Introduction
Many real-world datasets—such as server logs, financial transactions, or clinical visits—do not manifest as smooth time series. Instead, they appear as "event streams" characterized by irregular bursts of activity. To model these sequences, researchers rely on Graphical Event Models (GEMs), which describe how the occurrence of one event type influences the probability (or intensity) of another.
A major bottleneck in existing GEMs is their reliance on human experts to specify "basis functions." Essentially, users must guess the relevant time intervals for event dependencies (e.g., determining if an event happens within 5 minutes or 5 hours of its parent). Incorrect guesses lead to poor model performance, while attempting to include every possible interval renders computation intractable. To address this, researchers introduce Proximal Graphical Event Models (PGEMs), a data-driven approach that removes the human from the loop by automatically learning these critical time windows directly from the data.
Methods
The core framework of the Proximal Graphical Event Model (PGEM) posits that the rate of an event $X$ depends on a simple binary condition: has a parent event $Z$ occurred at least once within a specific "proximal" window $w_{zx}$?
Key technical contributions of this approach include:
- Optimal Window Learning: The authors mathematically prove that the optimal time window is not an arbitrary continuous value. Instead, it belongs to a finite set of "candidate" windows derived directly from the actual inter-event times present in the dataset.
- Structure Learning: The model employs a Forward-Backward Search (FBS) algorithm, guided by the Bayesian Information Criterion (BIC), to discover parent-child relationships between event types.
- Theoretical Bounds: The researchers provide a mathematical proof bounding the number of parents for any event by $O(\log N)$, where $N$ is the total number of events. This crucial bound prevents the search space from exploding as datasets grow.
- Efficient Heuristics: To handle large-scale data, two optimized versions of the search are proposed: FBS-IW (Independent Windows) and FBS-CW (Coordinate-wise optimization).
Key Insights
The success of PGEMs relies on a few core insights. First is the "Proximal" Assumption: by assuming only the most recent occurrence of a parent matters, the model remains highly interpretable, effectively answering, "If $A$ happened recently, how much more likely is $B$?"
Second is the concept of Data-Adaptive Windows. By searching for optimal windows at the jump points of the likelihood function (the inter-event times), the model identifies the exact temporal scale of influence without requiring manual tuning. Finally, using BIC combined with derived parent-set bounds ensures strict Complexity Control, preventing the model from overfitting by adding spurious edges to the dependency graph.
Datasets and Key Results
The authors evaluated PGEMs across a mix of controlled and real-world datasets:
- Synthetic Datasets: Six distinct PGEM structures with varying numbers of event labels ($M=5$ and $M=10$) were used to test ground-truth recovery.
- Book Datasets: Word occurrences in Thomas Hobbes’ Leviathan and the Bible (treating individual words as discrete events).
- ICEWS (Integrated Crisis Early Warning System): A massive dataset of political events across five countries (Argentina, Brazil, Colombia, Mexico, and Venezuela).
In synthetic tests, PGEMs consistently outperformed the SIP baseline and matched or exceeded the performance of CPCIM, a state-of-the-art model that requires manual basis functions. On the real-world book datasets, PGEMs achieved significantly higher log-likelihoods than CPCIM, often by margins of several thousand points. Furthermore, in the ICEWS dataset, PGEMs outperformed competing models in 4 out of 5 countries, successfully capturing complex political dynamics without prior domain knowledge of the specific regions.
Limitations and Future Work
While highly effective, PGEMs rely on a piece-wise constant assumption. This means the influence of a parent event is strictly "on" or "off" within the learned window. Consequently, it does not naturally model decaying influences, such as those seen in a Hawkes process where an event's effect gradually fades over time.
Future work may involve extending these models to handle more complex underlying temporal distributions or exploring how these learned causal relationships adapt over time in non-stationary environments.
Summary updated