New scalable machine learning methods: beyond classification and regression

The recent surge in data available has spawned a new and promising age of machine learning. Success cases of machine learning are arriving at an increasing rate as some algorithms are able to leverage immense amounts of data to produce great complicated predictions. Still, many algorithms in the toolbox of the machine learning practitioner have been render useless in this new scenario due to the complications associated with large-scale learning. Handling large datasets entails logistical problems, limits the computational and spatial complexity of the used algorithms, favours methods with few or no hyperparameters to be con gured and exhibits speci c characteristics that complicate learning. This thesis is centered on the scalability of machine learning algorithms, that is, their capacity to maintain their e ectivity as the scale of the data grows, and how it can be improved. We focus on problems for which the existing solutions struggle when the scale grows. Therefore, we skip classi cation and regression problems and focus on feature selection, anomaly detection, graph construction and explainable machine learning. We analyze four di erent strategies to obtain scalable algorithms. First, we explore distributed computation, which is used in all of the presented algorithms. Besides this technique, we also examine the use of approximate models to speed up computations, the design of new models that take advantage of a characteristic of the input data to simplify training and the enhancement of simple models to enable them to manage large-scale learning. We have implemented four new algorithms and six versions of existing ones that tackle the mentioned problems and for each one we report experimental results that show both their validity in comparison with competing methods and their capacity to scale to large datasets. All the presented algorithms have been made available for download and are being published in journals to enable practitioners and researchers to use them.