Modeling of high-density interconnect channels requires time-consuming, compute-intensive physics-based electromagnetic simulations. In this work, we use a supervised ensemble machine learning technique for this purpose to significantly reduce the design time and effort of such channels. A randomized dataset selection methodology has been proposed to significantly reduce the dataset required to train such models. The methodology uses the characteristic impedance profile and S-parameters to derive the reduced random dataset that follows a skewed distribution. The supervised ensemble model combining, K-nearest neighbor and decision tree regressor, is trained on this reduced dataset to predict S-parameters for given input physical parameters and material properties. The model is trained and tested on a single-ended stripline channel, and its accuracy is validated by comparing the predicted S-parameters to those obtained from the electromagnetic simulator. Training the model with a small dataset comprising merely 85 samples results in an accuracy of 97% in predicting the S-parameters, which improves significantly over the results obtained in the prior art.