Yan Ting Chok

Junior ML researcher specialised in drug discovery.
Department of Computer Science and Engineering, Korea University.

[Yan Ting Chok @yanting1412]

9. 🧵Here is a helpful explanation provided by the author to grasp the overall idea of the paper: https://t.co/q6ywOZP0fP

Take a look if you are interested!

[Yan Ting Chok @yanting1412]

Recently, I came across this interesting paper entitled ‘Edge Directionality Improves Learning on Heterophilic Graphs’. https://t.co/flFHMEMYnr
#LoG #LearningOnGraphs #GraphPaper

[Yan Ting Chok @yanting1412]

8. 🧵I believe that extensive research into edge directionality would, to some extent, benefit association analysis in the biomedical domain, which in turn, also boost the performance of any downstream tasks that utilize the representations obtained from these prior knowledge.

[Yan Ting Chok @yanting1412]

7. 🧵For example, in a drug-gene interaction network, an edge between a drug and a gene could represent inhibition or activation, and the inability to model this distinction could confuse the learning process.

[Yan Ting Chok @yanting1412]

6. 🧵The question of whether utilizing edge directionality and heterophily hurts graph learning has always intrigued me. In the biomedical domain, there are many directed, heterophilic graphs, such as drug-gene interaction graphs and the famous gene regulatory network (GRAND).

[Yan Ting Chok @yanting1412]

🧵5. To show the generalizability of the proposed framework, the author presents experimental results for various base models, such as GCN, GAT, and GraphSage, on several homophilic and heterophilic benchmark datasets.

[Yan Ting Chok @yanting1412]

4. 🧵The results demonstrate that utilizing edge directionality information increases the effective homophily of graphs overall and is particularly beneficial for learning on heterophilic graphs.

[Yan Ting Chok @yanting1412]

🧵3. To address this, the author proposes Dir-GNN, which accounts for edge directionality by performing separate aggregations of incoming and outgoing edges, thereby preserving expressivity and minimizing complexity.

[Yan Ting Chok @yanting1412]

🧵2. In the paper, the author mentions about the common practice of discarding the directionality of graphs due to some historical reasons and points out the limitations of previous works that use the same parameters for in- and out-neighbors, along with their scalability issues.

[Yan Ting Chok @yanting1412]

🧵1. I have always been interested in learning the improvement in expressivity in terms of the presence and absence of directionality and heterophily, so I would love to share some humble insights of me as a junior ML researcher specialized in drug discovery.