fabriziof (at) campus.technion.ac.il
Talks
Toward Reliable Language Models: Structure-Aware Learning on Computational Traces
Invited Talk
CIRM Workhops – Machine Learning and Signal Processing on Graphs · 2026/09
About this talk
LLM failure detectors mostly rely on simplistic linear models and handcrafted heuristics applied to hand-picked fragments of model internals. We argue that these 'computational traces' — such as activations, next-token distributions, and attention maps — are structured data modalities that deserve being processed by architectures with inductive biase matched to their structure, in the spirit of Geometric Deep Learning. In this talk I present a series of five works on hallucination detection, where we describe architectures to optimally process these computational traces, with a focus on employing attention-induced graphs and neural message-passing thereon as a flexibile, unifying approach. The resulting monitors are lightweight, outperform probing and heuristic baselines, and show promising transfer capabilities across tasks and LLMs.
Works covered
- Beyond Next Token Probabilities: Learnable, Fast Detection of Hallucinations and Data Contamination on LLM Output Distributions
- Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT
- Neural Message-Passing on Attention Graphs for Hallucination Detection
- LEMON: Equivariant Learning over Attention Maps for Zero-Shot Misbehavior Monitoring on New LLMs
- UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations
Structure-Aware Learning: From Graph Neural Networks to reliable Foundation Models
Invited Talk
ELLIS Institute Tübingen Scientific Symposium 2026 · 2026/03
About this talk
In this talk I encompass salient papers from my PhD through my Postdoc, and present a series of works on principled Geometric Deep Learning approaches for learning on structured data, built, in particular, around the design patterns of symmetry-awareness, expressiveness, and scalability. I start with equivariant graph architectures that provably overcome the expressiveness limits of message-passing while staying efficient, and then I extend the same principles to learning on LLM computational traces (such as attention maps and activations) to detect failure modes such as hallucinations. I close with the prospect for a broader research programme spanning foundation model monitoring systems bridging language modelling and relational learning, and structure-aware architectures for genomics.
Graph Learning, Equivariance and Expressiveness
Lecture
Course on Groups and Deep Learning at Technion · 2025/07
About this talk
In this lecture, I discuss learning on graphs and its key challenges, focusing on the interplay between equivariance, expressive power, and computational complexity. I present a coherent narrative that introduces these core aspects and highlights recent research aimed at navigating the trade-offs between them. We begin with feature augmentation techniques, including random feature identifiers, symmetrization methods, and substructure encodings. The lecture then transitions to the broader class of Subgraph GNNs, which aim to combine the strengths of these techniques while maintaining a domain-agnostic design.
Advances in Subgraph GNNs for Expressive and Efficient Learning on Graphs
Keynote
LoG 2024 Italian Meetup (Keynote Speaker) · 2024/12
About this talk
In this keynote talk, I discuss recent advances in Graph Representation Learning, with a focus on Subgraph Graph Neural Networks (Subgraph GNNs) as more expressive alternatives to traditional message-passing approaches. I begin by exploring the concept of expressive power and the limitations of conventional methods before introducing Subgraph GNNs. While these models offer greater flexibility and expressiveness, they come with the challenge of high computational complexity due to modeling graphs as bags of subgraphs. At the core of the talk, I present HyMN, a novel Subgraph GNN that balances efficiency and expressiveness through subgraph sampling and Structural Encodings (SEs). By leveraging walk-based node centrality measures, HyMN effectively reduces subgraph sizes while enhancing discriminative power and maintaining computational feasibility.
The future and challenges of Graph Learning
Panel
Graph Learning Social @ ICML 2024 (Panelist) · 2024/07
About this talk
I took part as a panelist at ICML 2024 Graph Learning Social's panel to discuss on the future and challenges related to Machine Learning on graphs.
Towards Expressive and Efficient Graph Neural Networks
Prof. Thomas Gärtner's ML research unit at TU Wien · 2024/03
About this talk
A presentation on the main work underpinning my PhD thesis. We discuss jointly, in a coherent and gradual fashion, the contributions I made on designing expressive and efficient Graph Neural Networks. We begin by exploring neural architectures on simplicial complexes, a generalization of graphs that captures group-wise interactions. By lifting graphs to simplicial complexes via complete subgraphs, we introduce a hierarchical message-passing approach that surpasses traditional node-centric MPNNs. Extending this, we consider regular cell complexes, incorporating structures like simple and induced cycles for enhanced expressiveness and performance, particularly in molecular modeling. Next, we present an alternative approach that models graphs as bags of subgraphs selected by predefined policies. We design an architecture that respects symmetries, achieving high expressiveness even with efficient, weak message-passing encoders. Finally, we provide a theoretical framework, unifying related works under the bags-of-subgraphs paradigm. A novel symmetry analysis offers an upper bound on expressiveness and defines a design space for future architectures.
