🤖 machine learning

Structured Neural Marked Point Processes for Interpretable Event Interaction Modeling

This paper proposes a Structured Neural Marked Point Process (SNMPP) that combines a signed interaction network and a delay-aware temporal network to simultaneously achieve high predictive accuracy and explicit, interpretable discovery of structured inter-event relationships in multi-class event streams.

Original authors: Zhitong Xu, Qiwei Yuan, Yinghao Chen, Shandian Zhe, Bin Shen

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhitong Xu, Qiwei Yuan, Yinghao Chen, Shandian Zhe, Bin Shen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a busy city square. People are walking in, buying coffee, dropping letters, and leaving. Sometimes, one person buying a coffee makes the line longer, causing more people to stop and chat (a "positive" effect). Other times, a sudden loud noise might make everyone scatter and stop talking for a while (a "negative" effect). Sometimes, a specific event happens, but its effect doesn't show up until ten minutes later, like a delayed reaction to a spicy meal.

This paper introduces a new tool called SNMPP (Structured Neural Marked Point Processes) to understand these kinds of busy, event-filled streams.

Here is the breakdown of what the paper does, using simple analogies:

The Problem: The "Black Box" vs. The "Transparent Map"

For a long time, computers have been good at predicting when the next event will happen in a stream like this. They use complex "neural networks" (think of them as super-smart, but opaque, black boxes).

The Issue: These black boxes are great at guessing the next event, but they are terrible at explaining why. They know that "Event A" usually leads to "Event B," but they hide that relationship inside layers of math that humans can't read. It's like a GPS that tells you to turn left but refuses to show you the map or explain that there is a roadblock ahead.
The Goal: The authors wanted a model that is just as good at predicting, but also acts like a transparent map. They wanted to see exactly how different events influence each other: Do they encourage each other? Do they stop each other? Is there a delay?

The Solution: The "Recipe" Approach

The authors built SNMPP, which is like a recipe for understanding event streams. Instead of one giant, messy black box, they built the model using two distinct, understandable ingredients that work together:

The "Relationship Chef" (Signed Interaction Network):
Imagine a chef who decides the type of relationship between two ingredients.
- If the chef sees "Coffee" and "Chat," they say, "These two excite each other" (Positive influence).
- If they see "Loud Noise" and "Silence," they say, "These inhibit each other" (Negative influence).
- If they see "Rain" and "Traffic," they might say, "These are neutral."
  This part of the model explicitly learns whether events push each other forward or hold them back.
The "Time-Lapse Camera" (Delay-Aware Temporal Network):
Imagine a camera that records how the effect of an event changes over time.
- Standard models assume the effect is strongest the instant something happens and then fades away immediately (like a firework).
- SNMPP realizes that sometimes the effect is weak at first, builds up, and peaks later (like a delayed medical treatment or a supply chain order).
- This part of the model learns the specific shape of that curve and even finds the exact "delay" time before the peak effect hits.

How They Trained It: The "Stratified Tasting"

Training these models is hard because you have to calculate the probability of events happening over time, which involves complex math integrals (summing up infinite tiny possibilities).

The Old Way: Imagine trying to taste a huge pot of soup by taking 100 random spoonfuls from the top. You might miss the salt at the bottom or the spice at the top.
The SNMPP Way: The authors invented a "Stratified Monte Carlo" method. Imagine cutting the pot of soup into 4 equal layers and taking exactly one spoonful from each layer. This ensures they taste every part of the soup equally. This makes the training much more stable and accurate without needing to taste the whole pot.

What They Tested It On

The authors didn't just talk about the theory; they tested it in three ways:

Fake Data (The Lab Test): They created artificial event streams where they knew the "ground truth" (e.g., "We know Event A triggers Event B after exactly 1 second"). SNMPP successfully found these hidden rules, including the delays and the "stop" signals, whereas other models missed them or couldn't explain them.
Real World Data (The Field Test): They tested it on real datasets like:
- Amazon reviews: When people buy one thing, do they buy another?
- Twitter retweets: Does a tweet make people retweet later?
- Taxi trips: Do pickups in one area lead to drop-offs in another?
- Medical records (MIMIC): What happens after a patient visits a doctor?
- Result: SNMPP was just as good at predicting the next event as the best "black box" models, but it also gave them the "map" of how events interact.
The Supply Chain Simulation (The Logic Test): They created a computer simulation of a warehouse. The simulation had strict rules (e.g., "If stock hits zero, stop selling"). The simulation wasn't built using the math the authors were testing; it was built on physical logic.
- The Result: Even though the data came from a logic-based simulation, SNMPP looked at the data and correctly figured out the hidden rules: "Orders cause restocking," "Restocking causes stock to arrive later," and "Running out of stock stops new orders." It successfully "reverse-engineered" the logic of the system.

The Bottom Line

This paper presents a tool that does two things at once:

It predicts the future with high accuracy.
It explains the structure of the past in plain language (e.g., "Event A encourages Event B, but only after a 4-hour delay").

It bridges the gap between powerful AI prediction and human understanding, allowing us to see the invisible connections and delays in complex systems like supply chains, social media, and medical records.

