A new framework for mining weighted periodic patterns in time series databases

摘要

Mining periodic patterns in time series databases is a daunting research task that plays a significant role at decision making in real life applications. There are many algorithms for mining periodic patterns in time series, where all patterns are considered as uniformly same. However, in real life applications, such as market basket analysis, gene analysis and network fault experiment, different types of items are found with several levels of importance. Again, the existing algorithms generate huge periodic patterns in dense databases or in low minimum support, where most of the patterns are not important enough to participate in decision making. Hence, a pruning mechanism is essential to reduce these unimportant patterns. As a purpose of mining only important patterns in a minimal time period, we propose a weight based framework by assigning different weights to different items. Moreover, we develop a novel algorithm, WPPM (Weighted Periodic Pattern Mining Algorithm), in time series databases underlying suffix trie structure. To the best of our knowledge, ours is the first proposal that can mine three types of weighted periodic pattern, (i.e. single, partial, full) in a single run. A pruning method is introduced by following downward property, with respect of the maximum weight of a given database, to discard unimportant patterns. The proposed algorithm presents flexibility to user by providing intermediate unimportant pattern skipping opportunity and setting different starting positions in the time series sequence. The performance of our proposed algorithm is evaluated on real life datasets by varying different parameters. At the same time, a comparison between the proposed and an existing algorithm is shown, where the proposed approach outperformed the existing algorithm in terms of time and pattern generation. (C) 2017 Elsevier Ltd. All rights reserved.