|
Abstract:
|
In supervised learning problems involving high-dimensional data, it is often desirable to reduce the number of variables. Biomarker data is usually high-dimensional and therefore a good candidate for variable selection approaches. Identifying a small number of important markers is not only essential to achieve high accuracy for the prediction model, but also necessary to allow an easy interpretation of the model from a biological point of view. Using biomarker data from an obesity study, we will illustrate the challenges involved in the model building and marker selection process. In particular, we will use a modified version of the RandomForest Wrapper Algorithm (Svetnik et al. 2004) to build models based on a small number of important markers and make comparisons to other approaches.
|
