Source-linked AI summary
Bayesian Evidence Accumulation for Graph-Based Learning in Dynamic Data-Driven Applications Systems
Moises Ponce, Sarah Harkins Dayton, Ioannis Schizas, Vasileios Maroulas
TL;DR
GNNs generally produce a single output, while DDDAS settings require decisions as observations stream in under uncertainty. B-EAGLe combines graph encoding with Bayesian evidence accumulation, updating class beliefs and deferring commitment until confidence is sufficient. It exceeds baselines on three of four tested datasets, with an approximately 13-point improvement on CIFAR10-SP under a 50-view observation cap.
Problem
GNN decision-making generally produces a single output from information available at inference time, whereas DDDAS requires decisions as observations arrive sequentially under uncertainty.
Source: §1 · p. 1Method
B-EAGLe combines graph observations with a Bayesian filter that updates posterior class beliefs sequentially and defers decisions until sufficient confidence is reached.
Source: §1 · p. 2, §2.2 · p. 4, §3.2 · p. 5Results
Takeaways & Limitations
B-EAGLe realizes the data-driven observation-to-decision direction of DDDAS feedback, while observation acquisition is not yet steered by the model.
Source: §5 · p. 8The tested system realizes only one direction of the DDDAS feedback loop; it does not steer which observations are acquired next.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Graph Neural Networks (GNNs) are widely used in modern machine learning, yet most existing frameworks remain fundamentally static, limiting their applicability in Dynamic Data Driven Application Systems (DDDAS) where observations arrive sequentially under uncertainty. We introduce B-EAGLe (Bayesian Evidence Accumulation for Graph-Based Learning), a probabilistic framework inspired by CBGT-Net (cortico-basal ganglia-thalamic network) that reformulates graph classification as a sequential evidence accumulation process. B-EAGLe combines a GNN encoder with a Bayesian accumulator that incrementally updates posterior beliefs over class labels as new observations arrive, committing to a decision only when sufficient confidence is achieved. Experiments on CIFAR10 demonstrate competitive performance together with principled confidence estimation under sequential observations.
1 Introduction
Existing GNN decision-making typically produces one output from information available at inference time, while DDDAS requires decisions from sequential observations under uncertainty. B-EAGLe reframes graph classification as confidence-driven evidence accumulation, deferring decisions until posterior confidence is sufficient.
1 Introduction
- GNNs model graph classification, but their single-output decision paradigm assumes all relevant information is available at inference time.That assumption does not fit settings where observations arrive sequentially under uncertainty.Source: §1 · p. 1
- DDDAS incorporates streaming measurements at execution time and uses them to govern measurement and data acquisition.Applications named include unmanned vehicle control, sensor selection, and natural disaster forecasting.Source: §1 · p. 2
- Graph-structured data occurs in DDDAS settings, and GNNs could serve as a computational backbone if decision-making supports sequential streaming observations.The passage cites forecasting and skeleton-based action recognition as existing GNN applications.Source: §1 · p. 2
- B-EAGLe maintains posterior beliefs over class labels as graph observations arrive and defers decisions until a confidence threshold is reached.Its Bayesian belief update supports sequential, confidence-driven classification under DDDAS.Source: §1 · p. 2
2 Preliminaries and Related Works
The preliminaries define graph observations and describe how sequential graph classification fits the DDDAS feedback loop. The paper uses uniform node sampling and Bayesian accumulation to make decisions from streamed observations.
2.1 Graph Theory Notation
- An induced subgraph contains a selected node subset and all edges whose endpoints lie within it; B(v, L) denotes the L-hop neighborhood around v.For proof of concept, a node is selected uniformly at random; ideally DDDAS guides observations as decisions are made.Source: §2.1 · p. 2
2.2 DDDAS and Decision-Making
- At each timestep, policy π uniformly selects a node, and its L-hop neighborhood becomes the observation; this isolates accumulation gains from adaptive policy-making.The policy determines which region of the physical system is interrogated next.Source: §2.2 · p. 3
- The paper models DDDAS classification as feedback over sequential graph observations, with a fixed set of classes and a posterior belief state maintained over time.The graph stream acts as the sensor, π as sensor management, and the encoder plus Bayesian filter as the estimation algorithm; no physical system simulation is instantiated.Source: §2.2 · p. 3
- CBGT-Net treats input as a stream of observations, while its additive accumulator lacks a probabilistic interpretation; B-EAGLe replaces it with a Bayesian filter over posterior beliefs.The additive accumulator initializes at zero and adds each timestep’s evidence.Source: §2.2 · pp. 3–4
3 Problem Formulation and Methodology
B-EAGLe processes streamed graph neighborhoods through a GNN and Bayesian belief updates, then commits when a confidence threshold is reached. Its method assumes conditionally independent observations, an assumption that can be strained by overlapping neighborhoods.
