Publication Type

Journal Article

Publication Date

9-2017

Abstract

Studying developers’ behavior in software development tasks is crucial for designing effective techniques and tools to support developers’ daily work. In modern software development, developers frequently use different applications including IDEs, Web Browsers, documentation software (such as Office Word, Excel, and PDF applications), and other tools to complete their tasks. This creates significant challenges in collecting and analyzing developers’ behavior data. Researchers usually instrument the software tools to log developers’ behavior for further studies. This is feasible for studies on development activities using specific software tools. However, instrumenting all software tools commonly used in real work settings is difficult and requires significant human effort. Furthermore, the collected behavior data consist of low-level and fine-grained event sequences, which must be abstracted into high-level development activities for further analysis. This abstraction is often performed manually or based on simple heuristics. In this paper, we propose an approach to address the above two challenges in collecting and analyzing developers’ behavior data. First, we use our ActivitySpace framework to improve the generalizability of the data collection. ActivitySpace uses operating-system level instrumentation to track developer interactions with a wide range of applications in real work settings. Secondly, we use a machine learning approach to reduce the human effort to abstract low-level behavior data. Specifically, considering the sequential nature of the interaction data, we propose a Condition Random Field (CRF) based approach to segment and label the developers’ low-level actions into a set of basic, yet meaningful development activities. To validate the generalizability of the proposed data collection approach, we deploy the ActivitySpace framework in an industry partner’s company and collect the real working data from ten professional developers’ one-week work in three actual software projects. The experiment with the collected data confirms that with initial human-labeled training data, the CRF model can be trained to infer development activities from low-level actions with reasonable accuracy within and across developers and software projects. This suggests that the machine learning approach is promising in reducing the human efforts required for behavior data analysis.

Keywords

Condition Random Field, Developers’ interaction data, Software development

Discipline

Software Engineering

Research Areas

Cybersecurity

Publication

Empirical Software Engineering

First Page

1

Last Page

39

ISSN

1382-3256

Identifier

10.1007/s10664-017-9547-8

Publisher

Springer Verlag (Germany)

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.

Additional URL

http://doi.org./10.1007/s10664-017-9547-8

Share

COinS