos_sum / (pos_sum + hard_neg_sum + 1e-08))
            valid_mask = torch.sum(pos_mask, dim=1) > 0
            if valid_mask.sum() > 0:
                contrastive_loss = contrastive_loss[valid_mask].mean()
            else:
                contrastive_loss = torch.tensor(0.0, device=features.device)
            return contrastive_loss

    model = ContrastiveLoss(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output: {output}")
    print(f"Output shape: {output.shape}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 0, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
GraphAttentionFusion Ran
ran on a synthesized input · constructed an object; no result was computed
afofanah/PIMPC-GNN/models/model.py
code served (permissive licence) · get_code("5991001f972de41c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, GraphAttentionFusion: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Define the arguments for GraphAttentionFusion
ARGS = {
    "input_dims": [4, 5, 6],  # Three feature types with different input dimensions
    "output_dim": 8,           # Output dimension for projection
    "num_heads": 2             # Number of attention heads (must divide output_dim)
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type GraphAttentionFusion — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a GraphAttentionFusion output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
KuramotoOscillator Ran
ran on a synthesized input · constructed an object; no result was computed
afofanah/PIMPC-GNN/models/model.py
code served (permissive licence) · get_code("5fc16f8ff965adce")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, KuramotoOscillator: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import math

torch.manual_seed(0)

class KuramotoOscillator(nn.Module):

    def __init__(self, feature_dim: int, num_steps: int=60):
        super(KuramotoOscillator, self).__init__()
        self.feature_dim = feature_dim
        self.num_steps = num_steps
        self.frequency_network = nn.Sequential(nn.Linear(feature_dim, feature_dim, dtype=torch.float32), nn.LayerNorm(feature_dim), nn.Tanh(), nn.Linear(feature_dim, feature_dim // 2, dtype=torch.float32), nn.Tanh(), nn.Linear(feature_dim // 2, 1, dtype=torch.float32))
        self.coupling_network = nn.Sequential(nn.Linear(feature_dim, feature_dim // 2, dtype=torch.float32), nn.GELU(), nn.Linear(feature_dim // 2, 1, dtype=torch.float32), nn.Sigmoid())
        self.sync_analyzer = nn.Sequential(nn.Linear(feature_dim + 4, feature_dim, dtype=torch.float32), nn.LayerNorm(feature_dim), nn.GELU(), nn.Dropout(0.1), nn.Linear(feature_dim, feature_dim // 2, dtype=torch.float32), nn.GELU(), nn.Linear(feature_dim // 2, 1, dtype=torch.float32), nn.Sigmoid())
        self.global_coupling = nn.Parameter(torch.tensor(0.5, dtype=torch.float32))

    def enhanced_kuramoto_dynamics(self, node_features: torch.Tensor, adjacency_matrix: torch.Tensor):
        batch_size, num_nodes, feature_dim = node_features.shape
        natural_frequencies = self.frequency_network(node_features).squeeze(-1)
        local_coupling = self.coupling_network(node_features).squeeze(-1)
        global_coupling = torch.clamp(self.global_coupling, 0.1, 2.0)
        phase_init = torch.atan2(torch.sum(node_features * torch.sin(torch.arange(feature_dim, device=node_features.device).float()), dim=-1), torch.sum(node_features * torch.cos(torch.arange(feature_dim, device=node_features.device).float()), dim=-1))
        if phase_init.dim() == 3:
            phases = phase_init.squeeze(0)
        elif phase_init.dim() == 2:
            phases = phase_init
        else:
            phases = phase_init.view(batch_size, num_nodes)
        sync_trajectory = []
        dt = 0.01
        for step in range(self.num_steps):
            phase_cos = torch.cos(phases).unsqueeze(-1)
            phase_sin = torch.sin(phases).unsqueeze(-1)
            neighbor_cos = torch.bmm(adjacency_matrix, phase_cos).squeeze(-1)
            neighbor_sin = torch.bmm(adjacency_matrix, phase_sin).squeeze(-1)
            coupling_effect = torch.atan2(neighbor_sin, neighbor_cos + 1e-08) - phases
            coupling_strength = local_coupling * global_coupling
            phase_derivative = natural_frequencies + coupling_strength * torch.sin(coupling_effect)
            phases = phases + dt * phase_derivative
            phases = phases % (2 * math.pi)
            if step % 10 == 0:
                complex_phases = torch.complex(torch.cos(phases), torch.sin(phases))
                order_param = torch.abs(torch.mean(complex_phases, dim=1))
                sync_trajectory.append(order_param)
        return (phases, natural_frequencies, local_coupling)

