PaperlayerView exact revision on arXiv

Source-linked AI summary

How Much Graph Is Too Much? Graph Construction and Exposure in Spatio-Temporal Forecasting

Saj Maru, Qiyu Han, Krishna Mamidala

arXiv:2610.34611v1stat.MLcs.LG

TL;DR

The paper asks whether graph neighbors add predictive information beyond each node’s own history, and whether forecast accuracy depends on neighbor choice or message integration. It tests these choices through conditional comparisons on three forecasting tasks. Late gated fusion has the lowest mean error among tested integration designs, while residual correlations have the lowest mean error among tested graphs under late gated integration.

  • Problem

    The paper asks whether graph neighbors provide predictive information beyond each node’s own history, and separates that question from how the temporal model uses their messages.

    Source: §1 · p. 1

  • Method

    It compares exposure designs with each domain graph fixed and graph constructions with late gated integration fixed, using a shared forecasting protocol.

    Source: §1 · p. 1, §2 · p. 2, §2.3 · p. 3

  • Results

    Across all three tasks, late gated fusion had the lowest mean error among tested designs; on electricity, it reached 5.27% WMAPE versus 9.21% without a gate.

    Source: §1 · p. 2, §4 · p. 6

  • Takeaways & Limitations

    The conditional comparisons support evaluating graph construction and integration together; randomized graphs performing similarly on unemployment limits evidence for a specific topology there.

    Source: §1 · p. 2, §4 · p. 6

    The comparisons cover selected construction–exposure combinations, and exposure variants also differ in computations and parameters, so differences cannot be attributed to timing alone.

    Source: §2.4 · p. 4

Abstract

from arXiv · show

Spatio-temporal graph neural networks combine a node's history with information from graph neighbors. The central question is whether graph neighbors provide predictive information beyond each node's own history. Forecast accuracy depends on both the choice of neighbors and the way graph messages enter the temporal model. We study these choices as graph construction and graph exposure on U.S. state unemployment, traffic speed, and electricity demand. Our diagnostic design uses two comparisons: varying exposure while holding the domain graph fixed, and varying construction while holding late gated integration fixed. A GRU using only each node's own history and randomized graph controls help assess what the graph contributes. Under a common change forecasting protocol, the same domain graph can improve or worsen error relative to the GRU across integration designs. Late integration alone is insufficient, while adding a fusion gate yields lower mean error in all three tasks. With late gated integration fixed, a graph based on residual correlations has the lowest mean error among the tested graphs, although randomized graphs perform similarly on unemployment. The findings support evaluating graph construction and integration together, with conclusions tied to the comparisons actually performed.

1 Introduction

The introduction asks whether graph neighbors add predictive information beyond each node’s own history, separating neighbor choice from how messages enter the temporal model. It reports that integration and topology both matter in the tested comparisons, with conclusions conditional on the design held fixed.

1 Introduction

  • Graph forecasting tests whether neighbors add information beyond each node’s own history, since a meaningful relation need not provide useful predictive information.The paper separates the choice of neighbors from how their information is used.Source: §1 · p. 1
  • Unemployment has a dominant common component explaining 93% of variation, while traffic and electricity retain more spatial dependence after common-component removal.Figure 1 reports PC1 variance share and the share of timesteps with significant residual Moran’s I.Source: §1 · p. 1, Figure 1 · p. 2
  • The diagnostic design varies exposure with each domain graph fixed and varies construction with late gated integration fixed, supporting conclusions conditional on the fixed choice.The comparisons distinguish neighbor selection from how the model uses graph messages.Source: §1 · p. 1
  • On electricity, late integration without a fusion gate has 9.21% WMAPE, versus 5.27% with the gate and 6.18% for the GRU baseline.The introduction concludes that delaying integration alone is insufficient in this setting.Source: §1 · p. 2
  • With late gated integration fixed, the residual correlation graph has the lowest mean error among tested graphs, while randomized graphs perform similarly on unemployment.The randomized-graph comparison limits evidence that gains establish the value of the intended topology.Source: §1 · p. 2

2 Problem Formulation and Diagnostic Design

The study separates graph construction from graph exposure and evaluates them with comparisons that hold one choice fixed. It uses a change-forecasting task to test whether graph messages add predictive information beyond each node’s history.

