An Overview of User Feedback Classification Approaches Rubens Santos, Eduard C. Groen, Karina Villela Fraunhofer IESE, Kaiserslautern, Germany rubens.santos@iese-extern.fraunhofer.de; {eduard.groen; karina.villela}@iese.fraunhofer.de Abstract Online user feedback about software products is a promising source of user requirements. To allow scaling analyses to large amounts of user feedback, research on Crowd-based Requirements Engineering (CrowdRE) seeks to tailor natural language processing (NLP) tech- niques to Requirements Engineering (RE). Various frameworks have been proposed, but it remains largely unclear why particular NLP techniques are better suited for CrowdRE than others, which makes it hard to make a well-founded choice for a technique. We found that CrowdRE research most often uses machine learning (ML) and has so far applied twelve clusters of ML algorithms and seven clusters of ML features. The prevalent algorithm–feature pair is Na¨ ıve Bayes with Bag of Words – Term Frequency (BOW-TF), followed by Support Vector Machines (SVM) with BOW-TF. An initial comparison of the reported precision and recall suggests that classifications in RE need further im- provement. Our research presents a preliminary overview of the current landscape of automated classification techniques for RE whose results may inspire researchers to apply new strategies to advance research in this field, or to include ML models they had not considered previously in their benchmarks. 1 Introduction Software products today typically have many geographically dispersed stakeholders who together form a large heterogeneous group of (potential) end-users known as a “crowd” [GSA + 17]. Using traditional requirements engineering (RE) approaches, such as interviews, focus groups, questionnaires, and field observations to elicit and validate requirements from a representative sample of such a crowd faces scalability problems. The discipline within RE that investigates solutions for how to derive validated requirements from the crowd by motivating its members to provide feedback, analyzing their feedback and behavior, and validating the outcomes is generally known as Crowd-based RE (CrowdRE) [GSA + 17]. One promising approach within CrowdRE is the use of online user feedback about software products for RE. This kind of feedback is widely and often publicly available on review platforms, social media, and tracking systems and forms a valuable resource for useful information from which potential software improvements and changed or new requirements can be derived [GSA + 17]. However, because the crowd providing this feedback is large, the amount of information they provide can quickly become too abundant to be analyzed manually [GSK + 18], causing similar scalability problems as those observed with traditional RE techniques. This is why Copyright c  2019 by the paper’s authors. Copying permitted for private and academic purposes.