    def forward(self, node_features: torch.Tensor, adjacency_matrix: torch.Tensor):
        batch_size, num_nodes, feature_dim = node_features.shape
        phases, frequencies, coupling = self.enhanced_kuramoto_dynamics(node_features, adjacency_matrix)
        phase_features = torch.stack([torch.cos(phases), torch.sin(phases), torch.cos(2 * phases), torch.sin(2 * phases)], dim=-1)
        sync_input = torch.cat([node_features, phase_features], dim=-1)
        desync_scores = self.sync_analyzer(sync_input).squeeze(-1)
        complex_phases = torch.complex(torch.cos(phases), torch.sin(phases))
        order_parameter = torch.abs(torch.mean(complex_phases, dim=1))
        phase_std = torch.std(phases, dim=1)
        return {'desynchronization_scores': desync_scores, 'order_parameter': order_parameter, 'phases': phases, 'frequencies': frequencies, 'coupling_strength': coupling, 'sync_features': sync_input, 'phase_coherence': 1.0 / (1.0 + phase_std)}


feature_dim = 4
num_steps = 10

model = KuramotoOscillator(feature_dim=feature_dim, num_steps=num_steps)

batch_size = 2
num_nodes = 3

node_features = torch.randn(batch_size, num_nodes, feature_dim, dtype=torch.float32)
adjacency_matrix = torch.randn(batch_size, num_nodes, num_nodes, dtype=torch.float32)

ARGS = {
    'feature_dim': feature_dim,
    'num_steps': num_steps
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type KuramotoOscillator — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a KuramotoOscillator output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
SpectralEmbedding Ran
ran on a synthesized input · constructed an object; no result was computed
afofanah/PIMPC-GNN/models/model.py
code served (permissive licence) · get_code("17ba07f05030e9e7")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, SpectralEmbedding: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Build minimal arguments for SpectralEmbedding
ARGS = {
    "feature_dim": 8,
    "num_eigen_vectors": 4,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type SpectralEmbedding — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a SpectralEmbedding output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
Themodynamic Ran
ran on a synthesized input · constructed an object; no result was computed
afofanah/PIMPC-GNN/models/model.py
code served (permissive licence) · get_code("97bbfc272543e4b2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, Themodynamic: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F

torch.manual_seed(0)

class Themodynamic(nn.Module):

    def __init__(self, feature_dim: int, num_diffusion_steps: int=30):
        super(Themodynamic, self).__init__()
        self.feature_dim = feature_dim
        self.num_diffusion_steps = num_diffusion_steps
        self.thermal_conductivity = nn.Parameter(torch.tensor(0.15, dtype=torch.float32))
        self.heat_capacity = nn.Parameter(torch.tensor(1.0, dtype=torch.float32))
        self.heat_source_network = nn.Sequential(nn.Linear(feature_dim, feature_dim, dtype=torch.float32), nn.LayerNorm(feature_dim), nn.GELU(), nn.Dropout(0.1), nn.Linear(feature_dim, feature_dim // 2, dtype=torch.float32), nn.GELU(), nn.Linear(feature_dim // 2, 1, dtype=torch.float32), nn.Softplus())
        self.anomaly_detector = nn.Sequential(nn.Linear(feature_dim * 2, feature_dim, dtype=torch.float32), nn.LayerNorm(feature_dim), nn.GELU(), nn.Dropout(0.15), nn.Linear(feature_dim, feature_dim // 2, dtype=torch.float32), nn.GELU(), nn.Linear(feature_dim // 2, 1, dtype=torch.float32), nn.Sigmoid())

