Time Series Discretization via MDL-based Histogram Density Estimation

In knowledge discovery from real-valued time series, discretization is often a key preprocessing that extends the applicability of sophisticated tools for symbolic data mining or logic-based machine learning. For finding meaningful discrete values that can be directly translated into some intuitive symbols, this paper proposes a novel discretization method based on density estimation using a two-dimensional (measurement vs. time) histogram of variable-width bins. We extend Kontkanen and Myllymaki's histogram construction method into our two-dimensional case, keeping the efficiency brought by dynamic programming. Experimental results with artificial and real datasets show the robustness and the usefulness of the proposed method.