Is deep learning better than traditional approaches in tag recommendation for software information sites?

Context: Inspired by the success of deep learning in other domains, this new technique been gaining widespread recent interest in being applied to diverse data analysis problems in software engineering. Many deep learning models, such as CNN, DBN, RNN, LSTM and GAN, have been proposed and recently applied to software engineering tasks including effort estimation, vulnerability analysis, code clone detection, test case selection, requirements analysis and many others. However, there is a perception that applying deep learning is a ”silver bullet” if it can be applied to a software engineering data analysis problem. Object: This motivated us to ask the question as to whether deep learning is better than traditional approaches in tag recommendation task for software information sites. Method: In this paper we test this question by applying both the latest deep learning approaches and some traditional approaches on tag recommendation task for software information sites. This is a typical Software Engineering automation problem where intensive data processing is required to link disparate information to assist developers. Four different deep learning approaches – TagCNN, TagRNN, TagHAN and TagRCNN – are implemented and compared with three advanced traditional approaches – EnTagRec, TagMulRec, and FastTagRec. Results: Our comprehensive experimental results show that the performance of these different deep learning approaches varies significantly. The performance of TagRNN and TagHAN approaches are worse than traditional approaches in tag recommendation tasks. The performance of TagCNN and TagRCNN approaches are better than traditional approaches in tag recommendation tasks. Conclusion: Therefore, using appropriate deep learning approaches can indeed achieve better performance than traditional approaches in tag recommendation tasks for software information sites.