Pack your subgraphs: A journey into subgraphs for powerful Graph Neural Networks
Yves-Alexandre de Montjoye's Group Talks (Imperial College London) · 2023/03
About this talk
Our latest two works on subgraphs for more expressive GNNs lined up to tell a coherent story. In the first part of the talk we show the design of a novel framework (ESAN) to process generic bags of subgraphs in an equivariant manner. We show that this approach effectively increases the expressive power of MPNNs. Then, we notice a surge in concurrent approaches which – sometimes unwittingly – make use of subgraphs for powerful graph representations. In the second part, a novel symmetry analysis allows to unify and better characterise these approaches when using a node-based selection policy. We prove an upper-bound on their expressive power and conceive a framework serving as a design space for equivariant node-based subgraph methods.
Pack your subgraphs: A journey into subgraphs for powerful Graph Neural Networks
Lecture
Course on Geometric Deep Learning at Oxford University · 2023/03
About this talk
Our latest two works on subgraphs for more expressive GNNs lined up to tell a coherent story. In the first part of the talk we show the design of a novel framework (ESAN) to process generic bags of subgraphs in an equivariant manner. We show that this approach effectively increases the expressive power of MPNNs. Then, we notice a surge in concurrent approaches which – sometimes unwittingly – make use of subgraphs for powerful graph representations. In the second part, a novel symmetry analysis allows to unify and better characterise these approaches when using a node-based selection policy. We prove an upper-bound on their expressive power and conceive a framework serving as a design space for equivariant node-based subgraph methods.
Exploring the practical and theoretical landscape of expressive Graph Neural Networks
Tutorial
Learning on Graphs Conference 2022 · 2022/12
Co-presented with: Beatrice Bevilacqua, Haggai Maron
About this talk
The tutorial reviews the most prominent expressive GNNs, categorises them into different families, and draws interesting connections between them. This is accomplished through a series of practical coding sessions and an organic overview of the literature landscape. We aim to convey the importance of studying the expressive power of GNNs and make this field more accessible to our community, especially practitioners and newcomers.
Understanding and Extending Subgraph GNNs by Rethinking their Symmetries
Learning on Graphs and Geometry Reading Group · 2022/11
Co-presented with: Beatrice Bevilacqua
About this talk
Many concurrent works seem to share a similar underlying intuition: to represent a graph by processing subgraphs generated via some simple selection policy. Inspired by a novel symmetry analyses for the most prominent of these methods – those where subgraphs are in a 1:1 correspondence with nodes in the original graph – we prove an upper-bound on their expressive power and conceive a framework serving as a design space for future new equivariant node-based subgraph architectures.
Pack your subgraphs: A journey into subgraphs for powerful Graph Neural Networks
Simone Scardapane's Group Talks (Sapienza) · 2022/11
Co-presented with: Beatrice Bevilacqua
About this talk
We line up our latest two works on subgraphs for more expressive GNNs to tell a coherent story. In the first part of the talk we show the design of a novel framework (ESAN) to process generic bags of subgraphs in an equivariant manner. We show that this approach effectively increases the expressive power of MPNNs. Then, we notice a surge in concurrent approaches which – sometimes unwittingly – make use of subgraphs for powerful graph representations. In the second part, a novel symmetry analysis allows to unify and better characterise these approaches when using a node-based selection policy. We prove an upper-bound on their expressive power and conceive a framework serving as a design space for equivariant node-based subgraph methods.
Pack your subgraphs: A journey into subgraphs for powerful Graph Neural Networks
Meta AI orgs Reading Meeting · 2022/07
Co-presented with: Beatrice Bevilacqua
About this talk
We line up our latest two works on subgraphs for more expressive GNNs to tell a coherent story. In the first part of the talk we show the design of a novel framework (ESAN) to process generic bags of subgraphs in an equivariant manner. We show that this approach effectively increases the expressive power of MPNNs. Then, we notice a surge in concurrent approaches which – sometimes unwittingly – make use of subgraphs for powerful graph representations. In the second part, a novel symmetry analysis allows to unify and better characterise these approaches when using a node-based selection policy. We prove an upper-bound on their expressive power and conceive a framework serving as a design space for equivariant node-based subgraph methods.
