Learning Stable Concepts in Domains With Hidden Changes in Context

This paper presents Splice, a batch meta-learning system, designed to learn locally stable concepts in domains with hidden changes in context. The majority of machine learning algorithms assume that target concepts remain stable over time. In many domains this assumption is invalid. For example, nan-cial prediction, medical diagnosis, and network performance are domains in which target concepts may not remain stable. Unstable target concepts are often due to changes in a hidden context. Existing works on learning in the presence of hidden changes in context use an incremental learning approach.