Current students


Section: Computer Science and Engineering
Nowadays, data come in large volumes and high variety. Streaming data are highly subjective to changes and a model trained on these data needs to be continuously adapted in order to keep high predictive performances. However, to properly change the model it is important to understand the characteristics of the drifts of the data, when the drift occurs, where it occurs and how. The monitoring on data distribution can be done on the received input and on the associated labels or on their joint distribution. According to the ways data change, the model can adapt in different ways. Sometimes it may be useful to have a detection mechanism that signals changes in the distributions of data and triggers adaptation of the model: these are called active approaches. They can be distinguished from passive approaches that continuously update the model by adjusting it on a sample-level or batch-level. 
This thesis addresses the problem of capturing the best way the model can handle these changes. It considers the problem in the supervised, semi-supervised, and reinforcement learning setting. An important topic that is addressed is the change of model complexity and type of model through the non-stationary changes. A special focus is also devoted to the ability of the system to recognize previously seen concepts and thus exploit prior knowledge to improve model performances. This can be also done through the learning of concepts' metadata such as its average duration and the way it switches to a new concept: learning this aspect allows to not only detect changes after they occur but also to predict when it will happen.