ABSTRACT Relational database designs evolve. Table columns are modified, deleted, or extended. Tools and techniques are available to assess the effects of these modifications on source code. Unfortunately, these changes have an impact on other tables of the same database as well. There are a few tools for analyzing this impact, and they mainly analyze the impact of modifications and deletions, while ignoring table extensions. We evaluate the effectiveness of alternative similarity comparison approaches applied to database tables for analyzing the impact of data model extensions. We use three similarity‐based and five distance‐based metrics that are applied in various domains. In addition, we evaluate the effectiveness of Siamese networks, which constitute a type of neural network architecture designed to measure similarity between pairs of input samples. We conduct a case study in an industrial setting to assess the accuracy of alternative metrics and approaches. We use two datasets collected from Customer Relationship Management (CRM) and Payment application domains. For the CRM dataset, the highest F1 score is achieved with a distance‐based metric, Bray–Curtis, as 98.20% on average. Average F1 scores obtained with Cosine similarity and Siamese networks are 98.10% and 96.33%, respectively. In the Payment dataset, the highest F1 score is achieved with Bray–Curtis as 81.85% on average, followed by PearsonR with 81.07% and Cosine similarity with 80.63%. Although their relative order varied across domains, Bray–Curtis consistently ranked as the best performing method. The use of Siamese networks requires an additional training step, although the training and validation of the model take less than a minute on commodity hardware. We share the developed tool that implements alternative approaches and the anonymized dataset for the replication of our experiments. Automated and accurate impact analysis for data model extensions is possible. Tool support can enable junior developers to conduct precise impact analysis without relying on assistance from domain experts.
Ersoy et al. (Fri,) studied this question.