We investigate the differentiation of cold and heat syndromes in Traditional Chinese Medicine with a special concern on the issue of data imbalance. Data imbalance occurs frequently in syndrome differentiation. In this study, we use a neural network classifier, fastText, to differentiate cold and heat syndromes, which have skewed distributions in the medical records and in the population. We investigate several sampling techniques to tackle the issued of data imbalance, including oversampling, under-sampling, and bagging. We further set thresholds of probabilities of the classifiers based on an observation on precision recall curves. The performances are evaluated using macro-averaged F1 score. It is disclosed the classifier we used, fastText, is insensitive to imbalanced data.