Ensemble Learning for Method-Call Recommendation
Code Completion is one of the advanced features of modern IDEs. However, most existing tools still just present an alphabetically sorted list of proposals. Using prediction models based on minded knowledge from existing code, the accuracy and predictive power of the code completion are strongly improved, as shown in prior work by Bruch et al., Amann, and Heinemann et al.. In our thesis, we discuss several context features suited to improve code completion. We evaluate models based on one of these features each. We show that which features are needed for good predictions is type-dependent. We discuss and evaluate several approaches to combine theses base models by linear combination. Thereby we show that a model combining the proposed features is able to outperform the base models.