A divide and conquer approach to using inductive logic programming for learning user models

The use of Inductive Logic Programming (ILP) for the construction of a user model allows possibilities such as the use of intensional background knowledge and the ability to present a set of comprehensible rules to the user. A few simple concepts have previously been modelled using standard ILP methods but a concept of the complexity attempted within this dissertation brings problems which have not been attempted within the usermodelling context. This dissertation describes the problems encountered (and their solutions) when using ILP to build a user model that will be used to make detailed predictions of successive entries within an individual user’s online diary. Each predicted entry must contain several values — not just the type of task being performed, but also its location and duration — thus increasing the complexity of the clauses required. Other problems to be dealt with are the vague nature of the underlying concept to be modelled and the possibility of high levels of noise within the data collected from the user. The complexity of the clauses required and the large amount of background knowledge that could be used renders the size of the search space of possible hypotheses too large for an exploratory ILP engine such as Aleph to search adequately within the limited time available. The methods required to overcome these problems are presented within the algorithm Dilum (Dimensional ILP for User Modelling). Dilum pre-processes its given data set to split the concept to be learnt into separate subconcepts, and uses a divide-and-conquer approach when working with each subconcept to reduce the complexity of each learning problem to a more manageable level. The main contribution within this disseration is the use of these two techniques to significantly reduce the amount of computation time required for the kind of learning problem encountered within the scenario described earlier. It will be shown within this dissertation that the ideas encapsulated within Dilum enable construction of an accurate, usable, and understandable user-model, and that if the concept to be learnt is both vague and complex then even a greedy learning algorithm will require partitioning as a data pre-processing stage. It will also be shown that using conditional probability distributions as a measure of uncertainty whilst querying the user model not only helps to reduce the effects of noisy data on the accuracy of the predictions made by the model but also increases its precision.