Automatic Feature Exploration and an Application in Defect Prediction

Many software engineering tasks heavily rely on hand-crafted software features,
e.g.
, defect prediction, vulnerability discovery, software requirements, code review, and malware detection. Previous solutions to these tasks usually directly use the hand-crafted features or feature selection techniques for classification or regression, which usually leads to suboptimal results due to their lack of powerful representations of the hand-crafted features. To address the above problem, in this paper, we adopt the effort-aware just-in-time software defect prediction (JIT-SDP), which is a typical hand-crafted-feature-based task, as an example, to exploit new possible solutions. We propose a new model, named
neural forest
(NF), which uses the deep neural network and decision forest to build a holistic system for the automatic exploration of powerful feature representations that are used for the following classification. NF first employs a deep neural network to learn new feature representations from hand-crafted features. Then, a decision forest is connected after the neural network to perform classification and in the meantime, to guide the learning of feature representation. NF mainly aims at solving the challenging problem of combining the two different worlds of neural networks and decision forests in an end-to-end manner. When compared with previous state-of-the-art defect predictors and five designed baselines on six well-known benchmarks for within- and cross-project defect prediction, NF achieves significantly better results. The proposed NF model is generic to the classification problems which rely on the hand-crafted features.