Improving Code Recommendations by Combining Neural and Classical Machine Learning Approaches

Abstract

Code recommendation systems for software engineering are designed to accelerate the development of large software projects. A classical example is code completion or next token prediction offered by modern integrated development environments. A particular challenging case for such systems are dynamic languages like Python due to limited type information at editing time. Recently, researchers proposed machine learning approaches to address this challenge. In particular, the Probabilistic Higher Order Grammar technique (Bielik et al., ICML 2016) uses a grammar-based approach with a classical machine learning schema to exploit local context. A method by Li et al., (IJCAI 2018) uses deep learning methods, in detail a Recurrent Neural Network coupled with a Pointer Network. We compare these two approaches quantitatively on a large corpus of Python files from GitHub. We also propose a combination of both approaches, where a neural network decides which schema to use for each prediction. The proposed method achieves a slightly better accuracy than either of the systems alone. This demonstrates the potential of ensemble-like methods for code completion and recommendation tasks in dynamically typed languages.