Species interaction models are an important tool in community ecology, but fitting a model to a particular community requires a lot of data, as there are a lot of interactions to quantify. The aggregation of species into groups is a common tool to reduce the number of parameters in these models. Current methods require choosing groupings a priori, and reliable methods of selecting appropriate groupings are not available. Herein, I describe a method for finding optimal groupings a posteriori, yielding a data-informed approach for using grouping to improve the parsimony of species interaction models. Applying this method to empirical data, I find that optimal species groupings are difficult to derive from information available a priori. Further, different groupings are needed to describe different components of species interactions within the same community species group differently in how they affect others than in how they respond to interactions from others. This violates the common assumption in species interaction models that a single grouping can be used to aggregate species in all regards.