Abstract
Graphical event models (GEMs) are representations of temporal point process dynamics between different event types. Many real-world applications however involve limited event stream data, making it challenging to learn GEMs from data alone. In this paper, we introduce approaches that can work together in a score-based learning paradigm, to augment data with potentially different types of background knowledge. We propose novel scores for learning an important parametric class of GEMs; in particular, we propose a Bayesian score for leveraging prior information as well as a more practical simplification that involves fewer parameters, analogous to Bayesian networks. We also introduce a framework for incorporating easily assessed qualitative background knowledge from domain experts, in the form of statements such as 'event X depends on event Y' or 'event Y makes event X more likely'. The proposed framework has Bayesian interpretations and can be deployed by any score-based learner. Through an extensive empirical investigation, we demonstrate the practical benefits of background knowledge augmentation while learning GEMs for applications in the low-data regime.
Introduction
Learning the causal and temporal dependencies between different types of events is a fundamental task in fields ranging from clinical healthcare to social media analysis. These dependencies are often modeled using Temporal Point Processes (TPPs) and represented via Graphical Event Models (GEMs). However, structure learning for GEMs is notoriously data-intensive. In many practical scenarios, researchers face "low-data regimes" where only a few event streams are available, making purely statistical approaches prone to noise and overfitting.
This work addresses this challenge by introducing a framework to augment sparse datasets with qualitative background knowledge. By integrating "soft" human expertise—such as an expert's belief that one event type influences another—into the "hard" statistical learning process, we can significantly improve model accuracy in data-scarce environments.
Methodology
The proposed framework utilizes a score-based learning paradigm, introducing two primary mechanisms for knowledge injection.
1. Bayesian Gamma (BG) Score
The authors derive a novel Bayesian score specifically for piece-wise constant GEMs. By employing the Gamma distribution as a conjugate prior for event intensities, the scoring function naturally incorporates prior beliefs about event rates. To enhance practical usability, a simplified version called the BGPi score is introduced. This score requires only two intuitive hyperparameters: the expected time duration to observe a single event ($D'$) and a complexity penalty parameter ($\kappa$).
2. Incompatibility Framework
To handle qualitative statements (e.g., "Event Y inhibits Event X"), the authors propose an incompatibility framework. This method modifies the standard log-likelihood ($LL$) of the model by adding a penalty term: $LL' = LL - w \cdot f(C, \mathcal{M})$, where $w$ represents the confidence weight of the background knowledge, and $f(C, \mathcal{M})$ quantifies the model's violation of the provided constraints. This "soft" constraint approach allows the model to eventually override incorrect expert knowledge if the empirical data is sufficiently strong. Mathematically, providing an inequality constraint acts like an adjustment to the "effective duration" of the data, pushing estimated event rates in the desired direction.
Experimental Evaluation
The framework was evaluated across several synthetic and real-world datasets to test its robustness in low-data settings.
Datasets
- Synthetic PGEMs: Three distinct graph structures (ranging from sparse to dense) were used to test structure recovery performance.
- Real-World Data: The study utilized employment events from LinkedIn, electronic health records from Mimic-II, user engagement badges from Stack Overflow, and a Diabetes dataset containing blood glucose and insulin levels paired with 11 expert-assessed causal relations.
Key Results and Insights
The empirical results highlight the effectiveness of background knowledge augmentation:
- Performance in Low-Data Regimes: On synthetic data with very few streams (5–10), the BGPi score significantly outperformed traditional criteria like BIC and AIC, as well as neural-based baselines like CAUSE, in terms of F1 scores for structure recovery.
- Impact of Qualitative Input: Adding even a single correct parental statement per node (e.g., "A depends on B") dramatically improved performance. In a setting with only two event streams, the F1 score rose from 0.2 to nearly 0.8.
- Robustness to Incorrect Expertise: A critical finding is the framework's resilience to "wrong" experts. While incorrect knowledge initially degrades performance, the likelihood term eventually dominates as more data is collected, allowing the model to recover the true underlying structure.
- Synergy of Data and Knowledge: In the Diabetes dataset, the combination of data and 11 expert statements yielded the best results. Using expert statements alone resulted in a poor F1 score (0.27), demonstrating that the framework successfully leverages the synergy between human intuition and empirical evidence.
Conclusion
This paper demonstrates that augmenting score-based GEM learning with qualitative background knowledge is a powerful strategy for low-data scenarios. While the BGPi score and incompatibility framework provide robust tools for this purpose, future work could focus on automating hyperparameter estimation and extending the framework to handle more complex temporal logic.
Summary updated