Adaptive Query Processing for Time-Series Data

Traditional query processing for time-series data is based on transforming data in the time domain into frequency domain. This approach however is less controllable by the end users, therefore isn't well suited for finding patterns with many dynamically specified user constraints. For many applications, there may not be any known patterns to search with. Instead, the users discover patterns by issuing queries with combinations of various dynamically specified constraints, such as the degrees of accuracy, the moving average window sizes, the aggregate time units, (e.g., total by day, by week, or by month) etc. We present a method to search time-series data that is adaptive to process queries with many dynamically specified constraints. Based on movements between contiguous data points in time series, our method first transforms time sequences into symbol strings of the discrete domain. Next, a suffix tree is built to index all suffixes of the symbol strings transformed form the time sequences. The focus of this paper is to present this method and to demonstrate how it can adapt to the processing of time-series queries with various dynamic constraints.