Pack your subgraphs: A journey into subgraphs for powerful Graph Neural Networks
Maks Ovsjanikov's Group Talks (École Polytechnique, France) · 2022/07
Co-presented with: Beatrice Bevilacqua
About this talk
We line up our latest two works on subgraphs for more expressive GNNs to tell a coherent story. In the first part of the talk we show the design of a novel framework (ESAN) to process generic bags of subgraphs in an equivariant manner. We show that this approach effectively increases the expressive power of MPNNs. Then, we notice a surge in concurrent approaches which – sometimes unwittingly – make use of subgraphs for powerful graph representations. In the second part, a novel symmetry analysis allows to unify and better characterise these approaches when using a node-based selection policy. We prove an upper-bound on their expressive power and conceive a framework serving as a design space for equivariant node-based subgraph methods.
Subgraphs for more expressive GNNs
Seminar
Course on Geometric Deep Learning at the African Master's on Machine Intelligence (GDL100) · 2022/07
About this talk
Many concurrent works seem to share a similar underlying intuition: to represent a graph by processing subgraphs generated via some simple selection policy. We discuss these methods together and try to find a common characterisation, while showing that the architecture proposed in our previous work 'Equivariant Subgraph Aggregation Networks' subsumes most of them.
Works covered
Graph Representation Learning on Simplicial and Cellular Complexes
Dagstuhl Seminars – Graph Embeddings: Theory meets Practice · 2022/03
About this talk
I gather together our works on message-passing on Simplicial and Cellular Complexes and present them together showing how these approaches in turn allow for more expressive graph representations.
Equivariant Subgraph Aggregation Networks
Author Interviews with Zak Jost · 2022/02
Co-presented with: Derek Lim, Beatrice Bevilacqua
About this talk
We explore the idea of modeling a graph as a bag of subgraphs generated by simple, domain agnostic selection policies, such as node-deletion, edge-deletion, ego-networks. We formally show that this approach lead to the design Graph Neural Networks which can be made strictly more powerful than standard Message Passing ones. We prove this by defining a novel WL variant and by constructing layers equivariant to the emerging symmetry group.
Works covered
Subgraph Networks
Third Nepal Winter School in AI · 2021/12
About this talk
Many concurrent works seem to share a similar underlying intuition: to represent a graph by processing subgraphs generated via some simple selection policy. We discuss these methods together and try to find a common characterisation, while showing that the architecture proposed in our previous work 'Equivariant Subgraph Aggregation Networks' subsumes most of them.
Works covered
Equivariant Subgraph Aggregation Networks
Learning on Graphs and Geometry Reading Group · 2021/12
Co-presented with: Derek Lim, Beatrice Bevilacqua
About this talk
We explore the idea of modeling a graph as a bag of subgraphs generated by simple, domain agnostic selection policies, such as node-deletion, edge-deletion, ego-networks. We formally show that this approach lead to the design Graph Neural Networks which can be made strictly more powerful than standard Message Passing ones. We prove this by defining a novel WL variant and by constructing layers equivariant to the emerging symmetry group.
Works covered
Anees Kazi on 'Graph Convolutional Network for Disease Prediction with Imbalanced Data'
Moderator
London Machine Learning Meetup · 2021/10
About this talk
I moderate Anees Kazi's talk at the London Machine Learning Meetup.
Weisfeiler and Lehman Go Cellular: CW Networks
Learning on Graphs and Geometry Reading Group · 2021/08
Co-presented with: Cristian Bodnar
About this talk
We extend message-passing from graphs and Simplicial Complexes to Cellular Complexes, combinatorial topological spaces generalising these two. By lifting graphs to Cellular Complexes we can naturally model molecules and unlock a series of advantages w.r.t. standard pairwise message-passing.
Works covered
Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
Seminar
Course on Geometric Deep Learning at the African Master's on Machine Intelligence (GDL100) · 2021/08
About this talk
We re-define colour refinement and message-passing on Simplicial Complexes, topological generalisation of graphs. This allows to model higher-order interactions, model meso-scale structures and increase the expressive power of Graph Neural Networks.
