Query Synthesis in One-Class Active Learning
|Termin||Fr 16. November 2018|
|Kurzfassung||Although machine learning plays an ever-increasing role in modern technology, there are still some parts where a human is needed to help with the learning process. With active learning a human operator is added to the process and helps with the classification of unknown samples. This improves the precision of the machine learning process. Although increasing the precision is important, the addition of a human operator introduces a problem known since the invention of the computer: Humans are slow compared to machines. Therefore it is essential to present the human operator with queries having the highest value of informativeness to optimize the learning process. The better the queries are chosen, the less time the learning process needs.
Current query selection strategies, use class label information to interpolate between opposite pairs at the decision boundary or select a query from a set of given unlabeled data points. However, these strategies cannot be used when no unlabeled and no negative observations are available. Then, one uses a query strategy function that rates the informativeness for any query candidate to synthesize the optimal query. While it is easy to calculate the most informative point in just a few dimensions, the curse of dimensionality quickly becomes a problem when searching for the most informative point in a high-dimensional space. This thesis takes a look at synthesizing queries in high-dimensional one-class cases via metaheuristics. The goal is to compare different metaheuristics experimentally with multiple data sets.