Spatial-temporal Motif Discovery with Variable-size Sliding Windows

Motif discovery in spatial-temporal data is very important to the comprehension of some specific phenomena in areas such as weather prediction, wind generation, and image recognition. Most motif discovery methods are generally based on support and similarity. Among them, similarity-based Matrix Profile has been adopted in domains recently. However, it generally requires defining motif length in advance, which greatly harms its applicable scenarios. Moreover, unequal lengths of subsequences can be regarded as an informative feature for mining. In this paper, we propose a Sliding Window-based Motif Discovery algorithm (SWMD), which adopts a variable-size sliding window to accurately locate motifs by detecting targets in the generated distance matrix. To achieve higher efficiency, a novel pruning strategy is proposed, where its optimal parameters are achieved through a sensitivity test. Since the proposed algorithm is fit in processing multiple dimensional data, we apply it to discover motifs from a typical and natural spatial-temporal dataset, i.e., flood. Comparative studies with three methods show that the proposed algorithm could find more meaningful motifs within an acceptable time expense.