The expressive power of GNNs by the WL test
Tutorial
London Geometry and Machine Learning Summer School 2021 · 2021/07
Co-presented with: Giorgos Bouritsas
About this talk
What is the current state of affair in provably more expressive Graph Neural Networks? Giorgos Bouritsas and I lead and moderate this interactive discussion on approaches to go beyond 1-WL discriminating power. Ultimately, we would like to tackle the following question: 'Is the Weisfeiler-Leman hierarchy suitable to study the representational power of learning models on graphs? Do we necessarily need to go beyond message-passing?'
Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
TopoNets 2021 – Networks Beyond Pairwise Interactions · 2021/06
Co-presented with: Cristian Bodnar
About this talk
We re-define colour refinement and message-passing on Simplicial Complexes, topological generalisation of graphs. This allows to model higher-order interactions, model meso-scale structures and increase the expressive power of Graph Neural Networks.
Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
Artificial Intelligence Research Group Talks (Cambridge University) · 2021/05
Co-presented with: Cristian Bodnar, Yu Guang Wang
About this talk
We re-define colour refinement and message-passing on Simplicial Complexes, topological generalisation of graphs. This allows to model higher-order interactions, model meso-scale structures and increase the expressive power of Graph Neural Networks.
Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
Graph Journal Club (Valence AI) · 2021/04
Co-presented with: Cristian Bodnar, Yu Guang Wang
About this talk
We re-define colour refinement and message-passing on Simplicial Complexes, topological generalisation of graphs. This allows to model higher-order interactions, model meso-scale structures and increase the expressive power of Graph Neural Networks.
Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
Math Machine Learning seminar MPI MIS + UCLA · 2021/04
Co-presented with: Cristian Bodnar, Yu Guang Wang
About this talk
We re-define colour refinement and message-passing on Simplicial Complexes, topological generalisation of graphs. This allows to model higher-order interactions, model meso-scale structures and increase the expressive power of Graph Neural Networks.
Improving Graph Neural Network Expressivity via Subgraph Isomorphism Counting
Graph Journal Club (Valence AI) · 2021/04
Co-presented with: Giorgos Bouritsas
About this talk
We study how isomorphism counting of small substructures can improve the expressive power of Message Passing Neural Networks, while retaining equivariance, a good inductive bias, and their tractable fwd-pass computational complexity.
Toward Reliable Language Models: Structure-Aware Learning on Computational Traces
CIRM Workhops – Machine Learning and Signal Processing on Graphs · 2026/09
About this talk
LLM failure detectors mostly rely on simplistic linear models and handcrafted heuristics applied to hand-picked fragments of model internals. We argue that these 'computational traces' — such as activations, next-token distributions, and attention maps — are structured data modalities that deserve being processed by architectures with inductive biase matched to their structure, in the spirit of Geometric Deep Learning. In this talk I present a series of five works on hallucination detection, where we describe architectures to optimally process these computational traces, with a focus on employing attention-induced graphs and neural message-passing thereon as a flexibile, unifying approach. The resulting monitors are lightweight, outperform probing and heuristic baselines, and show promising transfer capabilities across tasks and LLMs.
Works covered
- Beyond Next Token Probabilities: Learnable, Fast Detection of Hallucinations and Data Contamination on LLM Output Distributions
- Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT
- Neural Message-Passing on Attention Graphs for Hallucination Detection
- LEMON: Equivariant Learning over Attention Maps for Zero-Shot Misbehavior Monitoring on New LLMs
- UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations
Structure-Aware Learning: From Graph Neural Networks to reliable Foundation Models
ELLIS Institute Tübingen Scientific Symposium 2026 · 2026/03
About this talk
In this talk I encompass salient papers from my PhD through my Postdoc, and present a series of works on principled Geometric Deep Learning approaches for learning on structured data, built, in particular, around the design patterns of symmetry-awareness, expressiveness, and scalability. I start with equivariant graph architectures that provably overcome the expressiveness limits of message-passing while staying efficient, and then I extend the same principles to learning on LLM computational traces (such as attention maps and activations) to detect failure modes such as hallucinations. I close with the prospect for a broader research programme spanning foundation model monitoring systems bridging language modelling and relational learning, and structure-aware architectures for genomics.