Technical Summary: Structured Neural Marked Point Processes for Interpretable Event Interaction Modeling

Problem Statement
Multi-class event streams are ubiquitous in domains such as social networks, supply chains, and medical records. A central challenge in modeling these streams is balancing high predictive accuracy with the need to uncover structured, interpretable inter-event relationships. Existing neural point process models (e.g., Neural Hawkes Processes, Transformer Hawkes Processes) offer high expressivity by using deep neural networks to parameterize conditional intensities. However, they encode event interactions in a "black-box" manner within latent representations, preventing the explicit discovery of structured dependencies such as excitation, inhibition, or delayed effects. Conversely, classical models like standard Hawkes processes are interpretable but often restricted to excitatory effects and assume immediate influence decay, failing to capture delayed peaks or inhibitory dynamics common in real-world systems.

Methodology: Structured Neural Marked Point Process (SNMPP)
The authors propose SNMPP, a framework designed to achieve modeling flexibility while enabling explicit, data-driven discovery of event-wise and class-wise interaction structures.

Model Architecture:
The core of SNMPP is a product-form neural influence kernel that decomposes the influence of a past event of type $k_n$ on the intensity of type $k$ into two distinct neural components:
$f(e_{k_n}, e_k, \Delta t) = \psi(e_{k_n}, e_k) \cdot \phi(e_{k_n}, e_k, |\Delta t - d_{k_n, k}|)$
1. Signed Interaction Network ( $\psi$ ): This component takes embeddings of the source and target event types as input and outputs a signed scalar. It explicitly characterizes the interaction topology: positive values indicate excitation, negative values indicate inhibition, and values near zero indicate neutrality. This allows the model to learn directed relationships without separate parameters for every ordered pair.
2. Delay-Aware Monotonic Temporal Network ( $\phi$ ): This component models the temporal evolution of the influence strength. It is designed to be monotonically decreasing with respect to the time lag to capture decay patterns. Crucially, the input to this network is the absolute time difference offset by a learnable delay parameter ( $d_{k_n, k}$ ). This formulation allows the model to capture delayed peak effects, where influence is initially weak, peaks at a specific lag, and then decays.
3. Constraints: To ensure interpretability and stability, the temporal network's output is constrained to the bounded range $[0, 1]$ using a differentiable soft-clipping transformation. This separates the magnitude/sign (handled by $\psi$ ) from the temporal shape (handled by $\phi$ ).
Training Algorithm:
The likelihood of the observed event sequences involves integral terms that lack closed-form solutions. The authors develop a stratified Monte Carlo estimator for efficient training. Instead of drawing samples uniformly from the entire inter-event interval, the interval is partitioned into $Q$ non-overlapping segments, with one sample drawn from each. This stratification reduces the variance of the likelihood estimator while preserving sensitivity to local variations in the intensity function, enabling stable stochastic gradient descent.

Key Contributions

Explicit Interaction Discovery: Unlike black-box neural models, SNMPP explicitly recovers the interaction topology (excitation, inhibition, neutrality) and temporal dynamics (decay patterns, delay parameters) directly from data.
Flexible Temporal Modeling: The model accommodates diverse temporal behaviors, including delayed effects where influence peaks after a time lag, a feature often observed in medical treatments and supply chain replenishment but missed by standard exponential kernels.
Efficient Learning: The proposed stratified Monte Carlo estimator provides an unbiased, low-variance approximation of the likelihood integrals, facilitating robust training on complex datasets.

Experimental Results
The paper validates SNMPP through three distinct experimental settings:

Synthetic Data: On synthetic datasets featuring delayed triggering and inhibitory interactions, SNMPP accurately recovers the ground-truth conditional intensities, interaction signs, delay parameters, and kernel shapes. It outperforms standard Hawkes processes (which fail on inhibition/delay) and neural baselines (which fail to explicitly recover the structure).
Real-World Benchmarks: Evaluated on five datasets (MIMIC, Amazon, Retweet, Taxi, StackOverflow), SNMPP achieves state-of-the-art performance in next-event time prediction (lowest RMSE) across all datasets. While event-type prediction accuracy is competitive with the best neural baselines, it does not always surpass them, demonstrating that the structured inductive bias does not compromise predictive power.
Supply-Chain Simulation: In a simulated environment governed by physical constraints and decision rules (rather than a predefined point process), SNMPP successfully uncovers the underlying interaction structure. It correctly identifies excitation (orders triggering replenishment), inhibition (stock arrivals suppressing new orders), hard gating (stockouts suppressing demand), and the specific delay parameter (approx. 4.0 time units) corresponding to lead times.

Significance and Claims
The paper claims that SNMPP successfully bridges the gap between the high expressivity of neural point processes and the interpretability of classical models. By introducing a structured neural influence kernel, the method enables the explicit discovery of structured relationships (including inhibition and delayed effects) without sacrificing predictive accuracy. The authors argue that this structured parameterization acts as an effective inductive bias, regularizing the learning problem and improving temporal generalization. The work demonstrates that it is possible to move beyond black-box representations to uncover interpretable mechanisms—such as delayed peak effects and inhibitory topologies—in complex, multi-class event streams. Future work is noted to potentially extend the model to asymmetric temporal influence profiles.