We propose a single unified minimax entropy approach for user preference modeling with multidimensional knowledge. Our approach provides a discriminative learning protocol which is able to simultaneously a) leverage explicit human knowledge, which are encoded as explicit features, and b) model the more ambiguous hidden intent, which are encoded as latent features. A latent feature can be carved by any parametric form, which allows it to accommodate arbitrary underlying assumptions. We present our approach in the scenario of check-in preference learning and demonstrate it is capable of modeling user preference in an optimized manner. Check-in preference is a fundamental component of Point-of-Interest (POI) prediction and recommendation. A user's check-in can be affected at multiple dimensions, such as the particular time, popularity of the place, his/her category and geographic preference, etc. With the geographic preferences modeled as latent features and the rest as explicit features, our approach provides an in-depth understanding of users' time-varying preferences over different POIs, as well as a reasonable representation of the hidden geographic clusters in a joint manner. Experimental results based on the task of POI prediction/recommendation with two real-world checkin datasets demonstrate that our approach can accurately model the check-in preferences and significantly outperforms the state-of-art models.