Graph Learning, Equivariance and Expressiveness
Course on Groups and Deep Learning at Technion · 2025/07
About this talk
In this lecture, I discuss learning on graphs and its key challenges, focusing on the interplay between equivariance, expressive power, and computational complexity. I present a coherent narrative that introduces these core aspects and highlights recent research aimed at navigating the trade-offs between them. We begin with feature augmentation techniques, including random feature identifiers, symmetrization methods, and substructure encodings. The lecture then transitions to the broader class of Subgraph GNNs, which aim to combine the strengths of these techniques while maintaining a domain-agnostic design.
Advances in Subgraph GNNs for Expressive and Efficient Learning on Graphs
LoG 2024 Italian Meetup (Keynote Speaker) · 2024/12
About this talk
In this keynote talk, I discuss recent advances in Graph Representation Learning, with a focus on Subgraph Graph Neural Networks (Subgraph GNNs) as more expressive alternatives to traditional message-passing approaches. I begin by exploring the concept of expressive power and the limitations of conventional methods before introducing Subgraph GNNs. While these models offer greater flexibility and expressiveness, they come with the challenge of high computational complexity due to modeling graphs as bags of subgraphs. At the core of the talk, I present HyMN, a novel Subgraph GNN that balances efficiency and expressiveness through subgraph sampling and Structural Encodings (SEs). By leveraging walk-based node centrality measures, HyMN effectively reduces subgraph sizes while enhancing discriminative power and maintaining computational feasibility.
The future and challenges of Graph Learning
Graph Learning Social @ ICML 2024 (Panelist) · 2024/07
About this talk
I took part as a panelist at ICML 2024 Graph Learning Social's panel to discuss on the future and challenges related to Machine Learning on graphs.
Towards Expressive and Efficient Graph Neural Networks
Prof. Thomas Gärtner's ML research unit at TU Wien · 2024/03
About this talk
A presentation on the main work underpinning my PhD thesis. We discuss jointly, in a coherent and gradual fashion, the contributions I made on designing expressive and efficient Graph Neural Networks. We begin by exploring neural architectures on simplicial complexes, a generalization of graphs that captures group-wise interactions. By lifting graphs to simplicial complexes via complete subgraphs, we introduce a hierarchical message-passing approach that surpasses traditional node-centric MPNNs. Extending this, we consider regular cell complexes, incorporating structures like simple and induced cycles for enhanced expressiveness and performance, particularly in molecular modeling. Next, we present an alternative approach that models graphs as bags of subgraphs selected by predefined policies. We design an architecture that respects symmetries, achieving high expressiveness even with efficient, weak message-passing encoders. Finally, we provide a theoretical framework, unifying related works under the bags-of-subgraphs paradigm. A novel symmetry analysis offers an upper bound on expressiveness and defines a design space for future architectures.
Pack your subgraphs: A journey into subgraphs for powerful Graph Neural Networks
Yves-Alexandre de Montjoye's Group Talks (Imperial College London) · 2023/03
About this talk
Our latest two works on subgraphs for more expressive GNNs lined up to tell a coherent story. In the first part of the talk we show the design of a novel framework (ESAN) to process generic bags of subgraphs in an equivariant manner. We show that this approach effectively increases the expressive power of MPNNs. Then, we notice a surge in concurrent approaches which – sometimes unwittingly – make use of subgraphs for powerful graph representations. In the second part, a novel symmetry analysis allows to unify and better characterise these approaches when using a node-based selection policy. We prove an upper-bound on their expressive power and conceive a framework serving as a design space for equivariant node-based subgraph methods.
Pack your subgraphs: A journey into subgraphs for powerful Graph Neural Networks
Course on Geometric Deep Learning at Oxford University · 2023/03
About this talk
Our latest two works on subgraphs for more expressive GNNs lined up to tell a coherent story. In the first part of the talk we show the design of a novel framework (ESAN) to process generic bags of subgraphs in an equivariant manner. We show that this approach effectively increases the expressive power of MPNNs. Then, we notice a surge in concurrent approaches which – sometimes unwittingly – make use of subgraphs for powerful graph representations. In the second part, a novel symmetry analysis allows to unify and better characterise these approaches when using a node-based selection policy. We prove an upper-bound on their expressive power and conceive a framework serving as a design space for equivariant node-based subgraph methods.
Exploring the practical and theoretical landscape of expressive Graph Neural Networks
Learning on Graphs Conference 2022 · 2022/12
Co-presented with: Beatrice Bevilacqua, Haggai Maron
About this talk
The tutorial reviews the most prominent expressive GNNs, categorises them into different families, and draws interesting connections between them. This is accomplished through a series of practical coding sessions and an organic overview of the literature landscape. We aim to convey the importance of studying the expressive power of GNNs and make this field more accessible to our community, especially practitioners and newcomers.
