Reproduce “Distribution-Free Conformal Joint Prediction Regions for 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
Victor Dheur, Tanguy Bosser, Rafael Izbicki, Souhaib Ben Taieb
A conformal prediction framework generates distribution-free joint prediction regions for event arrival times and marks in neural Temporal Point Processes. The proposed Conformal Highest Joint Density Region (C-HDR) method yields sharper, more informative regions by accounting for interdependencies between time and mark, outperforming naive approaches.
Key insights
- Conformal prediction consistently achieves marginal coverage guarantees (≥ 1-α) for neural TPPs, a crucial reliability metric that heuristic uncertainty methods often fail to meet.
- The C-HDR method produces more efficient and sharper joint prediction regions by effectively leveraging the interdependencies between event arrival time and mark, unlike naive independent combinations.
- While all conformal methods ensure marginal coverage, C-HDR demonstrates competitive conditional coverage, indicating a better ability to adapt prediction regions to specific input histories.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.