Jesse Read will defend his habilitation thesis on Wednesday November 8th at 14h00, in room Henri Poincaré in the Alan Turing Building.

Abstract: In multi-label classification, multiple target variables are modelled, and multiple values predicted for each instance, as opposed to the traditional learning problem where each instance is associated with a single target variable. The main challenge is detecting and dealing with dependencies among labels, and incorporating these into learning methods, which should be computationally tractable to large problems. This habilitation thesis elaborates on new approaches for tackling this challenge, and also extending the challenge to the context of data streams. In data streams instances arrive continuously in a theoretically-infinite stream. A particular focus of this work is the case where temporal dependence is involved. A number of methods and metrics are proposed to overcome numerous challenges encountered in these scenarios.