Jan van Rijn
promotor: prof.dr. J.N. Kok (UL)
copromotors: dr. A.J. Knobbe (UL) and dr. J. Vanschoren (TU/e)
Date: 19 December, 2016, 12:30
We are surrounded by data. On a daily basis, we are confronted by many forms of it. Companies try to spread their commercials by means of billboards, commercials and online advertisements. We have instant access to our friends’ social lives using services as Facebook and Twitter, and we can obtain information about countless topics of interest by means of websites such as Wikipedia. In most cases, this is a double-edged sword. Companies and governments also collect information about us. For example, most websites store information about our browsing behaviour, banks know most about our financial transactions, and telecom providers even have access to our exact whereabouts, as our GPS coordinates are shared by our mobile phones.
Data is also gathered for scientific purposes. Large sensor networks and telescopes measure complex processes, happening around us on Earth or throughout the Universe. Any estimation of the amount of data that is being produced, transferred and gathered would be pointless, as it will be outdated some moments after publication.
All this data is valuable for the information, knowledge and eventually wisdom we could obtain from it. We could identify fraudulent transactions based on financial data, develop new medicines based on clinical data, or locate extraterrestrial life based on telescope data. This process is called learning. The scientific community has created many techniques for analysing and processing data. A traditional scientific tasks is modelling, where the aim is to describe the data in a simplified way, in order to learn something from it. Many data modelling techniques have been developed, based on various intuitions and assumptions. This area of research is called Machine Learning.
However, all data is different. For example, data about clinical trials is typically very sparse, but well-structured, whereas telescopes gather large amounts of data, albeit initially unstructured. We cannot assume that there is one algorithm that works for all sorts of data. Each algorithm has its own type of expertise. We have only little knowledge about which algorithms work well on what data.
The field of Machine Learning contains many challenging aspects. The data itself is often big, describing a complex concept. Algorithms are complex computer programs, containing many lines of code. In order to study the interplay between these two, we need data about the data and the algorithms. This data is called meta-data, and learning about the learning process itself is called meta-learning. It is possible to gain knowledge about the learning process when there is sufficient meta-data. Some effort has been devoted to building a large repository of this experimental data, called the ‘open experiment database’. It contains a large amount of publicly available Machine Learning results. This way, existing experimental data can be used to answer new research questions. Although this has proven extremely useful, there is still room for improvement. For example, sharing experiments was difficult: while all experimental data was accessible to the public, contributing new results towards the experiment database was only practically possible for a small circle of researchers. Furthermore, sensibly defining the types of meta-data that are being stored would expand the range of information and knowledge that can be obtained from the data. For example, storing all evaluation measures per cross-validation fold enables statistical analysis on the gathered data, and storing the individual predictions of the algorithms enables instance-level analysis. This thesis is about building upon the existing work of experiment databases, and demonstrate new opportunities for Machine Learning and meta-learning.