Stream Mining via Density Estimators: A Concrete Application

Many real-world applications share the property that the data they process arrives in streams. The transient and volatile nature of these streams renders the application of common processing and analysis techniques difficult. In particular, the mining of streams has proved to be a difficult task due to the rigid processing requirements that must be met within the data stream scenario. We propose to exploit kernel density estimation for stream mining. Kernel density estimation as a technique from the area of mathematical statistics found its way into many mining related topics and applications. However, its heavy computational cost makes the direct application to streams impossible. On account of this, we developed in a previous work a sophisticated adaptation that continuously computes kernel density estimators over streams. By means of these estimators, we can tackle a variety of mining tasks over streaming data. We illustrate some of them against the background of a concrete application in a medical environment. More precisely, we will demonstrate a prototypical implementation of a medical monitor that visualizes an online analysis of the vital signs of patients. Besides demonstrating the monitor, we will also present the concepts underlying its implementation.