Missing values are an old problem that is very common in real data bases. We describe the damages caused by missing values on condensed representations of patterns extracted from large data bases. This is important because condensed representations are very useful to increase the efficiency of the extraction and enable new uses of frequent patterns (e.g. rules with minimal body, clustering, classification). We show that, unfortunately, such condensed representations are unreliable in presence of missing values. We present a method of treatment of missing values for condensed representations based on δ-free or closed patterns, which are the most common condensed representations. This method provides an adequate condensed representation of these patterns. We show the soundness of our approach, both on a formal point of view and experimentally. Experiments are performed with our prototype MV_(MINER) (for Missing Values miner), which computes the collection of appropriate δ-free patterns.
展开▼
