Abstract:
"
Pulsar stars are rare type of remnants from the core of extinguished stars that emit two
beams of energy per rotation. Energy beams of pulsar stars are detectable through radio
telescopes. However, such telescopes also detect other signals that could be considered
as noise such as background internet radio waves. Filtering pulsar signals among noise
has previously been done manually. The development of telescopes resulted in an
increase in the consolidation of observed data. Machine learning algorithms were used
successfully to classify pulsar star candidates. Usual batch learning algorithms were
effective for early surveys. Recent pulsar data collections roughly have a velocity of
nearly 0.5-1 TB of information per second. Thus, the need of developing algorithms
that can work with data streams was observed eliminating the need of store and classify
data. Subsequently, stream classification algorithms were applied to address the issue
of big data. This leads to the crux of this research.
Stream classification algorithms are specially developed to match with the
characteristics of data streams such as concept drift and limited memory. Extremely
Fast Decision Tree is one of the stream classification algorithms that can learn
incrementally when it meets new data. Learning incrementally from a pulsar candidate
stream should not be harmed by imbalances of candidate data. This dissertation is a
result of the research of improving Extremely Fast Decision Tree, for imbalanced data
streams. The resulted algorithm, Gaussian Hellinger Extremely Fast Decision Tree
(GH-EFDT) is accurate, fast and avoids the pitfalls of class imbalance and concept drift.
Benchmarking with existing stream classification algorithms for candidate selection,
GH-EFDT perform better with the two standard datasets HTRU1 and HTRU2. Not only
for the pulsar candidate selection, GH-EFDT performs equally with similar algorithms
for other imbalanced data streams. The developed system using GH-EFDT for the
candidate selection, Asips, is a web application with number of other features and a
user-friendly GUI. Also, these features are available as an API called Asips-for-Pulsar Astronomy.
"
