Temporal Point Processes in Theory & Practice
Temporal point processes provide a mathematical framework for modeling discrete events in continuous time, with applications ranging from social networks and finance to healthcare and neuroscience. This presentation introduces the core theory behind TPPs, explores both classical intensity-based and modern intensity-free neural approaches, and demonstrates how recent methodological advances are enabling more accurate, efficient, and interpretable event sequence modeling across diverse real-world domains.Script
When do earthquakes strike, trades execute, or neurons fire? These aren't evenly spaced ticks of a clock. Temporal point processes give us the mathematics to model events that happen in continuous time, capturing not just when they occur, but why the timing itself carries meaning.
At the heart of every temporal point process is the conditional intensity function, lambda star of t. It tells us the instantaneous probability of an event occurring right now, given everything that's happened before. This history dependence is what makes point processes so powerful for capturing cascades, bursts, and self-exciting dynamics.
Modern neural temporal point process models come in two flavors. Intensity-based models learn the rate function directly, often using recurrent networks or transformers to encode event history. Intensity-free models, like CuFun, take a different path: they parameterize the cumulative distribution of inter-event times with monotonic neural networks, eliminating the need for numerical integration and achieving both speed and stability.
CuFun's architecture enforces monotonicity by construction: positive weights, sigmoid nonlinearities, and an element-wise product that fuses temporal and historical information. Because the cumulative distribution function is learned directly, you can compute likelihoods via automatic differentiation with no numerical quadrature. On benchmarks from finance to social networks, this approach delivers state-of-the-art accuracy and remarkable numerical robustness.
The real test of any framework is how it performs in the wild. Temporal point processes are now modeling patient disease progression in intensive care units, predicting high-frequency trades on the New York Stock Exchange, forecasting user interactions on social platforms, and decoding neural spike trains. Hybrid models that integrate temporal logic rules with neural intensity functions have even achieved interpretable, clinically valid predictions that clinicians can trust and audit.
Temporal point processes have evolved from hand-crafted kernels to universal neural approximators, unlocking expressiveness and performance once thought impossible. Yet the frontier remains wide open: scaling to millions of events, integrating multimodal marks, ensuring interpretability, and adapting online to streaming data. If you're curious to learn more about these methods or create your own video summaries of cutting-edge research, visit EmergentMind.com and explore the mathematics of time itself.
