Reliable estimation of prediction confidence remains a significant challenge in machine learning. We usually expect past performance to indicate future performance. When we deal with risk-sensitive systems – where the cost of a bad decision can be very high, and prediction accuracy is not the only objective; we need a multidimensional perspective about the forecast models. So, if the user is given a confidence of each new predictions made by the model, then a more meaningful action can be taken.

In this talk we will discuss about Conformal Prediction Framework and how it can be leveraged across various machine learning algorithms used in Storage Industry (e.g. Disk Drive Failure detection and storage demand forecasting). Furthermore, as an example, we will describe how this framework can be translated to time-series, classification and regression problem which will give a confidence (indication of the quality of each prediction) and credibility (filter mechanism with which we can “reject” certain predictions).

Learning Outcomes

a. Conformal Prediction framework for reliable time-series forecasting and application in regression and classification problems

b. Fundamentals of On-line learning approach and handling concept drift

Presented by Supriya Kannery, Dell EMC

свернуть