Design Principles of Event Mining Systems

to assist researchers, analysts, and decision makers in extracting knowledge from such a variety of data. As the speed and scale of this data generation will increase even further in the future, we require new frameworks that support highperformance computing, processing techniques that fuse heterogeneous data and uncover hidden patterns and unknown correlations, scalable software tools, and useful visualizations that help analysts and decision makers understand the results better. Modeling complex, mysterious, and at least partly unknowable systems involves many complicated decisions such as determining a model selection strategy, defin­ ing a model structure, defining a criteria for model goodness, selecting data and the transformations applied to the data, tuning learning parameters, and so on. Most of these decisions involve a reliance on theoretical or empirical results, that is, expert domain knowledge, and cannot be learned by a system itself solely from available input data. Moreover, many spurious associations might arise from learned models, resulting in false scientific discoveries and false statistical infer­ ences [Calude and Longo 2017]. A promising approach for modeling complex phe­ nomenon is to adopt a human-in-the-loop approach in the data processing step. This integrates high-level expert knowledge into the modeling process by acquiring an expert’s relevance judgments regarding a set of initial retrieval results. Despite the apparent benefits of such a perspective, frameworks that facilitate a seam­ less interaction between a domain expert and a traditional knowledge discovery process are not well studied. Figure 4.1 shows the human-in-the-loop in a modeling process rooted in event mining. Design Principles of Event Mining Systems