3 Problem Formulation and Methodology
- Each streamed graph yields an L-hop induced subgraph around a selected node, serving as the observation passed into the model.This observation choice reduces the need to compute over whole graphs.Source: §3 · p. 4
- The architecture has three modules: a GNN encoder, a Bayesian Evidence Accumulation Module, and a Confidence Threshold Module.Source: §3 · p. 4
3.1 GNN Encoder Module
- The GNN encoder uses GAT-based layers [19] and GNN+ architectural choices [20], with graph-level mean pooling or node-wise readout before softmax classification.The encoder also integrates edge features, batch normalization, dropout, residual connections, and feed-forward transformations.Source: §3.1 · pp. 4–5
3.2 Bayesian Evidence Accumulation Module
- The Bayesian accumulator maintains class posteriors using conditionally independent observations and a uniform initial prior.Source: §3.2 · p. 5
- GNN softmax outputs are treated as normalized observation likelihoods, allowing sequential Bayesian filter inference.Source: §3.2 · p. 5
- The accumulator multiplies each observation's likelihood into the prior belief and normalizes the class probabilities at each update.A small ε prevents zero-probability terms.Source: §3.2 · pp. 5–6
- Overlapping observations can reuse evidence, making posterior beliefs overconfident and potentially satisfying the stopping criterion too early.A stringent threshold can offset this effect, but an overly stringent threshold may increase timeouts.Source: §3.2 · p. 6
3.3 Confidence Threshold Module
- The model stops at the first timestep when a class belief meets the confidence threshold and commits to a classification.Until then, it continues collecting evidence; the decision is bounded by Tmax.Source: §3.3 · p. 6
4 Case Study and Results
Experiments evaluate B-EAGLe on super-pixel graph-classification and stochastic-block-model node-classification datasets, using neighborhood observations and a fixed confidence threshold. B-EAGLe outperformed baselines on three of four datasets, while the τ = 0.99 threshold produced near-total timeouts on node-classification tasks.
4 Case Study and Results
- The evaluation used MNIST-SP and CIFAR10-SP for graph classification and PATTERN and CLUSTER for node classification.The datasets are Super-pixel and Stochastic Block Model datasets, respectively.Source: §4 · p. 6
- Graph-classification observations used 3-hop neighborhoods, while node-classification observations used 1-hop neighborhoods.The settings capture broader structural context for graph classification and local structure for node classification.Source: §4 · p. 7
- Experiments used τ = 0.99 and Tmax = 50; episodes that did not reach the threshold were classified by the final-timestep arg max.The threshold triggers a decision when any class posterior reaches or exceeds 99%.Source: §4 · p. 7, Table 1 · p. 7
- B-EAGLe outperformed state-of-the-art baselines on three of the four datasets.The comparison is reported in Table 2.Source: §4 · p. 7, Table 1 · p. 7
- Node-classification datasets had timeout percentages near 100%, suggesting that τ = 0.99 can be too stringent.Graph-classification datasets did not have high timeout rates at this threshold.Source: §4 · p. 7
5 Conclusion and Future Work
B-EAGLe exceeded baselines on three of four tested datasets, including an approximately 13-point improvement on CIFAR10-SP under a 50-view cap. The authors note that the DDDAS feedback loop is only partly realized and identify live-system evaluation and other analyses as future work.
5 Conclusion and Future Work
- Approximately 13 percentage points of improvement on CIFAR10-SP was achieved with an observation budget capped at 50 views per graph.Table 2 reports classification accuracy and observation and timeout statistics.Source: §5 · p. 7, Table 2 · p. 8
- PATTERN improved by almost 7 percentage points and CLUSTER by 2.The seed-to-seed variance on CIFAR10-SP was higher than the baseline and was potentially attributable to predicted-class variance rather than the accumulation module.Source: §5 · p. 8
- B-EAGLe exceeded baselines on three of four tested datasets, but only one direction of the DDDAS feedback loop was fully realized.A complete system would also steer which observations are acquired next.Source: §5 · p. 8
- Future work includes live instrumentation evaluation of decision latency and improving data selection during low-certainty episodes.The authors also propose threshold sensitivity or a dynamic threshold, belief calibration, and matched-observation-budget ensemble baselines.Source: §5 · p. 8