Sarah D'Ettorre of Microsoft

In today's world, most data are collected in streams of continuously flowing instances with some temporal nature. Many industries, therefore, are faced with solving dynamic Machine Learning problems; problems where models must adapt to changing patterns, or concept drift, in order to remain useful. A common strategy is to predefine some arbitrary or domain informed retrain cadence, at which time a new model is trained from scratch on some amount of recent data. The trouble with this solution is that it is very uninformed (it has no information about when patterns actually change), prone to error (it has no information about what types of drifts occurred or how to appropriately adapt) and can be inefficient or even completely impractical. Fear not; there is a better way. Enter adaptive learning. Adaptive learning strategies learn and update incrementally on each instance, or in batches, with user in the loop or automatically where appropriate. This keeps models up to date and relevant. Additionally, concept drift detection mechanisms can inform the system when and how it should adapt, depending on concept drift characteristics like magnitude and duration. These solutions ensure that models remain relevant through time, reduce manual analysis during retrain and increase the longevity of solutions. This talk will discuss data streams, concept drift and concept drift detection methods. It will walk through various model adaption strategies, and finally, online evaluation techniques.

Biography:

Sarah D'Ettorre is a Senior Machine Learning Engineer on the Commercial Software Engineering team at Microsoft. She and her team work with Microsoft's top strategic customers to help them innovate and compete in their domains. Sarah's area of expertise is in online, adaptive learning and time series. She has lead various projects including credit card fraud detection, churn prediction, demand forecasting and user behaviour pattern discovery. Sarah completed her Master of Computer Science specializing in Machine Learning at the University of Ottawa, Canada. There, she and her colleagues developed a novel, unsupervised Machine Learning algorithm to rapidly detect evolving concepts in streaming data, and appropriately adapt the model. She is passionate about all things ML, but remains especially captivated by adaptive learning.