Finding Contrast Patterns for Mixed Streaming Data

Contrast set mining identifies patterns in the data that can best distinguish between groups. Most of the existing work focuses on categorical and batch data, and they do not scale well for large datasets. In this work, we focus on finding contrast patterns for mixed (quantitative and categorical) and streaming data. We adapt a discretization methodology, Supervised Dynamic and Adaptive Discretization, to identify meaningful bin boundaries. We then use the discretization result to find contrast patterns on streaming data. In order to achieve this, we identify frequent items and then contrast them to a group of interest. To handle potential concept drift, we propose an update strategy to keep the frequent items relevant. In addition, our algorithm samples feature combinations based on their "sampling" score and user feedback to reduce the search space as well as retrieving more interesting patterns.