Implementation of "Towards Efficient Fusion for Graph Neural Networks"
Humans experience the world around them from multiple sensory modalities. Since real-world data is quite often multimodal, understanding different modalities is meaningful for ML models. Although both Multimodal deep learning and Graph Machine Learning has significantly developed in recent years, but multimodal learning with graphs was not been fully explored. In this paper, we transfer the idea of fusion to graphs and create a universal framework for fusion among multiple entities. Since interaction between a few graphs commonly occurs in Drug Discovery, we evaluate our framework on drug pair scoring tasks ( drug synergy prediction and drug-drug interaction). Our method reached state-of-the-art results on the O’Neil dataset on Bliss, Loewe, HSA, and ZIP synergy scores with much fewer parameters in comparison with the previous solutions.
The project report can be found here
Oleksii Tsepa, Bohdan Naida