Predicting future outcomes based on past observational data is a common application in data mining. While the primary goal is usually to achieve the highest possible prediction accuracy, the interpretation of the resulting prediction model is important to understand its shortcomings for further improvements.
Throughout this paper we focus on branch prediction, where the (binary) outcome of a test is needed for enhancing the performance of pipelined computer architectures. Many research has been done in this domain and different branch prediction solutions are described in the literature.
The quality of a prediction model is highly dependent on the quality of the available data. Especially the choice of the related variables or features to base the prediction on is important. In this paper we evaluate the predictive power of different branch prediction features using the metric Gini-index,
which is used as feature selection measure in the construction of decision trees. We observe that through this Gini-metric an explanation can be provided for the performance of existing branch predictors. We show that the Gini-index is a good metric for comparing branch prediction features. Further, we found that a feature can have good discriminative capacities, although this does not result in very good accuracies because of shortcomings in the predictor implementation.