Understanding and Extending Subgraph GNNs by Rethinking their Symmetries
Learning on Graphs and Geometry Reading Group · 2022/11
Co-presented with: Beatrice Bevilacqua
About this talk
Many concurrent works seem to share a similar underlying intuition: to represent a graph by processing subgraphs generated via some simple selection policy. Inspired by a novel symmetry analyses for the most prominent of these methods – those where subgraphs are in a 1:1 correspondence with nodes in the original graph – we prove an upper-bound on their expressive power and conceive a framework serving as a design space for future new equivariant node-based subgraph architectures.
Pack your subgraphs: A journey into subgraphs for powerful Graph Neural Networks
Simone Scardapane's Group Talks (Sapienza) · 2022/11
Co-presented with: Beatrice Bevilacqua
About this talk
We line up our latest two works on subgraphs for more expressive GNNs to tell a coherent story. In the first part of the talk we show the design of a novel framework (ESAN) to process generic bags of subgraphs in an equivariant manner. We show that this approach effectively increases the expressive power of MPNNs. Then, we notice a surge in concurrent approaches which – sometimes unwittingly – make use of subgraphs for powerful graph representations. In the second part, a novel symmetry analysis allows to unify and better characterise these approaches when using a node-based selection policy. We prove an upper-bound on their expressive power and conceive a framework serving as a design space for equivariant node-based subgraph methods.
Pack your subgraphs: A journey into subgraphs for powerful Graph Neural Networks
Meta AI orgs Reading Meeting · 2022/07
Co-presented with: Beatrice Bevilacqua
About this talk
We line up our latest two works on subgraphs for more expressive GNNs to tell a coherent story. In the first part of the talk we show the design of a novel framework (ESAN) to process generic bags of subgraphs in an equivariant manner. We show that this approach effectively increases the expressive power of MPNNs. Then, we notice a surge in concurrent approaches which – sometimes unwittingly – make use of subgraphs for powerful graph representations. In the second part, a novel symmetry analysis allows to unify and better characterise these approaches when using a node-based selection policy. We prove an upper-bound on their expressive power and conceive a framework serving as a design space for equivariant node-based subgraph methods.
Pack your subgraphs: A journey into subgraphs for powerful Graph Neural Networks
Maks Ovsjanikov's Group Talks (École Polytechnique, France) · 2022/07
Co-presented with: Beatrice Bevilacqua
About this talk
We line up our latest two works on subgraphs for more expressive GNNs to tell a coherent story. In the first part of the talk we show the design of a novel framework (ESAN) to process generic bags of subgraphs in an equivariant manner. We show that this approach effectively increases the expressive power of MPNNs. Then, we notice a surge in concurrent approaches which – sometimes unwittingly – make use of subgraphs for powerful graph representations. In the second part, a novel symmetry analysis allows to unify and better characterise these approaches when using a node-based selection policy. We prove an upper-bound on their expressive power and conceive a framework serving as a design space for equivariant node-based subgraph methods.
Subgraphs for more expressive GNNs
Course on Geometric Deep Learning at the African Master's on Machine Intelligence (GDL100) · 2022/07
About this talk
Many concurrent works seem to share a similar underlying intuition: to represent a graph by processing subgraphs generated via some simple selection policy. We discuss these methods together and try to find a common characterisation, while showing that the architecture proposed in our previous work 'Equivariant Subgraph Aggregation Networks' subsumes most of them.
Works covered
Graph Representation Learning on Simplicial and Cellular Complexes
Dagstuhl Seminars – Graph Embeddings: Theory meets Practice · 2022/03
About this talk
I gather together our works on message-passing on Simplicial and Cellular Complexes and present them together showing how these approaches in turn allow for more expressive graph representations.
Equivariant Subgraph Aggregation Networks
Author Interviews with Zak Jost · 2022/02
Co-presented with: Derek Lim, Beatrice Bevilacqua
About this talk
We explore the idea of modeling a graph as a bag of subgraphs generated by simple, domain agnostic selection policies, such as node-deletion, edge-deletion, ego-networks. We formally show that this approach lead to the design Graph Neural Networks which can be made strictly more powerful than standard Message Passing ones. We prove this by defining a novel WL variant and by constructing layers equivariant to the emerging symmetry group.
