Gene expression profiles provide hidden biological knowledge and key information that can be used to distinguish different types of cancer. Due to their high dimensionality and redundancy, gene expression data are often preprocessed by dimensionality reduction (DR) methods. Conventional supervised DR methods use only labeled samples to train the model, leading to a limited performance due to small number of labeled samples in the real world. This paper proposes a transductive local Fisher discriminant analysis (TLFDA) method that uses the available unlabeled data in the learning process. On the one hand, the label information is utilized to maximize the inter-class distance in the embedding space. On the other hand, the local structural information of all data samples is taken into consideration to maintain the smoothness property. In this way, the TLFDA provides more discriminative power than state-of-the-art supervised or semi-supervised DR methods, even when the number of labeled samples is very limited. Our experimental results on benchmark GCM and Acute Leukemia datasets show its promising performance on gene expression profile-based cancer classification.