Personalization is one of the most expected features in the current educational systems. User modeling is supposed to be the first stage of this process, which may incorporate learning style as an important part of the model. Learning style, which is a non-stable characteristic in the case of children, differentiates students in learning preferences. This paper identifies a new hybrid method to initiate and update the information of children's learning style in an educational system. At the start-up phase, children's learning style information is gathered through the modified Murphy-Meisgeier Type Indicator for Children (MMTIC) questionnaire, which is based on the well-known Myers-Briggs Type Indicator (MBTI). This primary information will be tuned by tracking children's behaviors during the learning process. Analytical data mining helped us to cluster these behaviors and find their patterns. The proposed method was applied on 81 fourth grade children in elementary school. Delivering results suggest that this method provides a good precision in recognizing children learning style and may be an appropriate solution for non-stability problems in their preferences.