Features for Learning Local Patterns in Time-Stamped Data

Time-stamped data occur frequently in real-world databases. The goal of analysing time-stamped data is very often to find a small group of objects (customers, machine parts,...) which is important for the business at hand. In contrast, the majority of objects obey well-known rules and is not of interest for the analysis. In terms of a classification task, the small group means that there are very few positive examples and within them, there is some sort of a structure such that the small group differs significantly from the majority. We may consider such a learning task learning a local pattern. Depending on the goal of the data analysis, different aspects of time are relevant, e.g., the particular date, the duration of a certain state, or the number of different states. From the given data, we may generate features that allow us to express the aspect of interest. Here, we investigate the aspect of state change and its representation for learning local patterns in time-stamped data. Besides a simple Boolean representation indicating a change, we use frequency features from information retrieval. We transfer Joachim’s theory for text classification to our task and inestigate its fit to local pattern learning. The approach has been implemented within the MiningMart system and was successfully applied to real-world insurance data.