Download files
Abstract
Advances of computational power, data collection and storage techniques are making new data available every day. This situation has given rise to hypothesis generation research, which complements conventional hypothesis testing research. Hypothesis generation research adopts techniques from machine learning and data mining to autonomously uncover causal relations among variables in the form of previously unknown hidden patterns and models from data. Those patterns and models can come in different forms (e.g. rules, classifiers, clusters, causal relations).
In some situations, data are collected without prior supposition or imposition of a specific research goal or hypothesis. Sometimes domain knowledge for this type of problem is also limited. For example, in sensor networks, sensors constantly record data. In these data, not all forms of relationships can be described in advance. Moreover, the environment may change without prior knowledge. In a situation like this one, hypothesis generation techniques can potentially provide a paradigm to gain new insights about the data and the underlying system.
This thesis proposes a general hypothesis generation framework, whereby assumptions about the observational data and the system are not predefined. The problem is decomposed into two interrelated sub-problems: (1) the associative hypothesis generation problem and (2) the causal hypothesis generation problem. The former defines a task of finding evidence of the potential causal relations in data. The latter defines a refined task of identifying casual relations.
A novel association rule algorithm for continuous domains, called functional association rule mining, is proposed to address the first problem. An agent based causal search algorithm is then designed for the second problem. It systematically tests the potential causal relations by querying the system to generate specific data; thus allowing for causality to be asserted.
Empirical experiments show that the functional association rule mining algorithm can uncover associative relations from data. If the underlying relationships in the data overlap, the algorithm decomposes these relationships into their constituent non-overlapping parts. Experiments with the causal search algorithm show a relative low error rate on the retrieved hidden causal structures.
In summary, the contributions of this thesis are:
(1) a general framework for hypothesis generation in continuous domains, which relaxes a number of conditions assumed in existing automatic causal modelling algorithms and defines a more general hypothesis generation problem;
(2) a new functional association rule mining algorithm, which serves as a probing step to identify associative relations in a given dataset and provides a novel functional association rule definition and algorithms to the literature of association rule mining;
(3) a new causal search algorithm, which identifies the hidden causal relations of an unknown system on the basis of functional association rule mining and relaxes a number of assumptions commonly used in automatic causal modelling.