Carlo Bertrand, Daniel, Wei- Hua Suvash, Yan
We have not lifted any functions out of this paper's repositories yet, so there is nothing we have run. We have no repository linked to it either. That is a statement about our graph, not about the paper.
Graph Neural Networks (GNNs) have emerged as the primary tool for learning on graphstructured data. Yet, many GNNs are primarily suited for toy academic benchmarks, often limiting their efficacy in real-world scenarios. This thesis proposes novel approaches to enable the use of GNNs in real-world applications, including but not limited to social networks and recommender systems. Our contributions are as follows: • Scalability: We introduce Scalable Inception Graph Neural Networks (SIGN) (Chapter 3,[202]). By decoupling graph propagation from model learning, SIGN obtains comparable accuracy to previous models, but with up to 30x faster inference, enabling efficient GNNs on web-scale graphs. • Temporality: We propose Temporal Graph Networks (TGN) (Chapter 4, [200]), a comprehensive framework for learning on evolving graphs. Representing a temporal graph as a series of events, TGN dynamically updates node representations through event-driven messages prior to their aggregation across the graph. TGN outperforms previous methods on both temporal node-classification and link-prediction tasks. • Edge Directionality: Our research (Chapter 5, [201]) shows the importance of leveraging edge directionality, especially in heterophilic tasks. We introduce Directed Graph Neural Networks (Dir-GNN), a general framework to extend any spatial GNN to directed graphs. Incorporating edge directionality with Dir-GNN leads to large accuracy gains in heterophilic scenarios, while leaving performance unchanged on homophilic graphs. • Missing Data: Addressing data incompleteness, Feature Propagation (FP) (Chapter 6, [203]) imputes missing node attributes by propagating existing ones through the graph. Paired with a downstream GNN, FP mitigates accuracy degradation to about 4% even when 99% of the features are absent, vastly outperforming previous methods. • Structural Inference: We propose NuGget (Chapter 7, [205]), a method to reconstruct unknown graphs in a game-theoretic setup. NuGget recovers the network structure of a game from players' equilibrium actions, outperforming previous models without relying on explicit utility functions. v First and foremost, I would like to express my gratitude to Michael: not only for creating the opportunity to pursue a PhD under your supervision while we were both at Twitter, but also for the support and trust you demonstrated throughout the years. Your approach, granting me the freedom to navigate my own path and learn through experience, all while ensuring your timely presence when necessary, was invaluable. I happily remember the whiteboard brainstorming sessions in the first year of my PhD, as well as the long conversations about research over lunch. I hold a special place in my heart for my former team at Fabula, and subsequently at Twitter-Fabrizio, Davide, Federico, and Ben. The countless days we spent in the office, sharing insightful discussions, long lunches and intense table tennis matches, were both fun and a source of continuous learning. Each one of you has been a special collaborator and friend. I owe particular gratitude to Fabrizio, my companion on this PhD journey: your constant presence and collaborative spirit, despite our research diverging, was of great support. I'm grateful our paths reconnected professionally on my latest paper, marking a fruitful reunion after three years. To Nils, my manager at Twitter: thank you for having been an emphatic mentor as much as a manager. You showed me the path to thriving in professional relationships and aligning my research with Twitter's strategic initiatives. To my other collaborators from all over the world-
The same record, over MCP at https://syntology.ai/mcp:
get_harvested_code_for_paper("2510.21994")
get_code_for_paper("2510.21994")
have("2510.21994")
The run record, dated, one paper per request, free:
curl https://syntology.ai/api/ran/2510.21994.json
A badge for a README (the split and the date, never a ratio):
[](https://syntology.ai/paper/2510.21994)
Connect an agent — have() is free.