Machine Learning on Dynamic Graphs and Temporal Graph Networks | MLSys 2021

If you have any copyright issues on video, please send us an email at [email protected] 0:00 Introduction 0:23 Graphs are everywhere 0:42 Graph Neural Networks 1:10 Problem: Many Graphs are Dynan 3:48 From Static to Dynamic Grapl 7:24 CTDGS: Many Types of Event 8:22 Why is Learning on Dynamic Graphs Different? Model needs to: - Handle different types of events - Use the time information of the events - Efficiently and incrementally incorporate new events - Different tasks: predict when something will happen 10:08 Temporal Graph Model 11:06 Encoding a Temporal Graph 17:31 TGN: Temporal Graph Networ 20:04 Modules: Message Function - Each event generates a message • Messages will be used to update the memory 21:17 Modules: Memory Updater 21:47 Modules: Graph Embedding 23:38 Future Link Prediction 24:23 Scalability 25:10 Experiments: Future Edge Predicti 26:35 Experiments: Dynamic Node Classificati 27:44 Predicting when events will happen - Qualitatively different question from other tasks - A decoder which makes use of Temporal Point Processes is needed 28:28 Future Work 29:10 Conclusion 29:35 Questions? 32:34 Ablation Study (Future edge prediction)