When modeling real-world event sequences—such as electronic health records, financial transactions, or seismic activity—we want to answer two questions: what will happen next, and when? Traditionally, the mathematical tool of choice is the Temporal Point Process (TPP). Classical TPPs rely on a conditional intensity function $\lambda(t | \mathcal{H}_t)$, which represents the instantaneous rate at which events occur given the historical sequence of events $\mathcal{H}_t$.
While elegant, intensity-based models suffer from a fundamental computational bottleneck during training and simulation. A growing body of research shifts this paradigm toward intensity-free temporal modeling, transforming event prediction from rate estimation into direct conditional density estimation.
The Integration Bottleneck
To understand why intensity-free modeling works, we must first look at why traditional intensity-based models struggle. The likelihood of observing a sequence of events up to time $T$ is given by:
$$L = \prod_{i=1}^{N} \lambda(t_i | \mathcal{H}_{t_i}) \cdot \exp\left( -\int_{0}^{T} \lambda(s | \mathcal{H}_s) \, ds \right)$$
The right-hand term represents the survival probability—the likelihood that no events occurred in the intervals between observations. To compute this, we must integrate the intensity function over time.
For simple parametric models like the Hawkes process, this integral can be solved analytically. However, when we use deep neural networks to capture complex, non-linear history dependencies, the intensity function becomes highly intricate. As a result, calculating the survival probability requires computationally expensive numerical integration (such as Monte Carlo approximations).
Simulating new events from an intensity-based model is equally painful. Because we only have the event rate, we must rely on iterative rejection sampling algorithms (like Lewis-Shedler thinning). These algorithms simulate time ticking forward step-by-step and accept or reject candidate events based on the intensity at that moment, leading to high computational overhead.
Shifting to Direct Density Estimation
As surveyed by Angelica Liguori et al. in Modeling Events and Interactions through Temporal Processes - A Survey, intensity-free modeling bypasses this "integration nightmare." Instead of modeling the instantaneous rate $\lambda(t)$, these frameworks directly estimate the probability density function (PDF) $p(\tau | \mathcal{H})$ of the inter-event times $\tau_i = t_i - t_{i-1}$.
Traditional Intensity-Based TPP:
History (H) ──> Conditional Intensity λ(t) ──> [Numerical Integration] ──> Likelihood / Sampling
Intensity-Free TPP:
History (H) ──> Direct Density p(τ) ────────> [Closed-Form Formulas] ───> Likelihood / Sampling
By parameterizing the probability distribution of the delays directly, the cumulative distribution function (CDF) $P(\tau)$ and the survival function $S(\tau) = 1 - P(\tau)$ become readily accessible. This shift yields two massive advantages:
- Closed-Form Likelihoods: The log-likelihood simplifies to evaluating $\sum \log p(\tau_i | \mathcal{H})$, completely eliminating numerical integration during training.
- Single-Forward-Pass Generation: Generating an event sequence no longer requires rejection sampling. We can sample a delay $\tau$ directly from $p(\tau | \mathcal{H})$ in a single forward pass, accelerating simulation speeds by orders of magnitude.
Parameterizing the Inter-Event Density
To make direct density estimation work, the chosen distribution must be flexible enough to capture multi-modal behaviors while respecting the fact that time intervals are strictly positive ($\tau > 0$).
Log-Normal Mixtures
In their seminal paper Intensity-Free Learning of Temporal Point Processes, Oleksandr Shchur et al. demonstrated that a mixture of log-normal distributions acts as a universal approximator for one-dimensional temporal distributions. Because log-normal distributions are naturally skewed and strictly positive, they perfectly fit physical process delays—a property also leveraged by Michael Joya in An Event-Based Framework for Process Inference.
Beyond flexibility, Vinayak Gupta et al. show in Learning Temporal Point Processes with Intermittent Observations that log-normal mixtures allow the use of the reparameterization trick. This makes it easy to backpropagate through event times, which is particularly useful when handling missing or hidden historical data.
Normalizing Flows and Splines
For even greater expressivity, researchers use normalizing flows to warp simple base distributions into complex temporal profiles. Felix Draxler et al. utilize rational-quadratic splines in Transformers for Mixed-type Event Sequences to map inter-event times directly, maintaining high computational stability.
When dealing with multivariate point processes, Guilherme Augusto Zagatti et al. show in Learning multivariate temporal point processes via the time-change theorem that using invertible triangular maps ensures the transformation's Jacobian remains easy to compute. This preserves exact likelihood calculations while scaling linearly with sequence length.
Alternative Intensity-Free Paradigms
Directly modeling the PDF is not the only way to achieve intensity-free learning. Several other generative frameworks have been adapted to bypass the intensity function entirely.
Quantile Function Regression
Rather than predicting a probability density and inverting it to sample an event, Souhaib Ben Taieb proposes a different route in Learning Quantile Functions for Temporal Point Processes with Recurrent Neural Splines. By modeling the conditional quantile function directly, the model maps a uniformly sampled probability level $\alpha \in [0, 1]$ directly to a time offset $\tau$. This treats uncertainty as a primary output and completely eliminates both numerical integration and iterative root-finding during inference.
Generative Adversarial Networks (GANs)
The very first intensity-free framework, introduced by Shuai Xiao et al. in Wasserstein Learning of Deep Generative Point Process Models, relies on an adversarial setup (WGANTPP). Instead of calculating explicit likelihoods, a generator network takes a sequence of base noise and transforms it into a target event sequence in a single forward pass, while a discriminator evaluates the sequence realism using the Wasserstein distance.
Denoising Diffusion Models
For high-dimensional marked point processes (where each event is accompanied by complex metadata like text, images, or spatial coordinates), calculating joint densities becomes intractable. To solve this, Yuan Yuan et al. developed Spatio-temporal Diffusion Point Processes, which treats event prediction as an iterative denoising task.
By reversing a Gaussian noise process conditioned on event history, models like the one proposed by Shixiang Zhu et al. in Conditional Generative Modeling for High-dimensional Marked Temporal Point Processes can generate high-fidelity event times and high-dimensional marks simultaneously, completely sidestepping the integration nightmare over complex mark spaces.
Key Foundational Work
Wasserstein Learning of Deep Generative Point Process Models (2017) — contributed 2 concept elements to this direction.
An Event-Based Framework for Process Inference (2011) — contributed 2 concept elements to this direction.
Visualization of patient specific disease risk prediction (2017) — contributed 1 concept elements to this direction.