Selection of tuning parameters in pliable lasso models via modified Bayesian type criteria for high-dimensional data sets

High-dimensional data analysis is important across many research domains. In this setting, selecting suitable tuning parameters is critical for accurate model development. This study examines the selection of the tuning parameters lambda and alpha in pliable lasso models. We also investigate the optimal settings of the criterion-specific hyperparameters within the Bayesian-type information criteria EBIC, mBIC, and mBIC2. A wide range of lambda and alpha values is systematically evaluated to identify the trade-off between predictive accuracy and parsimony in high-dimensional, structured datasets. Through extensive simulations and real-world datasets, we assess the performance of these criteria under varying conditions of correlation, sparsity, and sample size. The results indicate that EBIC at low to moderate penalty strength, and mBIC at low to moderate prior inclusion probabilities, achieve the most favorable balance between parsimony and predictive performance. When this probability is increased to 0.75, mBIC attains the lowest prediction error in most settings, offering a useful prediction-oriented alternative at the expense of model parsimony. By contrast, mBIC2 consistently selects too many predictors across all settings. This work provides a systematic empirical benchmark of Bayesian-type information criteria for selecting lambda and alpha in the pliable lasso, offering practical guidelines for hyperparameter selection in high-dimensional data.