Dealing with Incomplete Data in Pattern Recognition

The ability to deal with partial or uncertain information is a fundamental requirement for systems working in the real world. In many applications, it is important to learn relations between the data even when they appear incomplete or corrupted by noise. In others, it is necessary to know how to react in presence of missing or unreliable inputs, for instance when sensors fail or provide noisy measurements. In this paper reviews some techniques and algorithms proposed in the literature for dealing with incomplete data in areas related to statistical pattern recognition, density estimation, and neural networks. Particular attention is given to methods based on statistical models which allow to cope also with noisy data