Deep Learning on Real-World Graphs

AI summary

Overview

Research area: Graph machine learning — specifically Graph Neural Networks (GNNs) applied to real-world, large-scale graph-structured data. This is a doctoral thesis (single author, Emanuele Rossi), so it aggregates several related model contributions rather than a single result.

Technical level: Advanced. The work assumes familiarity with GNNs, message passing, and the standard benchmark setting that the thesis argues is insufficient for industrial use.

Scope: The thesis proposes a set of GNN models that each target one specific gap between academic graph benchmarks and real-world graph systems — scalability, temporality, directionality, missing features, and structural uncertainty.

What This Paper Is About

GNNs work well on curated benchmark graphs, but real systems rarely look like those benchmarks. They are huge, they change over time, their edges often point in a meaningful direction, their node attributes are frequently absent or incomplete, and the graph structure itself is only partially known or uncertain.

The thesis's goal is to close that gap by introducing one model per limitation, so that GNNs become usable on the kinds of graphs that social and recommendation platforms actually operate.

Key Contributions

1. SIGN — a model aimed at scalable graph learning, addressing the cost of applying GNNs to very large graphs.
2. TGN — a model for temporal graphs, addressing the fact that real graphs evolve over time rather than sitting static.
3. Dir-GNN — a model that handles directed edges explicitly, including settings where the network is heterophilic (connected nodes tend to differ rather than resemble each other).
4. Feature Propagation (FP) — a method for learning when node features are missing, and NuGget, a game-theoretic approach to inferring structure when that structure is uncertain.

Main Findings

• The gap is the problem being named: the abstract's central claim is that scalability, temporality, directionality, data incompleteness, and structural uncertainty are the blockers keeping GNNs out of real-world deployment.
• One model per limitation: each challenge is matched to a dedicated contribution rather than a single general-purpose fix.
• Feasibility at industrial scale: the thesis claims these contributions "bridge the gap between academic benchmarks and industrial-scale graphs," enabling GNN use in domains such as social and recommender systems.
• No quantitative results in the abstract: the abstract reports no accuracy figures, speedups, dataset sizes, or comparisons against baselines, so the strength of each claim cannot be assessed from it. The details are not available here.

Methodology in Plain English

The approach is a portfolio strategy. Rather than proposing one architecture and arguing it handles everything, the author isolates five distinct ways real-world graphs break the standard assumptions, then designs a targeted solution for each:

• For size, restructure how computation is done so the model does not need to repeatedly traverse the whole graph.
• For time, treat the graph as a stream of events and model how nodes and edges change, instead of freezing it into a snapshot.
• For direction, let the model treat an incoming edge differently from an outgoing one, which also helps when neighbors are dissimilar.
• For missing features, propagate known feature information across the graph to reconstruct what is absent, instead of discarding nodes that lack attributes.
• For uncertain structure, use a game-theoretic formulation to infer what the underlying structure likely is.

The unifying thread is that each method relaxes one unrealistic assumption baked into standard GNN benchmarks.

Why This Matters

Research impact: The thesis reframes GNN evaluation around messy, real conditions rather than clean benchmarks, arguing that the field's standard testbed hides the obstacles that matter in practice. It also treats directionality, missing features, and structural uncertainty as first-class modeling problems rather than preprocessing afterthoughts.

Real-world applications the abstract points to:

• Social systems — networks that are large, constantly changing, and only partially observed.
• Recommender systems — user–item graphs where interactions are directional and time-stamped, and where many users have sparse or absent profile features.
• Applications implied by the stated challenge areas but not named in the abstract: any large directed network, and any setting where records are incomplete or the graph must be inferred rather than read off directly.

Industry relevance: The five limitations listed in the abstract are exactly the conditions on real production graphs, which is why the thesis frames its value in terms of industrial-scale deployment rather than benchmark scores.

Future Directions

• Reconciling the five solutions: the abstract presents separate models; whether they can be combined into one architecture that handles size, time, direction, missing features, and uncertainty simultaneously is left open.
• Evaluation beyond benchmarks: the thesis argues existing benchmarks are inadequate, which raises the question of what a realistic, reproducible evaluation standard for industrial-scale graphs would look like.
• Handling all limitations at once: real graphs typically exhibit several of these problems together, so robustness under combined conditions is a natural next step.
• Extending structural inference: NuGget's game-theoretic view of structure suggests further work on when structure can be recovered reliably versus when it must remain uncertain.

Target Audience

Researchers and graduate students in graph machine learning, and practitioners — particularly ML engineers working on social or recommendation systems — who need GNNs to function on large, dynamic, partially observed graphs. Readers looking for a broad survey or for concrete benchmark numbers will not find them in this abstract; it best serves those already comfortable with GNN fundamentals who want a structured account of the field's real-world limitations.

Authors’ abstract

Graph Neural Networks (GNNs) have become a central tool for learning on graph-structured data, yet their applicability to real-world systems remains limited by key challenges such as scalability, temporality, directionality, data incompleteness, and structural uncertainty. This thesis introduces a series of models addressing these limitations: SIGN for scalable graph learning, TGN for temporal graphs, Dir-GNN for directed and heterophilic networks, Feature Propagation (FP) for learning with missing node features, and NuGget for game-theoretic structural inference. Together, these contributions bridge the gap between academic benchmarks and industrial-scale graphs, enabling the use of GNNs in domains such as social and recommender systems.