Dynamic discriminant functions with missing feature values

Datasets with missing feature values are often encountered, especially in biometric databases. A common solution is to fill in the missing values by imputation. Unfortunately there is no universally best imputation method and the performance of a classifier can be degraded by poor imputations. In this paper, we propose a framework called the dynamic Fisher's linear discriminant that uses a quadratic classifier with a dynamically modified quadratic discriminant function. By eliminating imputations as far as possible, the proposed framework is useful for pattern classification. Satisfactory results are obtained from experiments conducted on four datasets from the UCI machine learning repository and the KEEL dataset repository, together with four fingerprint datasets.