Finding Relevant Causes in Complex Systems: a Generic Method Adaptable to Users and Contexts

Complex Adaptive Systems (CAS) feature intricate dynamics that are difficult to track and control. This paper addresses the challenge of providing causal explanations for unwanted or unexpected events in CAS, e.g. “why did this happen?”. Previous work in cognitive science indicates that satisfactory explanations should identify relevant causes tailored to the user's query and context. While automated methods exist for generating causal explanations, they typically overlook relevance or rely on a static notion of “best cause”. We propose a novel two-step methodology to provide relevant causes: (1) identify a wide range of causes; (2) rank and filter them using a flexible evaluation framework that can be tailored to users. We rely on an existing cause-detection method (1); and focus on selecting the most relevant ones (2). We propose three relevance metrics and a flexible method for combining them. We validate our framework using a flocking simulation where agents encounter obstacles. Experiments demonstrate how the flexibility of our method allows us to generate diverse causal explanations, each highlighting the most relevant aspects to the user.