2.1 Forecasting Task and Training Protocol

  • The shared task predicts standardized changes, anchoring each forecast at the node’s last observation rather than targeting future levels directly.This targets future movement beyond the shared state already visible across nodes.Source: §2.1 · pp. 2–3

2.2 Graph Construction

  • Five construction rules compare domain structure, empirical correlation, and randomized topology, with graphs estimated from training data and held fixed thereafter.The residual-correlation graph removes the cross-node mean first difference at each time before selecting positively correlated neighbors.Source: §2.2 · p. 3, Table 1 · p. 3

2.3 Graph Exposure

  • Graph exposure specifies where messages enter the temporal model, how often graph operations occur, and whether a separate gate combines local and graph representations.The strategies range from mixing inputs before temporal encoding to applying graph operations after the input window has been encoded.Source: §2.3 · p. 3, Table 2 · p. 4
  • In late gated fusion, the temporal encoder represents each node’s own history before a graph representation is combined through a separate fusion gate.The gate controls local and graph contributions and is distinct from the GRU’s update and reset gates.Source: §2.3 · p. 3

2.4 Diagnostic Comparisons

  • The diagnostic design varies exposure with the domain graph fixed, then varies construction with late gated fusion fixed.These comparisons support conclusions conditional on the choice held fixed.Source: §2 · p. 2, §2.4 · pp. 3–4
  • The comparisons cover selected construction–exposure combinations, not every graph under every integration strategy.Exposure variants also differ in computations and parameters, so their differences cannot be attributed to timing alone.Source: §2.4 · p. 4

3 Experiments

Across unemployment, traffic, and electricity, the experiments compare exposure with the domain graph fixed and construction under late gated fusion. Residual-correlation graphs performed best on traffic and electricity, while exposure results varied with integration design.

3.1 Setup

  • The experiments span common-factor unemployment, road-network traffic, and grid- and weather-shaped electricity demand, using chronological 70/15/15 splits.The study reports WMAPE across datasets with different native units and uses a graph-free GRU as its primary baseline.Source: §3.1 · p. 4

3.2.1 Does Graph Exposure Matter?

  • Early and recurrent designs harmed traffic and electricity and were unstable on unemployment; on electricity, they exceeded 9% WMAPE versus 6.18% for the graph-free GRU.The same domain graph can yield different errors under different integration designs.Source: §3.2.1 · p. 4, Table 4 · p. 5
  • Late integration alone was insufficient: electricity WMAPE was 9.21% without the fusion gate, versus 5.27% with it and 6.18% for the GRU.On traffic, the corresponding errors were 7.96%, 7.64%, and 7.74%.Source: §3.2.1 · p. 4, Table 4 · p. 5

3.2.2 Does Graph Construction Matter?

  • 5.00% WMAPE on electricity and 6.87% on traffic: residual-correlation graphs beat domain graphs and GRU baselines, while unemployment graphs cluster within 1.23–1.27%.On unemployment, randomized graphs also performed comparably.Source: Table 5 · p. 5, §3.2.2 · p. 5
  • The study compares five graph-construction rules under late gated fusion and also evaluates reference methods with their own graph mechanisms and architectures.None of CaST, GAT, or AGCRN outperformed the late-gated residual-correlation recipe on any dataset.Source: §3.2.2 · pp. 4–5, Table 5 · p. 5

4 Conclusion and Future Extensions

Across three forecasting tasks, graph results depended on integration design, and residual-correlation graphs had the lowest mean error among tested graphs with late gated fusion fixed. The study’s scope is limited to three domains and intentionally simple diagnostic graph probes.

  • The comparisons vary exposure with each domain graph fixed and vary construction with late gated fusion fixed, so conclusions remain conditional on those choices.Source: §4 · p. 5
  • Late gated fusion had the lowest mean error among tested integration designs, while residual-correlation graphs had the lowest mean error among tested graphs with late gated fusion fixed.Source: §4 · p. 6
  • Randomized graphs performed similarly on unemployment, so those results provide limited evidence for the value of a specific topology.Source: §4 · p. 6
  • The study covers three domains and uses intentionally simple diagnostic graph probes, leaving other relational mechanisms and richer graph constructions for future work.Proposed extensions include adaptive, lagged, nonlinear, or exogenous-feature-aware graphs, and settings where the relevant graph changes over time.Source: §4 · p. 6

Related papers

Loading 2610.34611v1…