The data-based design of multi-model linear inferential (soft) sensors (MIS) is studied. These promise increased prediction accuracy yet simplicity of the model structure and training. The state-of-the-art approach to the MIS design consists of three steps: 1) data labeling (establishing training subsets for individual models), 2) data classification (creating a switching logic for the models), and 3) training of individual models. There are two main issues with this concept as steps 2)&3) are separate: (i) discontinuities can occur when switching between the models; (ii) data labeling disregards the quality of the resulting model. Our contribution aims at both the mentioned problems, where, for the problem (i), we introduce a novel support vector method (SVM)-based model training coupled with switching logic identification and, for the problem (ii), we propose a direct optimization of data labeling. The proposed methodology and its benefits are illustrated on an example from the chemical engineering domain.