| Summary: | The data stream mining problem has been studied extensively in recent years, due to the greatease in collection of stream data. The essential to a data stream mining algorithms is that we can only read data once. Unfortunately, most of traditional data mining algorithms do not have such single-scan property. Usually, data stream is considered as semi-in¯nite. It is impossible
to store all the past data with limited resources. Thus, mining high dimensional data streams
is a challenging task. In this report, we are going to propose some interesting observations on
feature quality stream(FQS), which is obtained from data stream in real time, and a frame-
work to analyze such stream. The analysis results of FQS are used to reduce the dimension of
data streams. We will also propose a data stream mining framework called MR-Stream. It is
a e±cient data stream clustering framework with the following properties: (1) computes and
updates synopsis information in constant time; (2) allows users to discover clusters at multiple
resolutions; (3) determines the right time for users to generate clusters from the synopsis in-
formation; (4) generates clusters of higher purity than existing algorithms; and (5) determines
the right threshold function for density-based clustering based on the fading model of stream
data. MR-Stream can be extend to solve classi¯cation problem. The classi¯cation results ob-
tained from the online component of MR-Stream framework are in realtime. The result given
by MR-Stream is presented as a probability distribution table over di®erent classes.
|
|---|