A Multi-Channel Neural Graphical Event Model with Negative Evidence

Tian Gao, Dharmashankar Subramanian, Karthikeyan Shanmugam, Debarun Bhattacharjya, Nicholas Mattei

Abstract

The Multi-Channel Neural Graphical Event Model (MCN-GEM) is a non-parametric deep learning framework designed to model complex, irregular event streams. By incorporating "fake epochs" to capture negative evidence—the information gained from the absence of events—the model learns history-dependent arrival rates without assuming rigid functional forms like exponential decay. Experimental results on synthetic and real-world datasets demonstrate that MCN-GEM significantly outperforms existing state-of-the-art models in log-likelihood and intensity modeling.

Introduction

Many real-world systems, ranging from electronic health records to social media interactions, generate "event streams": sequences of diverse event types occurring at irregular intervals. Modeling these streams requires estimating conditional intensity functions, which describe the instantaneous rate of event occurrence based on past history.

Traditional approaches, such as Hawkes Processes, face two primary challenges. First, they often rely on parametric assumptions, assuming that the influence of past events decays exponentially. However, real-world dependencies can be delayed or follow complex non-linear patterns. Second, these models typically update their internal state only when an event occurs, effectively ignoring the "negative evidence" provided by periods of silence.

Methodology

MCN-GEM addresses these issues through a fully data-driven, non-parametric architecture that combines Recurrent Neural Networks (RNNs) with a novel attention mechanism.

Multi-Channel Architecture and Attention

Instead of a single hidden state, the model partitions the hidden vector into "channels," where each channel tracks the state of a specific event type. A spatio-temporal attention mechanism maintains a memory bank of past hidden states, determining which previous events (spatial/type-based) and specific points in time (temporal/lag-based) are most influential for the current arrival rate.

The MCN-GEM architecture, illustrating the integration of fake epochs (red path) to capture negative evidence.

Negative Evidence via Fake Epochs

The core innovation of MCN-GEM is the injection of "fake epochs" (auxiliary labels) at uniform intervals between real events. These fake epochs force the LSTM to update its hidden state during periods of inactivity. This allows the model to approximate the continuous-time evolution of the system and more accurately calculate the probability that no event occurred during an interval.

Experimental Evaluation

The authors evaluated MCN-GEM across three distinct data domains:

  1. Synthetic Datasets: Five Proximal Graphical Event Models (PGEMs) were used to generate ground-truth streams.
  2. Political News (ICEWS): Socio-political event data from five South American countries.
  3. Healthcare (MIMIC-II): Clinical visit records and diagnoses from intensive care patients over seven years.

Key Results and Insights

MCN-GEM consistently outperformed state-of-the-art baselines, including the Neural Hawkes Process (NHP). On synthetic data, the model improved log-likelihood scores by 20% to 35%. On real-world datasets (ICEWS and MIMIC-II), it showed improvements between 5% and 20%.

Visualizing the intensity landscape: MCN-GEM accurately identifies event clusters and periods of inactivity.

A critical insight from the ablation studies was the "Power of One": adding just a single fake epoch in the middle of an inter-event interval provides a massive boost in model fitting. Furthermore, the non-parametric nature of the model allows it to learn "delayed excitation"—where an event's influence is only felt after a specific lag—which standard Hawkes processes cannot represent.

Conclusion

MCN-GEM provides a robust framework for modeling irregular event streams by effectively utilizing negative evidence. While the injection of fake epochs introduces some computational overhead, the resulting "sharper" intensity landscapes and improved log-likelihoods represent a significant advancement over parametric and standard neural point process models. Future work will focus on optimizing these mechanisms for real-time forecasting.

Summary updated

Abstract

Event datasets are sequences of events of various types occurring irregularly over the time-line, and they are increasingly prevalent in numerous domains. Existing work for modeling events using conditional intensities rely on either using some underlying parametric form to capture historical dependencies, or on non-parametric models that focus primarily on tasks such as prediction. We propose a non-parametric deep neural network approach in order to estimate the underlying intensity functions. We use a novel multi-channel RNN that optimally reinforces the negative evidence of no observable events with the introduction of fake event epochs within each consecutive inter-event interval. We evaluate our method against state-of-the-art baselines on model fitting tasks as gauged by log-likelihood. Through experiments on both synthetic and real-world datasets, we find that our proposed approach outperforms existing baselines on most of the datasets studied.
BibTeX
@inproceedings{gao2020multi,
  author = {Tian Gao and Dharmashankar Subramanian and Karthikeyan Shanmugam and Debarun Bhattacharjya and Nicholas Mattei},
  title = {{A Multi-Channel Neural Graphical Event Model with Negative Evidence}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i04.5810},
}
Figure