Reproduce “Preventing Conflicting Gradients in Neural Marked Temporal Point Processes” on OpenResearch

About this paper

Tanguy Bosser, Souhaib Ben Taieb

This research identifies the problem of conflicting gradients during the training of neural Marked Temporal Point Processes (MTPPs) and introduces architectural modifications to mitigate it. The proposed framework, which enables disjoint parameter learning for time and mark prediction tasks, consistently improved predictive accuracy and probabilistic calibration on multiple real-world event sequence datasets.

Key insights

• Conflicting gradients frequently occur in standard neural MTPP training, leading to degraded predictive performance on both time and mark prediction tasks.
• Separating model parameters for time and mark prediction at the architectural level effectively prevents conflicting gradients and improves optimization stability.
• Fully disjoint training of time and mark prediction tasks, particularly with separate history encoders, enhances mark prediction accuracy and calibration, indicating that mark prediction is often more susceptible to shared-parameter conflicts.

Summary by alphaXiv.