Similarity weighted ensembles for relocating models of rare events

Spatially distributed regions may have different influences that affect the underlying physical processes and make it inappropriate to directly relocate learned models. We may also be aiming to detect rare events for which we have examples in some regions, but not others. A novel method is presented for combining classifiers trained on regions with known sensor data and predicting rare events in new regions, specifically the closure of shellfish farms. The proposed similarity weighted ensemble method demonstrates an average 10 fold improvement in accuracy over One Class classification and 3 fold improvement over rules hand-crafted by an expert.