This thesis is concerned with the merging of two active research domains: Knowledge Discovery in Databases (KDD), more precisely the Association Rule Mining technique, and Knowledge Engineering (KE) with a main interest in knowledge representation languages developed around the Semantic Web. In Data Mining, the usefulness of association rule technique is strongly limited by the huge amount and the low quality of delivered rules. Experiments show that rules become almost impossible to use when their number exceeds 100. At the same time, nuggets are often represented by those rare (low support) unexpected association rules which are surprising to the user. Unfortunately, the lower the support is, the larger the volume of rules becomes. Thus, it is crucial to help the decision maker with an efficient technique to reduce the number of rules. To overcome this drawback, several methods have been proposed in the literature such as itemset concise representations, redundancy reduction, filtering, ranking and post-processing. Even though rule interestingness strongly depends on user knowledge and goals, most of the existing methods are generally based on data structure. For instance, if the user looks for unexpected rules, all the already known rules should be pruned. Or, if the user wants to focus on specific family of rules, only this subset of rules should be selected. In this context, we address two main issues: the integration of user knowledge in the discovery process and the interactivity with the user. The first issue requires defining an adapted formalism to express user knowledge with accuracy and flexibility such as ontologies in the Semantic Web. Second, the interactivity with the user allows a more iterative mining process where the user can successively test different hypotheses or preferences and focus on interesting rules. The main contributions of this work can be summarized as follows: (i) A model to represent user knowledge. First, we propose a new rule-like formalism, called Rule Schema, which allows the user to define his/her expectations regarding the rules through ontology concepts. Second, ontologies allow the user to express his/her domain knowledge by means of a high semantic model. Last, the user can choose among a set of Operators for interactive processing the one to be applied over each Rule Schema (i.e. pruning, conforming, unexpectedness, . . . ). (ii) A new post-processing approach, called ARIPSO (Association Rule Interactive Post-processing using rule Schemas and Ontologies), which helps the user to reduce the volume of the discovered rules and to improve their quality. It consists in an interactive process integrating user knowledge and expectations by means of the proposed model. At each step of ARIPSO, the interactive loop allows the user to change the provided information and to reiterate the post-processing phase which produces new results. (iii) The implementation in post-processing of the proposed approach. The developed tool is complete and operational, and it implements all the functionalities described in the approach. Also, it makes the connection between different elements like the set of rules and rule schemas stored in PMML/XML files, and the ontologies stored in OWL files and inferred by the Pellet reasoner. (iv) An adapted implementation without post-processing, called ARLIUS (Association Rule Local mining Interactive Using rule Schemas), consisting in an interactive local mining process guided by the user. It allows the user to focus on interesting rules without the necessity to extract all of them, and without minimum support limit. In this way, the user may explore the rule space incrementally, a small amount at each step, starting from his/her own expectations and discovering their related rules. (v) The experimental study analyzing the approach efficiency and the discovered rule quality. For this purpose, we used a real-life and large questionnaire database concerning customer satisfaction. For ARIPSO, the experimentation was carried out in complete cooperation with the domain expert. For different scenarios, from an input set of nearly 400 thousand association rules, ARIPSO filtered between 3 and 200 rules validated by the expert. Clearly, ARIPSO allows the user to significantly and efficiently reduce the input rule set. For ARLIUS, we experimented different scenarios over the same questionnaire database and we obtained reduced sets of rules (less than 100) with very low support.
| Identifer | oai:union.ndltd.org:CCSD/oai:tel.archives-ouvertes.fr:tel-00912580 |
| Date | 26 October 2010 |
| Creators | Marinica, Claudia |
| Publisher | Université de Nantes |
| Source Sets | CCSD theses-EN-ligne, France |
| Language | English |
| Detected Language | English |
| Type | PhD thesis |
Page generated in 0.0021 seconds