Recommendation Modelling based on Contextual Weather Data in an Adaptive Coaching System: A Semantic and Explainable Approach

Physical activities are of two types-indoor and outdoor. Outdoor exercise is an enjoyable way for individuals to improve their fitness; however, it depends on the weather. Staying active year-round can help to maintain strength, control weight gain, and improve overall health. The weather has been identified as a barrier to physical activity participation. The adverse weather conditions are most important to lead to reduced physical activity. As cold, rainy, foggy, and snowy weather settles in, individual exercise and physical activity may begin to decrease. The effects of heat stress and/or dehydration may lead to decreased physical activities. Physical activity is stressful in high ambient temperature and humidity. Cold injuries are a common result of exposure to cold during physical activities. Many people engage in physical fitness programs year-round in cold, wet or windy conditions (or a combination thereof), putting themselves at risk of injury from the cold. In this regard, a continuous evidence-based and goal-oriented monitoring system can be beneficial to guide people with automatic personalized recommendation generation to perform either indoor or outdoor activities based on the probabilistic distribution of external weather data. In this research, we have conceptualized the design of an automatic coaching system (eCoaching) that can generate recommendations based on the processing of contextual external weather data. Moreover, we have proposed an algorithm to annotate collected data, process and classify them, generate recommendations, and semantically represent the obtained knowledge using an Ontology. For the overall experiment, we collected weather data variations for one and a half years from thirteen cities of the Southern-Norway. Followed by, we added rule for indoor and outdoor activity types. The classification performance has been evaluated using the metrics such as accuracy, precision, recall, F1-score, brier-score, and Mathew’s coefficient (MCC) and achieves 99.99% accuracy with DecisionTree Classifier. Moreover, we used SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) to explain the classifications. The consistency of the Ontology model has been evaluated with reasoners and reasoning time. The Ontology model has been useful for semantic representation of collected and processed data, and the rule-based recommendation modeling with machine learning algorithm, a knowledge base and description logics.