This paper proposes a mini-batch discriminative feature weighting methodology for minimization of classification error in datasets with considerable number of classes and poor intra class information. Presented approach improves the classification system by enhancing the components more relevant to the recognition. It is based on the maximization of interclass Euclidean distance by utilization of information from all classes. A weighted nearest neighbor classifier is used for the classification. A mini-batch principle is implemented into the training process in order to boost the learning speed, which is a bottleneck for traditional batch algorithms. We report how the weighting can be applied to the task of Local Binary Patterns-based face recognition. The performance of the algorithm is evaluated on a color FERET database.