Sebastian Hönel

Sebastian Hönel, Jonas Nordqvist

The detection and localization of quality-related problems in industrially mass-produced products has historically relied on manual inspection, which is costly and error-prone. Machine learning has the potential to replace manual handling. As such, the desire is to facilitate an unsupervised (or self-supervised) approach, as it is often impossible to specify all conceivable defects ahead of time. A plethora of prior works have demonstrated the aptitude of common reconstruction-, embedding-, and synthesis-based methods in laboratory settings. However, in practice, we observe that most methods do not handle low data quality well or exude low robustness in unfavorable, but typical real-world settings. For practitioners it may be very difficult to identify the actual underlying problem when such methods underperform. Worse, often-reported metrics (e.g., AUROC) are rarely suitable in practice and may give misleading results. In our setting, we attempt to identify subtle anomalies on the surface of blasted forged metal parts, using rather low-quality RGB imagery only, which is a common industrial setting. We specifically evaluate two types of state-of-the-art models that allow us to identify and improve quality issues in production data, without having to obtain new data. Our contribution is to provide guardrails for practitioners that allow them to identify problems related to, e.g., (lack of) robustness or invariance, in either the chosen model or the data reliably in similar scenarios. Furthermore, we exemplify common pitfalls in and shortcomings of likelihood-based approaches and outline a framework for proper empirical risk estimation that is more suitable for real-world scenarios.

Sebastian Hönel

Detecting the presence of project management anti-patterns (AP) currently requires experts on the matter and is an expensive endeavor. Worse, experts may introduce their individual subjectivity or bias. Using the Fire Drill AP, we first introduce a novel way to translate descriptions into detectable AP that are comprised of arbitrary metrics and events such as logged time or maintenance activities, which are mined from the underlying source code or issue-tracking data, thus making the description objective as it becomes data-based. Secondly, we demonstrate a novel method to quantify and score the deviations of real-world projects to data-based AP descriptions. Using fifteen real-world projects that exhibit a Fire Drill to some degree, we show how to further enhance the translated AP. The ground truth in these projects was extracted from two individual experts and consensus was found between them. We introduce a novel method called automatic calibration, that optimizes a pattern such that only necessary and important scores remain that suffice to confidently detect the degree to which the AP is present. Without automatic calibration, the proposed patterns show only weak potential for detecting the presence. Enriching the AP with data from real-world projects significantly improves the potential. We also introduce a no-pattern approach that exploits the ground truth for establishing a new, quantitative understanding of the phenomenon, as well as for finding gray-/black-box predictive models. We conclude that the presence detection and severity assessment of the Fire Drill anti-pattern, as well as some of its related and similar patterns, is certainly possible using some of the presented approaches.

Sebastian Hönel, Morgan Ericsson, Welf Löwe et al.

Source code is changed for a reason, e.g., to adapt, correct, or adapt it. This reason can provide valuable insight into the development process but is rarely explicitly documented when the change is committed to a source code repository. Automatic commit classification uses features extracted from commits to estimate this reason. We introduce source code density, a measure of the net size of a commit, and show how it improves the accuracy of automatic commit classification compared to previous size-based classifications. We also investigate how preceding generations of commits affect the class of a commit, and whether taking the code density of previous commits into account can improve the accuracy further. We achieve up to 89% accuracy and a Kappa of 0.82 for the cross-project commit classification where the model is trained on one project and applied to other projects. Models trained on single projects yield accuracies of up to 93% with a Kappa approaching 0.90. The accuracy of the automatic commit classification has a direct impact on software (process) quality analyses that exploit the classification, so our improvements to the accuracy will also improve the confidence in such analyses.