New perspectives and methods for stream learning in the presence of concept drift.
López Lobo, Jesús
MetadataShow full item record
Applications that generate data in the form of fast streams from non-stationary environments, that is,those where the underlying phenomena change over time, are becoming increasingly prevalent. In thiskind of environments the probability density function of the data-generating process may change overtime, producing a drift. This causes that predictive models trained over these stream data become obsoleteand do not adapt suitably to the new distribution. Specially in online learning scenarios, there is apressing need for new algorithms that adapt to this change as fast as possible, while maintaining goodperformance scores. Examples of these applications include making inferences or predictions based onfinancial data, energy demand and climate data analysis, web usage or sensor network monitoring, andmalware/spam detection, among many others.Online learning and concept drift are two of the most hot topics in the recent literature due to theirrelevance for the so-called Big Data paradigm, where nowadays we can find an increasing number ofapplications based on training data continuously available, named as data streams. Thus, learning in nonstationaryenvironments requires adaptive or evolving approaches that can monitor and track theunderlying changes, and adapt a model to accommodate those changes accordingly. In this effort, Iprovide in this thesis a comprehensive state-of-the-art approaches as well as I identify the most relevantopen challenges in the literature, while focusing on addressing three of them by providing innovativeperspectives and methods.This thesis provides with a complete overview of several related fields, and tackles several openchallenges that have been identified in the very recent state of the art. Concretely, it presents aninnovative way to generate artificial diversity in ensembles, a set of necessary adaptations andimprovements for spiking neural networks in order to be used in online learning scenarios, and finally, adrift detector based on this former algorithm. All of these approaches together constitute an innovativework aimed at presenting new perspectives and methods for the field.