Outlier Region Detection in Time Series

Outlier regions are subsequences of the time series whose dynamics are not consistent with the remaining data. Finding these regions is important in many applications so general-purpose algorithms need to be developed. Self-supervision, which to date is used primarily in computer vision, is advantageous, because it provides a training objective for neural networks as domain-specific feature extractors.
Self-supervision is a promising paradigm in machine learning. The data is transformed into carefully designed tasks on which predictive models can be trained. These can then be used as feature extractors, for outlier detection, or for other forms of data-driven decision making.
This thesis will explore self-supervision for time series data, specifically for the detec-tion of outlier regions. Existing work in computer vision is not readily applicable, and the following questions need to be addressed: (a) which self-supervision tasks are most suitable for time series data and take care of both local dynamic and global context, and are able to score individual time steps (b) can we use the principle of inlier priority to compute an inlier/outlier score for each time step (c) can we use a probabilistic reasoning about self-supervision tasks to combine the score of individual time points into inlier and outlier regions.
Finally, qualitative and quantitative evaluation criteria need to be developed: The thesis explores whether the learned representations and domain specific tasks can give qualitative insights about the data. Experiments on various time series datasets test quantitatively whether the approach outperforms existing outlier detection algorithms.