Understanding the bigger picture: Batch-free exploration of streaming sequential patterns with accurate prediction

Finding sequential patterns in data streams has been an attractive research topic recently. Available approaches are able to bound the error of found patterns by using a static PrefixSpan approach. This usage forced a batch-based method to divide the stream into manageable chunks. However, discovering sequential patterns within batches of a stream encounters additional errors when compared to the continuous, non-batch way. First, a lot of patterns contain items from two consecutive batches and thus will be lost when each batch is processed individually. Second, some patterns may not be frequent in one batch, and thus will be pruned, even though they will appear frequently when considering multiple batches. In this paper, we present the BFSPMiner, a Batch-Free Sequential Pattern Miner algorithm that accurately explores patterns in streaming data. The proposed algorithm can efficiently find useful frequent patterns that are otherwise lost when applying batch-based approaches. In addition to addressing the above-mentioned issues, we show through extensive evaluations over multiple real-world datasets the high predictability of found patterns when compared with those generated from state-of-the-art batch-based algorithms.