Quantification of individual dataset contributions to prediction accuracy in cooperative learning

We consider cooperative learning, a recently proposed technique to leverage the predictive power of several datasets (also called data views) to improve the prediction accuracy of an outcome of interest. Cooperative learning uses a Lasso-type penalty to fit several datasets to a given outcome, while a so-called agreement penalty enforces that the predictions made by the individual datasets agree. We are interested in the question of whether the predictive power of each individual dataset can be quantified using cooperative learning. In this work, we demonstrate that certain trace plots, analogously to the ones for the classic Lasso, allow one to quantify the predictive power of each individual dataset. Importantly, this allows one to detect datasets which do not carry any predictive power on the outcome. In an experimental study, we quantify the predictive power of three real datasets in the context of the Childhood Asthma Management Program (CAMP), with the three datasets containing information on epidemiological variables, metabolites, and clinical data, respectively.