An Improved Piecewise Aggregate Approximation Based on Statistical Features for Time Series Mining

摘要

Piecewise Aggregate Approximation (PAA) is a very simple dimensionality reduction method for time series mining. It minimizes dimensionality by the mean values of equal sized frames, which misses some important information and sometimes causes inaccurate results in time series mining. In this paper, we propose an improved PAA, which is based on statistical features including a mean-based feature and variance-based feature. We propose two versions of the improved PAA which have the same preciseness except for the different CPU time cost. Meanwhile, we also provide theoretical analysis for their feasibility and prove that our method guarantees no false dismissals. Experimental results demonstrate that the improved PAA has better tightness of lower bound and more powerful pruning ability.