Article Preview
TopIntroduction
Knowledge Discovery in Databases (KDD) is a highly complex, iterative and interactive process, with a goal-driven and domain dependent nature (Fayyad et al., 1996). It involves three main steps (data preprocessing, data mining and post-processing) with many decisions made by the analyst (Figure 1). The complexity of KDD is mainly due to the nature of the analyzed data (distributed, incomplete, heterogeneous, etc.) and the nature of the process itself (since the KDD is by definition interactive and iterative).
Figure 1. Interaction between KDD and the two types of domain knowledge. Analyzed domain knowledge is used during the early stages of the process; mainly to understand and prepare data. Analyst domain knowledge is used during the latter stages to choose, configure and execute data mining methods, and to evaluate extracted patterns (Behja et al., 2005).
Given this complexity of KDD, the analyst faces two major challenges. On the one hand, he must manipulate prior domain knowledge to better understand the data and the business objective. On the other hand, he must be able to choose, configure, compose and execute tools and methods from various fields (e.g., machine learning, statistics, artificial intelligence, databases) to achieve goals. The first challenge involves analyzed domain knowledge, while the second involves the analyst domain knowledge (Figure 1).
A multi-view KDD process is usually held by one or more experts who consequently manipulate several types of knowledge and know-how. They will have different objectives and preferences, different competences, and different visions of analyzed data, KDD methods and functions. In brief, they have different viewpoints. In this context, the KDD process will be guided by the analyst’s viewpoint (Behja et al., 2005) and several types of knowledge and expertise are incorporated.
Figure 2 shows an example of a multi-view analysis of data from an e-learning system (mainly: log files, database, and courses material). These data can be analyzed by different actors of the system (learners, teachers, administrator, marketing …). The objective of a teacher (e.g., evaluation of a course) is not the same as the administrator’s one (e.g., ensuring system reliability). Attributes used for evaluating a course are different from those used for studying the reliability. Similarly, chosen data mining methods, techniques and tools will be different, and the interpretation of data mining results depends on the analyst’s viewpoint. Therefore, it is fundamental to take into account the viewpoint of each analyst and incorporate the two types of domain knowledge in the KDD process.
Figure 2. Multitude of viewpoints to analyze data from an e-learning system. The teacher may have as an objective the “evaluation of learning rate of a course,” description as KDD task, and may use (IP, UserLogin, Date, URL, Status, and Referrer) as attributes. While the administrator may have as an objective “ensuring the reliability of the system,” prediction as KDD task, and may use (IP, Date, URL, Status, and UserAgent) as attributes.
In this paper we propose to assist the users of a multi-view KDD process. Our objective is to enhance both the reusability of the process and coordination between its different users. We propose a formalization of viewpoint notion in KDD following a knowledge engineering approach: eliciting, structuring, and formalizing information and knowledge involved in a multi-view analysis (Schreiber et al., 2000). Elicitation will be based on CRISP-DM standard (Chapman et al., 1999) to identify a set of generic criteria that characterize a viewpoint in KDD. Knowledge involved in a multi-view analysis will be structured as a knowledge model containing four hierarchical sub-models: domain model, task and method model, viewpoint model and viewpoint organizational model. The viewpoint sub-model will be formalized using ontologies in OWL (Web Ontology Language) language (Bechofer, van Harmelen, Hendler, Horrocks, McGuinness, Patel-Schneider, & Stein, 2004).