Predicting Future Developer Behavior in the IDE Using Topic Models

Interaction data, gathered from developers' daily clicks and key presses in the IDE, has found use in both empirical studies and in recommendation systems for software engineering. We observe that this data has several characteristics, common across IDEs: 1) exponentially distributed - some events or commands dominate the trace (e.g., cursor movement commands), while most other commands occur relatively infrequently; 2) noisy - the traces include spurious commands (or clicks), or unrelated events, that may not be important to the behavior of interest; 3) comprise of overlapping events and commands - specific commands can be invoked by separate mechanisms, and similar events can be triggered by different sources. These characteristics of this data are analogous to the characteristics of synonymy and polysemy in natural language corpora. Therefore, this paper (and presentation) presents a new modeling approach for this type of data, leveraging topic models typically applied to streams of natural language text.