Towards Better Evaluation for Dynamic Link Prediction

Explore dynamic link prediction in graphs through this 55-minute conference talk by Andy Huang and Farimah Poursafaei from Valence Labs. Delve into the challenges of learning from time-evolving graphs and discover new evaluation procedures designed to better reflect real-world scenarios. Examine visualization techniques for understanding edge reoccurrence patterns over time and learn about EdgeBank, a memorization-based baseline that highlights shortcomings in current evaluation methods. Investigate novel negative sampling strategies aimed at improving robustness and relevance to real-world applications. Gain insights into six new dynamic graph datasets from diverse domains, offering fresh challenges for future research. The talk covers an introduction to dynamic temporal graph learning, dynamic link prediction, evaluation methods, visualization techniques, negative sampling approaches, and the EdgeBank baseline, concluding with a Q&A session.

- Intro
- Part 1: Intro to Dynamic Temporal Graph Learning
- Dynamic Link Prediction
- Part 2: Towards Better Evaluations for Dynamic Link Predictions
- Understanding Dynamic Graphs: TEA and TET Plots
- Random Negative Sampling
- Introducing EdgeBank
- Q+A