Learning on Graphs Conference, 2026
Organizers: Saurabh Sihag, Andrea Cavallo, Gonzalo Mateos, Elvin Isufi, Alejandro Ribeiro
Date: November 20, 2026
| US West Coast | Boston (EST) | Europe (London) | Asia (Beijing) |
|---|---|---|---|
| 12:30 | 15:30 | 20:30 | 04:30 (Nov 21) |
Length: 2.5 hours
Abstract:
Covariance matrices capture the redundancies between the features of a dataset: brain regions that activate together, stock prices that fluctuate in tandem, traffic volumes that are correlated across intersections. They are a cornerstone of multivariate data analysis, most prominently through principal component analysis (PCA). Yet PCA is a fixed, task-agnostic linear transformation, and learning from sample covariance matrices raises four key challenges: how to use covariance information for representation learning on a downstream task; whether inference is stable to the finite-sample perturbations of the estimated covariance; whether models are transferable across datasets that describe the same phenomenon with a different number of features; and how to make learning interpretable.
This tutorial treats the covariance matrix as a relational inductive bias, a data-driven graph, and introduces coVariance neural networks (VNNs): graph neural networks that operate on sample covariance matrices. Starting from the connection between PCA and the graph Fourier transform, we show that coVariance filters implicitly perform PCA-based inference without computing eigenvectors, and that, unlike PCA, VNNs are provably stable to covariance estimation errors and transferable across scales. We then show how these properties enable principled, anatomically interpretable brain age gap prediction in network neuroscience, and we illustrate how using pairwise statistics as inductive biases underpins a broader set of data processing methods beyond VNNs, including sparse VNNs, precision neural networks and covariance scattering transforms.
Website: https://learningwithcovariances.github.io/
Setup requirements: The hands-on part is a Colab notebook accessible at the following link: https://colab.research.google.com/drive/19EtO9WTkyehwlB4vh4HZFMQqBrXvwBKb?usp=sharing. It will not be discussed during the tutorial timeslot but is intended as additional material. To use it: 1. Open the notebook with the button below. 2. Create your own copy: in Colab, choose File → Save a copy in Drive (requires a Google account). 3. Run your copy: choose Runtime → Run all, or run the cells one by one.
Organizers: Franco Terranova, Victor-Alexandru Darvariu
Date: November 21, 2026
| US West Coast | Boston (EST) | Europe (London) | Asia (Beijing) |
|---|---|---|---|
| 12:30 | 15:30 | 20:30 | 04:30 (Nov 22) |
Length: 2.5 hours
Abstract:
Graph-based combinatorial optimization (GCO) problems, which involve finding a discrete structure that optimizes a given objective function, play a critical role in many real-world systems. Their applications span classic benchmark problems as well as a wide range of real-world applications. However, as graph sizes grow, exact optimization quickly becomes computationally prohibitive, motivating the use of heuristic and approximation methods that often require substantial problem-specific expertise.
Machine learning (ML) methods have steadily gained popularity for solving GCO problems. In particular, reinforcement learning (RL) has demonstrated the ability to automatically learn, using a reward signal, heuristics for constructing GCO solutions. Such approaches can achieve competitive solution quality with fast inference, while automating the complex trial-and-error process of heuristic design and retaining applicability across a wide range of problems. However, applying RL to GCO problems is technically complex. It requires understanding concepts from traditionally disjoint communities (RL, graph learning, GCO) that are presented in fragmented resources. This is the gap our tutorial aims to fill.
The first half of our tutorial begins by introducing GCO problems, traditional methods for solving them, and modern ML techniques. We subsequently introduce the Markov Decision Process (MDP) formalism and show how to formulate GCO problems as MDPs. Next, we introduce deep RL and graph representation learning methods and discuss their integration into GCO pipelines. We discuss evaluation methodologies and key challenges, including scalability and generalization. Lastly, we cover applications of RL for GCO, with a particular focus on real-world applications.
Website: https://victor.darvariu.me/log2026.html
Tutorial slides: Google Slides
Setup requirements: The repository for the practical part of the tutorial is https://github.com/terranovafr/RLforGCO-LoG2026. Prior to the tutorial, participants should follow the setup instructions in the README file of this repository to install Docker and locally build the container with the appropriate dependencies. Please note that this requires downloading about 2GB of data, which may be challenging to do during the live tutorial over conference WiFi. Therefore, we advise participants to complete this step before attending.
Organizers: Zixing Song
Date: November 22, 2026
| US West Coast | Boston (EST) | Europe (London) | Asia (Beijing) |
|---|---|---|---|
| 10:00 | 13:00 | 18:00 | 02:00 (Nov 23) |
Length: 1.5 hours
Abstract:
Large language models are increasingly deployed not as isolated models, but as multi-agent systems (MAS), where heterogeneous agents communicate, specialize, debate, verify, and coordinate toward a shared objective. In these systems, performance is determined not only by the capabilities of individual agents, but also by how agents are connected and how information flows among them. This makes graph structure a first-class object for learning and optimization.
This tutorial develops a graph-centric view of LLM-based multi-agent systems. We model agents as nodes, communication channels as directed edges, messages as edge-associated features, and multi-round interactions as dynamic graphs. We then survey the emerging progression from manually designed topologies to graph sparsification with learned adjacency matrices. Representative methods include agent communication pruning, GNN-based topology design, and autoregressive generation of task-specific agent organizations.
The tutorial emphasizes the connection between these emerging LLM-based multi-agent systems and core topics in graph machine learning: message passing, graph structure learning, sparsification, graph generation, dynamic graphs, and heterogeneous graphs. A hands-on component will let participants optimize a small multi-agent communication graph under performance and communication-cost constraints. We conclude with open problems toward scalable, adaptive, and trustworthy graph structures for LLM-based multi-agent systems.
Website: TBD
Setup requirements: TBD