Brain-computer interface (BCI) has attracted attention as a means of assisting patients with intractable neurological diseases to communicate their intentions. However, there are no reports examining the mental tasks that are effective in acquiring physiological signals for BCI, based on evaluation of classification performance and subjective usability of the mental tasks. This study aimed to investigate preferable mental task combinations that could be utilized in the task selection process for near-infrared spectroscopy (NIRS)-based BCIs. We evaluated the classification model performance of brain activation responses using NIRS signals and subjective usability of mental tasks. NIRS signals were measured in 10 healthy adult participants while they performed mental arithmetic task (MA), mental singing task (MS), mental writing task (MW), and mental figure rotation task (MFR), using a block design consisting of a 30-s rest and 30-s task period. For six combinations of mental tasks, binary classification models were constructed using random forest with a dimension reduction method, and the classification performance was evaluated using three-fold cross-validation. We also measured the subjective usability of the mental tasks. The results showed that MW vs. MFR was classified with an average accuracy of 71.5%, and six participants achieved over 70% accuracy. The subjective ratings of state anxiety and acceptability showed that all the mental tasks were rated as acceptable for use in BCI applications. These findings may be utilized as the preliminary communication tasks in the process of selecting appropriate cognitive tasks for BCI for individual users.