Abstract:
|
Time is involved in almost every scienti c eld one can think on.
Observations of a phenomena are collected with the aim of study or
explain its behavior. This collections lead to organized data called time
series.
Data mining community has spent a reasonable amount of time
studying time series, in order to extract all meaningful knowledge from
them. Humans are generally good comparing time series, but still, our
capabilities are not scalable and we need to design algorithms and
techniques that allow us to deal with high dimensional data and other
problems.
In this work we will focus in a speci c problem, extracting valid
features of unlabeled time series obtained from aircraft sensors. These
must serve as a summary of a ight and they also must include relevant
details that serve to characterize it. This information will be used to
feed an algorithm which can learn to classify ights in groups, reducing
the number of necessary labeled data to obtain the desired accuracy
using an active learning approach. |