Works covered
Subgraph Networks
Third Nepal Winter School in AI · 2021/12
About this talk
Many concurrent works seem to share a similar underlying intuition: to represent a graph by processing subgraphs generated via some simple selection policy. We discuss these methods together and try to find a common characterisation, while showing that the architecture proposed in our previous work 'Equivariant Subgraph Aggregation Networks' subsumes most of them.
Works covered
Equivariant Subgraph Aggregation Networks
Learning on Graphs and Geometry Reading Group · 2021/12
Co-presented with: Derek Lim, Beatrice Bevilacqua
About this talk
We explore the idea of modeling a graph as a bag of subgraphs generated by simple, domain agnostic selection policies, such as node-deletion, edge-deletion, ego-networks. We formally show that this approach lead to the design Graph Neural Networks which can be made strictly more powerful than standard Message Passing ones. We prove this by defining a novel WL variant and by constructing layers equivariant to the emerging symmetry group.
Works covered
Anees Kazi on 'Graph Convolutional Network for Disease Prediction with Imbalanced Data'
London Machine Learning Meetup · 2021/10
About this talk
I moderate Anees Kazi's talk at the London Machine Learning Meetup.
Weisfeiler and Lehman Go Cellular: CW Networks
Learning on Graphs and Geometry Reading Group · 2021/08
Co-presented with: Cristian Bodnar
About this talk
We extend message-passing from graphs and Simplicial Complexes to Cellular Complexes, combinatorial topological spaces generalising these two. By lifting graphs to Cellular Complexes we can naturally model molecules and unlock a series of advantages w.r.t. standard pairwise message-passing.
Works covered
Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
Course on Geometric Deep Learning at the African Master's on Machine Intelligence (GDL100) · 2021/08
About this talk
We re-define colour refinement and message-passing on Simplicial Complexes, topological generalisation of graphs. This allows to model higher-order interactions, model meso-scale structures and increase the expressive power of Graph Neural Networks.
The expressive power of GNNs by the WL test
London Geometry and Machine Learning Summer School 2021 · 2021/07
Co-presented with: Giorgos Bouritsas
About this talk
What is the current state of affair in provably more expressive Graph Neural Networks? Giorgos Bouritsas and I lead and moderate this interactive discussion on approaches to go beyond 1-WL discriminating power. Ultimately, we would like to tackle the following question: 'Is the Weisfeiler-Leman hierarchy suitable to study the representational power of learning models on graphs? Do we necessarily need to go beyond message-passing?'
Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
TopoNets 2021 – Networks Beyond Pairwise Interactions · 2021/06
Co-presented with: Cristian Bodnar
About this talk
We re-define colour refinement and message-passing on Simplicial Complexes, topological generalisation of graphs. This allows to model higher-order interactions, model meso-scale structures and increase the expressive power of Graph Neural Networks.
Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
Artificial Intelligence Research Group Talks (Cambridge University) · 2021/05
Co-presented with: Cristian Bodnar, Yu Guang Wang
About this talk
We re-define colour refinement and message-passing on Simplicial Complexes, topological generalisation of graphs. This allows to model higher-order interactions, model meso-scale structures and increase the expressive power of Graph Neural Networks.
Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
Graph Journal Club (Valence AI) · 2021/04
Co-presented with: Cristian Bodnar, Yu Guang Wang
About this talk
We re-define colour refinement and message-passing on Simplicial Complexes, topological generalisation of graphs. This allows to model higher-order interactions, model meso-scale structures and increase the expressive power of Graph Neural Networks.
Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
Math Machine Learning seminar MPI MIS + UCLA · 2021/04
Co-presented with: Cristian Bodnar, Yu Guang Wang
About this talk
We re-define colour refinement and message-passing on Simplicial Complexes, topological generalisation of graphs. This allows to model higher-order interactions, model meso-scale structures and increase the expressive power of Graph Neural Networks.
Improving Graph Neural Network Expressivity via Subgraph Isomorphism Counting
Graph Journal Club (Valence AI) · 2021/04
Co-presented with: Giorgos Bouritsas
About this talk
We study how isomorphism counting of small substructures can improve the expressive power of Message Passing Neural Networks, while retaining equivariance, a good inductive bias, and their tractable fwd-pass computational complexity.
My personal website has been realised through Jekyll and Github Pages. Original theme is by orderedlist, but it has been customised quite a bit.