enterpriseandobject-orientation（7页）.doc

A Preprocessing Method to Treat Missing Values in Knowledge Discovery in Databases Arnaud Ragel Many of analysis tasks have to deal with missing values and have developed specific and internal treatments to guess them. In this paper we present an external method, MVC (Missing Values Completion), to improve performances of completion and also declarativity and interactions with the user for this problem. Such qualities will allow to use it for the data cleaning step of the KDD process[6]. The core of MVC, is the RAR algorithm that we have proposed in[15]. This algorithm extends the concept of association rules[1] for databases with multiple missing values. It allows MVC to be an efficient preprocessing method: in our experiments with the c4.5[13] decision tree program, MVC has permitted to divide, up to two, the error rate in classification, independently of a significant gain of declarativity. 1 Introduction The missing values problem is an old one for analysis tasks[9][12]. The waste of data which can result from casewise deletion of missing values, obliges to propose alternatives approaches. A current one is to try to determine these values[10],[4]. However, techniques to guess the missing values must be efficient, otherwise the completion will introduce noise. With the emergence of KDD for industrial databases, where missing values are inevitable, this problem has become a priority task [6] also requiring declarativity and interactivity during treatments. At the present time, treatments are often specific and internal to the methods, and do not offer these qualities. Consequently the missing values problem is still a challenging task of the KDD research agenda[6]. To complete missing values a solution is to use relevant associations between the attributes of the data. The problem is that it is not an easy task to discover relations in data containing missing values. The association rules algorithms [2] [1]are a recent and efficient approach to extract quickly