Towards Neural Situation Evolution Modeling: Learning a Distributed Representation for Predicting Complex Event Sequences

Abstract

In real-world monitoring tasks, a situation can be understood as a sequence of causally related events of interest. In road traffic control, such a situation could be a rear-end collision at the end of a traffic jam, which worsens congestion and requires clearing operations and potentially rerouting. Whereas conventional event sequence prediction focuses on sequences of individual events $\langle e_{1}, \ldots, e_{n}\rangle$ , evolving situations thus can be conceived as sequences of states composed of multiple concurrent events, i.e., complex events: $\langle\{e_{1},\ldots, e_{m}\}, \ldots, \{e_{l},\ldots, e_{n}\}\rangle$ . Situation (evolution) prediction thus requires learning a transition model for these complex events to provide the expectations for potential successor event types. In previous work, this was represented by a Markov Chain defined on the observed complex events. However, using the entire event composite as “atomic” situation state representation does not allow capturing patterns between its individual events (e.g., events of type “accident” share similar successor event types across different event composites), nor generalizing behaviors between similar event types or incorporating additional features. Hence, we propose a neural modeling approach to learn a distributed representation of a given situation dataset. By encoding the input states as conjunction of their individual comprised events, the devised model can learn associations (i.e., enable an “information flow”) between individual event types, allowing to capture similar behaviors across different situations. We test our approach on both synthetic and real-world datasets.