Discussion of politics today often revolves around a polarized and partisan contest, which is reflected in voting models. Voting in political bodies, however, is more complex. Legislative bills involve juggling items that can elicit or detract support to build coalitions. We develop a machine learning model that captures the two factors of a voter’s preferences and items that are being voted on. The latter, collective aspect is more important for the rise of polarization in the US Senate. This, however, breaks with the unidimensional and zero-sum game formulation of political voting and suggests other opportunities for overcoming the political divide.