Lazy meta-learning: creating customized model ensembles on demand

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