Machine learning based data-driven solutions have potential to significantly improve quality of incident management process and make it cost effective. We present our experiences while addressing a spectrum of interrelated problems encountered in practice including identifying semantically related incidents, assignee recommendation, and mapping of incidents to business processes. We argue that despite long-standing research, it is not always straightforward to adopt recommendations from research into practice due to variability and complexity of business constraints and nature of data. We discuss need for meta-level analysis and suggest our own recommendations towards designing pragmatic solutions with low barriers to adoption and addressing right level of challenges.

Software defect prediction is still a challenging task in industrial settings. Noisy data and/or lack of data make it hard to build successful prediction models. In this study, we aim to build a change-level defect prediction model for a software project in an industrial setting. We combine various probabilistic models, namely matrix factorization and topic modeling, with the expectation of overcoming the noisy and limited nature of industrial settings by extracting hidden features from multiple resources. Commit level process metrics, latent features from commits, and semantic features from commit messages are combined to build the defect predictors with the use of log filtering and feature selection techniques, and two machine learning algorithms Naive Bayes and Extreme Gradient Boosting (XGBoost). Statistical tests show that collecting data from various sources and applying data pre-processing techniques show an improvement in terms of probability of detection by up to 24% when compared to a base model with process metrics only.

A lot of research in Software Engineering (SE) automatically extract topics of the text data and use the results directly or as a feature for a machine learning method. Research has shown that the majority of studies in SE use Latent Dirichlet Allocation (LDA) as the topic modeling approach. Similarly, there is a lot of work that apply LDA on GitHub data. However, there is no study that explores whether LDA is a good choice compared to other algorithms, nor is there any to investigate the effects of specific pre-processing steps on its performance. In this paper, we explore a large dataset of GitHub repositories and apply two main topic modeling algorithms, LDA (3 variants) and Non-Negative Matrix Factorization (NMF), in several experiments with different experimental settings. The results show that LDA results in a higher coherence score compared to NMF. However, care should be taken in the choice of LDA algorithm, setting its parameters, and the text pre-processing steps. The results of this paper benefit SE researchers who apply intelligent techniques using LDA.