This paper explores a combined global and local identification approach for linear parameter-varying systems. Ideally, the combined approach retains advantages of its two extremes - global and local - with the possibility to emphasize one or the other. Practically, it is prone to overfitting. This paper proposes a remedy based on the ℓ_2,1-norm regularization, describes its implementation within the nonlinear least squares framework, and gives an experimental validation. The results show a substantial decrease in the Euclidean norm of the model parameters, which resulted in a significantly smoother frequency response function surface and in overall, less-deviating model behavior.