LAUSR

Predicting future moods with machine learning can help plan appropriate strategies for mental health; this can be facilitated by collecting future event information from calendar applications. However, a model for mood prediction from calendar event information has not yet been developed. In addition, because the dataset of mood values is characterized by the tendency to collect values close to either the mean or the most recent value, a mood prediction model employing past moods as input is likely to overfit the distribution of the training dataset. This makes it difficult to capture changes in mood due to external factors such as events, especially on days when mood fluctuates greatly. In this study, we propose a mood prediction deep learning model with an event feature extractor. For introducing a multitask learning method, we add not only mood but also the effects of events on mood as objective variables, and we achieve higher prediction accuracy than existing machine learning models. Given the small initial weights in deep learning, the middle layer tends to output values near zero when learning is insufficient. With the introduction of multitask learning, the output values of the event feature extractor in the prediction model output non-zero values, and the prediction accuracy is improved on days featuring a large fluctuation in mood from the previous day. Our proposal is expected to be a more efficient learning method for deep learning models to predict human mental states from external factors.