The recent advances in data mining have produced algorithms for extracting hidden and potentially useful knowledge in large data sets, which are assumed to be complete and reliable. However, data suitable for mining comes from various sources, has different formats, and can have missing or incorrect values. Incomplete data sets significantly distort mining results. Therefore, data preparation to taking care of missing or out-of-range values is very critical to knowledge discovery. This paper proposes a generic framework for missing data imputation using neural networks, where two-stage filling algorithms are implemented. An empirical evaluation of this method through a large credit card data set is performed.