Instance Selection by Identifying Relevant Events Using Domain Knowledge and Minimal Human Involvement

Billions of events (image, video, tweet, purchase, delivery, or failure) are captured in the era of Big Data and used as training data in business analytic models. But, the event relevance and the consequent effect on a target variable are difficult to ascertain. Contemporaneous events, and not specified market latencies, might lead to a fuzzily defined training data set and poor classification results. Instance selection (IS) methods aim to choose relevant training examples, while reducing a training set, to a subset. Both goals contribute in Big Data projects. But, traditional approaches identified by a literature review are only heuristic and do not consider fuzzy effects and therefore fail. Thus, relevant instances must classified by experts. But, the content of events changes quickly and manual assignment to maintain a realistic model is time and cost intensive. We propose an alternative approach. Here, the relevance of an event is deduced from the effect it causes on a target variable after its publication. Such additional business domain knowledge can be expected to allow a more precise selection of instances and thus a successful prediction. An application in the natural gas market is presented that identified more relevant tickers than other approaches. This approach contributes to a scientific discussion by integrating domain knowledge into IS. Until now, only a few approaches have been introduced, all of which demand human involvement and are expensive. Our approach is automatic, thereby saving analysts' time and expense.