Piero P. Bonissone
IEEE Members: Free
Non-members: FreeDuration: 00:56:22
Abstract. In the not so distant future, we expect analytic models to become a commodity. We envision having access to a large number of data-driven mod- els, obtained by a combination of crowdsourcing, crowdservicing, cloud-based evolutionary algorithms, outsourcing, in-house development, and legacy mod- els. In this new context, the critical question will be model ensemble selection and fusion, rather than model generation. We address this issue by proposing customized model ensembles on demand, inspired by Lazy Learning. In our ap- proach, referred to as Lazy Meta-Learning, for a given query we find the most relevant models from a DB of models, using their meta-information. After retrieving the relevant models, we select a subset of models with highly uncorrelated errors. With these models we create an ensemble and use their me- ta-information for dynamic bias compensation and relevance weighting. The output is a weighted interpolation or extrapolation of the outputs of the models ensemble. Furthermore, the confidence interval around the output is reduced as we increase the number of uncorrelated models in the ensemble. We have successfully tested this approach in a power plant management application.