Recognizing the transitions of intensity of facial expressions in daily life is crucial to monitoring the individual’s mental state. We have investigated a two-input facial expression recognizer specialized for individuals to evaluate facial expressions in detail with personalized scales. Although this recognizer can accurately recognize facial expressions based on comparisons, it has the limitation of requiring the manual provision of supervised data for each individual to train the recognizer. The goal of this study is to automatically generate the datasets required for training the two-input facial expression recognizer for smiles from videos that capture an individual’s facial changes. To achieve this goal, we propose a method for extracting facial image pairs with explicit expression differences using a conventional facial expression recognizer. Additionally, we introduce a three-input facial expression change recognizer to extract facial image pairs with subtle expression differences. Experimental results showed that our proposed framework could successfully generate a dataset for training the two-input facial expression recognizer.