Reproduce “Addressing Mark Imbalance in Integration-free Neural Marked Temporal Point Processes” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sishun Liu, Ke Deng, Yongli Ren, Yan Wang, Xiuzhen Zhang
Researchers at RMIT University and Macquarie University introduced a "mark-first-then-time" prediction strategy with a novel Integration-free Neural Marked Temporal Point Process (IFNMTPP) model to address mark imbalance in Marked Temporal Point Processes. This approach utilizes an F1-score optimized thresholding mechanism to enhance prediction accuracy for rare events, yielding superior performance over existing MTPP models and traditional imbalance techniques.
Key insights
- Mark imbalance is a critical, unaddressed problem in MTPPs that significantly impacts the prediction accuracy of rare, yet often important, events.
- Reversing the prediction order from "time-first-then-mark" to "mark-first-then-time" is essential for effectively applying thresholding to mitigate mark imbalance in MTPPs.
- Directly modeling a unified integral function (Γ*(m,t)) with a specialized neural network architecture allows for efficient computation and time sampling, overcoming the limitations of previous integration-heavy MTPP models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.