In many practical machine learning systems, the prediction/classification tasks involve the usage of heterogeneous data in semi-supervised settings, where the objective is to maximize the utility of multiple views (usually dual views) information from the data. In this work, we propose a general framework, Dual Uncertainty Minimization Regularization (DUMR), that maximizes the usage of heterogeneous data for a dual view semi-supervised classification/prediction. Through extending a recent uncertainty regularizer to a heterogeneous setting, we propose to optimize an objective which ensures the minimum uncertainty of the prediction over both views extracted from heterogeneous source. In specific, for different problem settings, we design two type of uncertainty regularizer with entropy and squared-loss mutual information, separately. The proposed framework is exploited in three datamining/multimeida analysis tasks, social role identification, legislative prediction and action recognition, and the comparison with other peer methods corroborate the superior performance of the proposed method.