| 内田 翔一朗 | D, 中間発表 | 計算システムズ生物学 | 金谷 重彦, | 松本 健一, | 小野 直亮, | MD.Altaf-Ul-Amin | |
|
title: Comparative Study of Integrative Analysis Methods for Soil Microbiome and Soil Chemistry Data
abstract: This study investigates computational methods for the integrative analysis of soil microbiome and soil chemistry data. Understanding the relationships between soil microbial communities and soil chemical properties is essential for ecological research and agricultural applications. However, conventional linear methods may not adequately capture the complex relationships between these heterogeneous data types. As an initial step, Canonical Correlation Analysis (CCA) and Deep Canonical Correlation Analysis (DCCA) were applied to paired soil microbiome and soil chemistry datasets, and the resulting latent representations were compared. In addition, axis-wise correspondence and latent-space similarity were evaluated, and permutation tests were performed to statistically assess the validity and stability of the shared structures identified by each method. Based on these analyses, the characteristics and applicability of each integrative analysis method are discussed. Future work will expand the comparison to additional computational approaches, including kernel-based and deep learning methods, with the goal of establishing an effective computational framework for the integrative analysis of soil microbiome and soil chemistry data. language of the presentation: Japanese 発表題目: 土壌微生物・土壌化学データの統合解析手法の比較検討 発表概要: 本研究では、土壌微生物群集データと土壌化学データの統合解析手法について検討する。土壌微生物群集と土壌化学性の関係を適切に理解することは、生態系の理解や農業・環境分野への応用において重要である。しかし、従来広く用いられている線形手法では、両データ間に存在する複雑な関係を十分に捉えられない可能性がある。そこで本研究では、Canonical Correlation Analysis（CCA）およびDeep Canonical Correlation Analysis（DCCA）を用いて、土壌微生物データと土壌化学データの統合解析を行い、各手法によって得られる潜在表現の比較と評価を実施した。さらに、軸ごとの対応関係や潜在空間全体の類似性を評価することで、各手法の特徴を検討した。今後は比較対象をカーネル法や深層学習を含む他の統合解析手法へ拡張し、土壌微生物・土壌化学データの解析に適した計算手法の確立を目指す。 | |||||||
| 渡邊 幹 | M, 2回目発表 | ソフトウェア設計学 | 飯田 元, | 松本 健一, | 柏 祐太郎, | Reid Brittany, | Ji Zhenlan |
|
title: *** Predicting Unnecessary Methods in Agentic Code Generation ***
abstract: *** Agentic Coding, powered by autonomous agents such as GitHub Copilot and Cursor, enables developers to generate code, tests, and pull requests from natural language instructions alone. While this accelerates implementation, it produces larger volumes of code per pull request, shifting the burden from implementers to reviewers. In practice, a notable portion of AI-generated code is eventually deleted during review, yet reviewers must still examine such code before deciding to remove it. No prior work has explored methods to help reviewers efficiently identify code that will be removed. In this paper, we propose a prediction model that identifies functions likely to be deleted during PR review. Our results show that functions deleted for different reasons exhibit distinct characteristics, and our model achieves an AUC of 87.1%. These findings suggest that predictive approaches can help reviewers prioritize their efforts on essential code. *** language of the presentation: *** Japanese *** 発表題目: *** エージェント型コード生成における不要なメソッドの予測 *** 発表概要: *** GitHub Copilot や Cursor といった自律的なエージェントによって実現されるエージェント型コーディング（Agentic Coding）は，自然言語による指示のみから，コード・テスト・プルリクエストを生成することを可能にする．これは実装を加速させる一方で，プルリクエストあたりのコード量を増大させ，負担を実装者からレビュアーへと移すことになる．実際には，AI が生成したコードの少なからぬ部分は最終的にレビューの過程で削除されるが，レビュアーは削除を決定する前にそうしたコードを精査しなければならない．削除されるコードをレビュアーが効率的に特定するための手法は，これまで検討されてこなかった．本論文では，PR レビューの過程で削除される可能性の高い関数を特定する予測モデルを提案する．実験の結果，異なる理由で削除された関数はそれぞれ異なる特徴を示すことが明らかになり，提案モデルは 87.1% の AUC を達成した．これらの結果は，予測に基づくアプローチが，レビュアーが本質的なコードに労力を集中させる助けになりうることを示唆している． *** | |||||||
| SHAKIZADA YAGUT | M, 2回目発表 | ソフトウェア設計学 | 飯田 元, | 松本 健一, | 柏 祐太郎, | Reid Brittany, | Ji Zhenlan |
|
title: The Roles of Humans in AI-Centric Development: A Large-Scale Study of Multi-Agent Pull Request Workflows
abstract: Large Language Models are increasingly used not just to generate code, but to author and review pull requests with minimal human involvement. This creates a multi-agent ecosystem where multiple AI systems and human developers interact throughout the PR lifecycle — a setting that remains poorly understood. We study this ecosystem using the AIDev dataset, which contains 931,785 pull requests authored by five AI coding agents (GitHub Copilot, OpenAI Codex, Claude Code, Cursor, and Devin). Across five research questions, we examine what AI combinations appear in authoring and reviewing (RQ1), whether different author–reviewer configurations lead to different levels of suggested changes (RQ2), how much human developers intervene (RQ3), how AI and human review feedback differs (RQ4), and whether AI-generated suggestions actually get implemented (RQ5). Our methods combine large-scale repository mining, an LLM-as-a-Judge classifier validated through inter-rater agreement with human annotators, and a taxonomy grounded in Bacchelli Bird’s code review framework. Three findings stand out. First, over 93% of AI-authored PRs receive no review interaction at all — and when review does happen, AI reviewers are involved more often than humans. Second, AI and human reviewers focus on different things: AI reviewers concentrate on defects and code improvements, while humans ask more clarification questions and transfer knowledge. Third, despite the "AI-authored" label, over 97% of these PRs contain at least one human commit, meaning AI currently functions more as a drafting tool than an autonomous contributor. Among suggestions with available diff evidence, roughly 40% are implemented in the final code. language of the presentation: English | |||||||
| LIU SIYUAN | M, 2回目発表 | ソフトウェア設計学 | 飯田 元, | 松本 健一, | 柏 祐太郎, | Reid Brittany, | Ji Zhenlan |
|
title: Evaluating Cross-Language Transfer for Refactoring Detection with Large Language Models
abstract: The ability to detect refactoring is crucial to understanding software evolution. Many studies have proposed automated refactoring detection tools, such as RefactoringMiner and PyRef, but most are language-specific. Developing tools for new languages requires creating custom rules for each language, which is time-consuming and labor-intensive. This paper empirically evaluates the potential of Cross-Language Transfer (CLT) Learning using Large Language Models (LLMs) as a more general and cost-effective alternative for refactoring detection. Specifically, we compare CLT prompting against zero-shot prompting and native few-shot prompting. In CLT prompting, examples from one language (Java) are used to detect refactorings in another language (Python). In native few-shot prompting, examples are written in the same language as the target. Our results show that cross-language (Java → Python) transfer is highly effective, significantly outperforming the zero-shot prompting (Avg. F1 0.501 vs. 0.430). Surprisingly, it nearly matches the performance of native few-shot prompting (Python → Python, Avg. F1 0.506). These findings suggest that CLT is a promising direction for bypassing language-specific tool development. language of the presentation: English | |||||||