dissemination
academic talks and poster presentations
talks
Hierarchical pooling in Graph Neural Networks
Keynote TalkAbstract
In Graph Neural Networks, pooling operations play a crucial role in capturing hierarchical structures and global properties of graphs. Through hierarchical pooling, we can progressively condense graph representations, enabling both the extraction of high-level features and the expansion of the effective range of message-passing operations.
This talk presents a comprehensive overview of hierarchical graph pooling, introducing a unified framework for understanding pooling operators. We explore the main approaches to graph pooling, discussing their theoretical foundations and practical implications. The presentation concludes with an analysis of evaluation methodologies used to assess pooling effectiveness and their impact on various graph learning tasks.
Graph neural networks for conditional de novo drug design
Invited TalkAbstract
Drug discovery is a complex, time-consuming, and costly process. De novo drug design, particularly the generation of molecules with specific desired properties, remains one of its most challenging aspects. This talk presents two contributions to this field.
First, we provide an overview of graph-based approaches for conditional molecular generation, introducing a novel taxonomy that categorizes existing methods based on their learning frameworks, generation processes, and conditioning techniques.
Then, we introduce AMCG (Atomic-Molecular Conditional Generator), a novel graph-based generative framework that addresses several key limitations of existing approaches. AMCG employs a dual atomic-molecular representation and innovative sampling strategies to generate valid, diverse molecules without size constraints. The model demonstrates competitive performance on benchmark datasets and offers flexible conditioning mechanisms for both structural features and molecular properties, representing an initial step toward more efficient and controlled molecular generation in drug discovery.
References:
posters
MaxCutPool: Differentiable Feature-Aware MAXCUT for Pooling in Graph Neural Networks
Abstract
We propose a novel approach to compute the MAXCUT in attributed graphs, i.e. graphs with node features. Our method, MaxCutPool, is a differentiable pooling layer for Graph Neural Networks that leverages both graph topology and node features to compute hierarchical representations.
Key Contributions
- A novel MAXCUT computation method for attributed graphs
- A new hierarchical pooling layer especially effective for heterophilic graphs
- A general scheme for node-to-supernode assignment
- The introduction of the first heterophilic dataset for graph classification
Impact
Experimental results demonstrate that MaxCutPool achieves state-of-the-art performance across various graph classification and node classification tasks, highlighted by perfect accuracy on expressiveness tests and significant improvements on heterophilic graph classification.
MaxCutPool: Differentiable Feature-Aware MAXCUT for Pooling in Graph Neural Networks
Abstract
We propose a novel approach to compute the MAXCUT in attributed graphs, i.e. graphs with features associated with nodes and edges, by exploiting heterophilic message passing to assign connected nodes to different partitions. The approach is fully differentiable, making it possible to find solutions that jointly optimize the MAXCUT along with other objectives.
Based on the obtained MAXCUT partition, we implement MaxCutPool, a hierarchical graph pooling layer for graph neural networks. The layer is sparse, differentiable, and particularly suitable for downstream tasks on heterophilic graphs.
Key Contributions
- A novel MAXCUT computation method for attributed graphs
- A new hierarchical pooling layer especially effective for heterophilic graphs
- A general scheme for node-to-supernode assignment
- The introduction of the first heterophilic dataset for graph classification
Impact
Experimental results demonstrate that MaxCutPool achieves state-of-the-art performance across various graph classification and node classification tasks, highlighted by perfect accuracy on expressiveness tests and significant improvements on heterophilic graph classification.
AMCG: a graph dual atomic-molecular conditional molecular generator
Abstract
Drug design is both a time consuming and expensive endeavor. Computational strategies offer viable options to address this task; deep learning approaches in particular are indeed gaining traction for their capability of dealing with chemical structures.
A straightforward way to represent such structures is via their molecular graph, which in turn can be naturally processed by Graph Neural Networks (GNNs).
In this poster we present AMCG, a recently introduced dual Atomic-Molecular, Conditional, latent-space, Generative model built around graph processing layers able to support both unconditional and conditional molecular graph generation.
Among other features, AMCG is a one-shot model allowing for fast sampling, explicit atomic type histogram assignation and property optimization via gradient ascent.
The model was trained on the Quantum Machines 9 (QM9) and ZINC datasets, achieving state-of-the-art performances. Together with classic benchmarks, AMCG was also tested by generating large-scale sampled sets, showing robustness in terms of sustainable throughput of valid, novel and unique molecules.