Resource-aware mining with variable granularities in data streams

Abstract

For data stream applications, both approximation and adaptability are important issues for effective mining. We explore in this paper a fundamental problem that how the limited resources, e.g., memory space and computation power, can be well utilized to produce accurate estimates. Two important features for tracking mined patterns with properly utilized resources are examined. The first issue is temporal granularity which refers to the phenomenon that as time advances, people are more interested in recent events, meaning that more resources can be utilized to explore more recent data with finer granularities. Second, with the mining task of discovering frequent temporal patterns, more resources are expected to be allocated to the processing of those borderline patterns whose statistics, e.g., occurrence frequencies, are close to the specified threshold so as to have proper frequent itemset identification. This feature is called mining with support count granularity. Consequently, algorithm RAM-DS (Resource-Aware Mining for Data Streams) is designed to not only reduce the memory required for data storage but also retain good approximation of target time series. Experimental results have shown that the memory required for storing significant wavelet coefficients is very small and the quality of approximation is stable when performing incremental data updates, indicating that algorithm RAM-DS is feasible and suitable for adaptive mining in data streams.