Recency-based sequential pattern mining in multiple event sequences

The standard sequential pattern mining scheme hardly considers the positions of events in a sequence, and therefore it is difficult to focus on more interesting patterns that represent better the causal relationships between events. Without quantifying how close two events are in a sequence, we may fail to evaluate how likely an event is caused by the others from the pattern, which is a severe drawback for some applications like prediction. Motivated by this, we propose the recency-based sequential pattern mining scheme together with a novel measure of pattern interestingness to effectively capture recency as well as frequency. To efficiently extract all the recency-based sequential patterns, we devise a mining algorithm, called Recency-based Frequent pattern Miner (RF-Miner), together with an effective prediction method to evaluate the quality of recency-based patterns in terms of their prediction power. The experimental results show that our RF-Miner algorithm can extract more diverse and important patterns that can be used to make prediction of the next event, and can be more efficiently performed by using the upper bounds of our measure than baseline algorithms.