Abstract:
The scarcity of labeled time-series data is a major
challenge in use of deep learning methods for Time Series
Classification tasks. This is especially important for the growing
field of sensors and Internet of things, where data of high dimensions and complex distributions coming from the numerous field
devices has to be analyzed to provide meaningful applications.
To address the problem of scarce training data, we propose
a heuristic combination of deep transfer learning and deep
active learning methods to provide near optimal training abilities
to the classification model. To mitigate the need of labeling
large training set, two essential criteria – informativeness and
representativeness have been proposed for selecting time series
training samples. After training the model on source dataset,
we propose a framework for the model skill transfer to forecast
certain weather variables on a target dataset in an homogeneous
transfer settings. Extensive experiments on three weather datasets
show that the proposed hybrid Transfer Active Learning method
achieves a higher classification accuracy than existing methods,
while using only 20% of the training samples