Facial expression analysis has been well studied in recent years; however, these mainly focus on domains of posed or clear facial expressions. Meanwhile, subtle/micro-expressions are rarely analyzed, due to three main difficulties: inter-class similarity (hardly discriminate facial expressions of two subtle emotional states from a person), intra-class dissimilarity (different facial morphology and behaviors of two subjects in one subtle emotion state), and imbalanced sample distribution for each class and subject. This paper aims to solve the last two problems by first employing preprocessing steps: facial registration, cropping and interpolation; and proposes a person-specific AdaBoost classifier with Selective Transfer Machine framework. While preprocessing techniques remove morphological facial differences, the proposed variant of AdaBoost deals with imbalanced characteristics of available subtle expression databases. Performance metrics obtained from experiments on the SMIC and CASME2 spontaneous subtle expression databases confirm that the proposed method improves classification of subtle emotions.