[Tech. Repo] Evaluating Transformer-Based Embeddings for Software Change Recommendation: From General Models to Specialized Models

Ms. Savira Ramadhanty, a second-year master student in our research group, presented her research at the 222nd meeting of the IPSJ Special interest group of Software Engineering (SIGSE), held over two days starting March 9 at the Ookayama Campus of our university. This presentation is based on a portion of her master thesis.

Authors：Savira Ramadhanty, Profir-Petru Par?achi，Takashi Kobayashi（Science Tokyo）
Title：Evaluating Transformer-Based Embeddings for Software Change Recommendation: From General Models to Specialized Models
Publication： The SIG Technical Reports of IPSJ, Vol. 2026-SE-222, No. 3, pp.1-8. [Link], Mar. 9, 2026.

Abstract：
As software evolves, dependencies between program elements become increasingly complex. This complexity often results in incomplete changes, leading to bugs due to the developers’ inability to determine all impacted elements. To address this issue, previous work recommended co-change candidates at commit time. They do so based on a composite similarity over commits using textual information and changed items. However, when calculating the similarity of textual information, the semantics of code changes, which can serve as additional context, are not considered. Our proposed methods incorporate change semantics by deriving them from textual information using code-task pre-trained models. The goal is to allow the recommendation system to capture the overall context and characteristics of changes more accurately than existing methods.