Quantitative structure-activity relationship (QSAR) studies on a series of selective inhibitors of the cyclin-dependent kinase 4 (CDK4) were performed by using two conventional global modeling methods (multiple linear regression (MLR) and support vector machine (SVM)), local lazy regression (LLR) as well as three consensus models. It is remarkable that the LLR model could improve the performance of the QSAR model significantly. In addition, due to the fact that each model can predict certain compounds more accurately than other models, the above three derived models were all used as submodels to build consensus models using three different strategies: average consensus model (ACM), simple weighted consensus model (SWCM) and hat weighted consensus model (HWCM). Through the analysis of the results, the HWCM consensus strategy, firstly proposed in this work, proved to be more reliable and robust than the best single LLR model, ACM and SWCM models.