Latent class models (LCMs) are widely used to explore dietary consumption behaviors in large populations by querying a wide range of foods. Yet, it is not clear if the dietary patterns should be described by a major food group or at a more granular level. This work capitalizes on a pre-specified group over the food items that reflect nutritional similarities. The main statistical challenge is to distinguish subpopulations of distinct dietary patterns where between-class differences may occur at group or more specific item levels. Extending classic LCMs, our model 1) hierarchically clusters individuals, 2) places classes on the leaves of an unknown tree and encourages nearby classes to share similar response probabilities, and 3) integrates item groups to allow the response profiles to vary in a subset of item groups. Building upon Dirichlet diffusion tree processes that specify a joint prior for the unknown tree over classes and response probabilities, we derive a scalable variational message passing algorithm for posterior inference. Using simulated and prospective cohort dietary data, we demonstrate the utility of our model and improvement compared to standard techniques.