Maintenance for Case Streams: A Streaming Approach to Competence-Based Deletion

Abstract

The case-based reasoning community has extensively studied competence-based methods for case base compression. This work has focused on compressing a case base at a single point in time, under the assumption that the current case base provides a representative sample of cases to be seen. Large-scale streaming case sources present a new challenge for competence-based case deletion. First, in such contexts, it may be infeasible or too expensive to maintain more than a very small fraction of the overall set of cases, and the current system snapshot of the cases may not be representative of future cases, especially for domains with concept drift. Second, the interruption of processing required to compress the full case base may not be practical for large case bases in real-time streaming contexts. Consequently, such settings require maintenance methods enabling continuous incremental updates and robust to limited information. This paper presents research on addressing these problems through the use of sieve streaming, a submodular data summarization method developed for streaming data. It demonstrates how the approach enables the maintenance process to trade off between maintenance cost and competence retention and assesses its performance compared to baseline competence-based deletion methods for maintenance. Results support the benefit of the approach for large-scale streaming data.