Neural Networks for Relational Data (Neurale netwerken voor relationele gegevens)
Author:
Abstract:
The topic of this dissertation is situated in the field of machine learning. In general, machine learning is concerned with the development of algorithms that can learn from empirical datasets containing input and target patterns. What will be considered within the context of this dissertation, is learning predictive models from relational data. While propositional data is represented by a single tuple for each input pattern, in relational data every input pattern consists of multiple tuples with relationships between the tuples. Learning approaches for this type of data should always take into account properties of the input pattern as a whole. This could be structural features, which are properties based on substructures of the input pattern, or aggregate features, which are properties based on subsets of the collection of input pattern tuples. The focus of this dissertation will be on aggregate features.Artificial neural networks are one particular method in the field of machine learning. Originally, they were developed as a method for simple data domains in which every input pattern is represented by a single input vector of fixed size. There has also been quite some interest in the use of neural networks for relational data domains, but most of this research has focused on learning structural features. The aspect of learning aggregate features with neural networks has received very little attention so far, although this topic might be more promising. Learning structural features with neural networks gives rise to a number of problems but learning aggregate features is a task that suits them very well. Therefore, a number of neural network approaches will be developed in this dissertation that deal with learning aggregate features from relational data.The first part of the dissertation focuses on methods for data containing only a single relation. This means that each input pattern can be represented as a single, homogenous collection of tuples, with all tuples of the same type. In the second part of the dissertation, methods for data containing multiple relations are considered. In this case, every input pattern is represented by multiple collections of tuples, with different types of tuples for different collections. For both cases, similar neural network methods are developed and empirically tested on a number of datasets. These approaches are compared, both from a theoretical and an experimental point of view, to related approaches such as multi-instance methods or neural networks for graph structures. The experimental results show that the new methods can be viable options for tackling specific kinds of problems and have a performance that is in many cases comparable to the performance achieved with other methods or in some cases even better.