Activity logging applications for mobile health have gained wide attentions recently and tracking of user health states is becoming an emerging trend in human daily life. Modeling user behaviors from users' temporal event records is a promising research topic for enhancing the practicability of personal health management. However, user modeling has long been a very challenging problem due to the complexity in activities of each individual, which consists of two key sub-issues, periodicity and diversity. In this paper, we propose a novel user modeling approach, namely Periodic User Representation Learning (PURL), to learn dynamic representations of user behaviors from large-scale temporal event records. To deal with the periodicity issue, we utilize periodic frequent pattern mining to capture the periodic user behaviors from users' temporal event records. Next, to cope with the diversity of user behaviors, we further build a periodic temporal pattern embedding module, which yields interpretable user representations according to the presence of each distinct periodic temporal pattern. To the best of our knowledge, PURL is the first user behavior modeling method that tackles the periodicity and diversity issues of users' temporal event records. Based on a real-world large-scale user activity logging dataset, experimental results demonstrate that PURL delivers up to 47% improvement in terms of the accuracy on health state prediction task compared with other existing methods.