关键词: 泛化均值, 滑动窗口, 数据流

Abstract: Data streams are data sources delivering indefinitely and at high rate new data， which causes computational and storage problems in the analysis of these sources. An inherent assumption in the analysis of a data stream is that the data stream distribution changes over time， which causes old data to be non-representative to the current data distribution. In case our objective is to obtain knowledge of the current data properties， analysis based on old and non-representative data will result in incorrect or incomplete knowledge of the current state. To cope with this problem， many data stream analyses are built on a fixed sliding window of the most “freshest” data points received. In this way， old data is forgotten and the space and time complexity of the analysis algorithm bounded. Due to the fact that no prior knowledge of the eventual changes is available， there exists no optimal window size and the quality of the analysis degrades. To increase the quality of analysis， it proposes some initial steps towards efficiently equipping sliding window algorithms with flexible windowing. Flexible windowing is obtained by the efficient maintenance of a statistic called， the maximum normalized mean. This statistic is used as a building block for algorithms in change detection and frequent item set mining. This paper shows that the maximum normalized mean can be maintained efficiently in time and space. Furthermore， it restates several algorithms in change detection and frequent item set mining， such that the algorithms execution is only dependent on the maintenance of the maximum normalized mean. And most importantly it investigates the gain in performance due to the flexible windowing over a fixed size window.

Key words: normalized mean, sliding window, data stream