Predicting Weight Loss With Enemble Methods

Overweight and obesity are ubiquitous around the world, and to make matters worse, many fatal diseases are related to obesity. Consequently, an increasing effort have been made to help people lose weight. Due to the cost-effective, convenient and wide-reaching features of mobile applications, smart phone weight loss applications have gained great public attention. Such weight loss applications typically allow users to track their weight change, serving as an important positive reinforcement factor to motivate more effort to lose weight. In addition, social networking functions in the weight loss applications make it very convenient for users to share or seek information as well as support from peers. Past studies have proven the effectiveness of mobile applications in weight loss intervention. However, such intervention cannot perform in a timely manner as users can only adjust their weight loss behaviors according to the history weight records. Thus, the effectiveness of weight loss applications will be largely improved if users' future weight can be automatically predicted. Although future weight prediction is challenging, the results of our previous work based on a large-scale dataset with near 10 million users have shown that by leveraging the abundant social information available in most weight loss applications, we can build a prediction model demonstrating a strong performance. However, due to the nature of pilot study, the model built in our previous work is relatively simple as it can only capture linear mappings from a small portion of users who provide complete information. Thus, in our ongoing research, we aim to build a stronger non-linear model which is able to incorporate users who only provide partial information. As a large portion of users in the dataset only provide partial information, we believe the new model that utilizes much more data will achieve significantly better performance than the previous one.