When public opinion constructs are multidimensional, it is ideal to avoid aggregating different dimensions into an attitudinal scale and instead categorize observations into attitudinal types. The available categorization techniques, however, are primarily unsupervised learning methods, which generate attitudinal types that make post hoc conceptualization difficult. This study proposes a novel concept-driven approach in which machines “learn” the ideal types of each attitude and specify them as constraints. The results are therefore grounded in a priori conceptualization. This study then applies the concept-driven clustering approach to analyze data from the Political Trust Module of the World Values Survey. It shows that respondents who select the same level of political trust fall into different attitudinal types. It also finds that when the trust level is held constant, the trust type significantly affects the propensity to engage in political activism and voting. Overall, this study (1) theoretically argues that multidimensional constructs should be operationalized according to their conceptual structures, (2) methodologically reveals the added value of machine learning in integrating conceptual constraints into empirical analysis, and (3) substantively contributes to a more accurate understanding of political trust and its behavioral implications.