Many of today's speech recognition applications can benefit from long-term speaker adaptation using speaker logs, and discriminative methods present a promising approach for that given their previous successes. This paper carries out largevocabulary speech recognition experiments to investigate performance of feature-space and model-space discriminative adaptation methods for long-term speaker adaptation. The experimental results suggest that though on average discriminative adaptation does not obtain a big gain over ML adaptation, there are still a number of test speakers that show significant improvements. Motivated by this observation, we further propose an efficient method to automatically select speakers which can obtain big improvements in discriminative adaptation. When 35%~65% of the whole test population are selected for discriminative adaptation, the relative WER reduction over ML adaptation can reach 4%∼5% if only these speakers' performance is inspected.