Time series data set is highly important to many business areas, that may be healthcare,
nance, or the industry. Often in these cases it is not the normal data that is important to
the users but instead the abnormal points. Etc. a doctor may not pay that much attention
to all the normal heartbeat at an electrocardiogram (ECG) but instead the cases where
the heartbeat is diering from normal. Therefor it is a huge research area to nd suitable
methods to nds these abnormal points.
These time series data set is often unsupervised meaning they do not contain any
labeled (information about the object state) training data sets to be used for classication.
That is the reason why normal classication can´t be used, in these cases and leads us
to the demand of a set of algorithms that can detect these abnormal points without that
kind of predened information.
This project will cover six dierent algorithms for unsupervised outlier detection. These
algorithms will in general base their ability to detect outliers on their ability to learn the
normal distribution of the data. And they will try to do this by dierent methods like:
• Try if some objects can be isolated faster than other objects.
• Using neural networks to learn the normal pattern.
• Using variational autoencoders to learn the normal pattern.