    def compute_enhanced_heat_diffusion(self, initial_heat: torch.Tensor, laplacian: torch.Tensor, dt: float=0.015) -> torch.Tensor:
        batch_size, num_nodes, feature_dim = initial_heat.shape
        laplacian_nodes = laplacian.shape[-1]
        if num_nodes != laplacian_nodes:
            if num_nodes < laplacian_nodes:
                laplacian = laplacian[:, :num_nodes, :num_nodes]
            else:
                pad_size = num_nodes - laplacian_nodes
                laplacian = F.pad(laplacian, (0, pad_size, 0, pad_size), mode='constant', value=0)
        current_heat = initial_heat.clone()
        thermal_conductivity = torch.clamp(self.thermal_conductivity, 1e-06, 0.5)
        heat_capacity = torch.clamp(self.heat_capacity, 0.1, 2.0)
        heat_trajectory = []
        for step in range(self.num_diffusion_steps):
            heat_gradient = torch.bmm(laplacian.expand(batch_size, -1, -1), current_heat)
            heat_gradient = torch.clamp(heat_gradient, -5.0, 5.0)
            gradient_norm = torch.norm(heat_gradient, dim=-1, keepdim=True)
            adaptive_dt = dt / (1 + 0.1 * gradient_norm)
            heat_change = adaptive_dt * thermal_conductivity * heat_gradient / heat_capacity
            heat_change = torch.clamp(heat_change, -0.5, 0.5)
            current_heat = current_heat + heat_change
            current_heat = torch.clamp(current_heat, -2.0, 2.0)
            if step % 5 == 0:
                heat_trajectory.append(current_heat.clone())
            if torch.isnan(current_heat).any():
                current_heat = initial_heat.clone()
                break
        return (current_heat, heat_trajectory)

    def forward(self, node_features: torch.Tensor, adjacency_matrix: torch.Tensor) -> dict:
        batch_size, num_nodes, feature_dim = node_features.shape
        degree_matrix = torch.diag_embed(torch.sum(adjacency_matrix, dim=-1) + 1)
        laplacian = degree_matrix - adjacency_matrix
        heat_sources = self.heat_source_network(node_features)
        initial_heat = node_features * heat_sources
        diffused_features, trajectory = self.compute_enhanced_heat_diffusion(initial_heat, laplacian)
        anomaly_input = torch.cat([node_features, diffused_features], dim=-1)
        anomaly_scores = self.anomaly_detector(anomaly_input).squeeze(-1)
        return {'anomaly_scores': anomaly_scores, 'diffusion_features': diffused_features, 'heat_sources': heat_sources.squeeze(-1), 'heat_trajectory': trajectory}


feature_dim = 4
num_diffusion_steps = 10
batch_size = 2
num_nodes = 3

node_features = torch.randn(batch_size, num_nodes, feature_dim, dtype=torch.float32)
adjacency_matrix = torch.randn(batch_size, num_nodes, num_nodes, dtype=torch.float32)

ARGS = {
    'feature_dim': feature_dim,
    'num_diffusion_steps': num_diffusion_steps
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Themodynamic — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Themodynamic output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
PIMPC_GNN Not yet run
unverified (harvested, no recorded run)
afofanah/PIMPC-GNN/models/model.py
code served (permissive licence) · get_code("5a583b0b20412a5d")
How it ran

no invocation recorded.

Also present in other papers' code

Of 7 samples lifted from this paper's linked repositories, 7 carry a function identity and 0 do not. Of the 7 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 7 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Graph neural networks (GNNs) often struggle in class-imbalanced settings, where minority classes are underrepresented and predictions are biased toward majorities. We propose PIMPC-GNN, a physics-informed multi-phase consensus framework for imbalanced node classification. Our method integrates three complementary dynamics: (i) thermodynamic diffusion, which spreads minority labels to capture long-range dependencies, (ii) Kuramoto synchronisation, which aligns minority nodes through oscillatory consensus, and (iii) spectral embedding, which separates classes via structural regularisation. These perspectives are combined through class-adaptive ensemble weighting and trained with an imbalance-aware loss that couples balanced cross-entropy with physics-based constraints. Across five benchmark datasets and imbalance ratios from 5-100, PIMPC-GNN outperforms 16 state-of-the-art baselines, achieving notable gains in minority-class recall (up to +12.7%) and balanced accuracy (up to +8.3%). Beyond empirical improvements, the framework also provides interpretable insights into consensus dynamics in graph learning. The code is available at https://github.com/afofanah/PIMPC-GNN.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2602.01920")
get_code_for_paper("2602.01920")
have("2602.01920")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2602.01920.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2602.01920.svg)](https://syntology.ai/paper/2602.01920)

Connect an agent — have() is free.