A pattern based predictor for event streams

Recently, new emerging applications, such as web click-stream mining, failure forecast and traffic analysis, introduced a new challenging data model referred to as data streams. Mining such data can reveal up-to-date patterns, which are useful for predicting future events. Consequently, pattern mining in data streams is a popular field in data mining that presents unique challenges. The data is large and endlessly keeps on coming, making it impossible to store it, or to re-analyse historical data once it has been discarded. To solve this, we first present a novel method for mining sequential patterns from a data stream, in which we maximise memory usage in order to achieve higher accuracy in terms of results. In a second step, we use the discovered patterns in order to try to predict future events. We propose a number of ways to assign a score to each pattern in order to generate predictions. The prediction performance of these scoring strategies is then extensively experimentally evaluated. The predictor offers an opportunity for a faster detection and response to an important, though perhaps unexpected, event, which will occur in the future.