【論文要約】Code-to-Code Search Based on Deep Neural Network and Code Mutation

コードクローンの関連でチェックしたい論文：

Code-to-Code Search Based on Deep Neural Network and Code Mutation (osaka-u.ac.jp)


Abstract of the paper in 2 lines:

The paper presents an approach for code-to-code search using a Deep Neural Network (DNN) model and code mutation to generate sufficient code fragments for each label.

The preliminary experiment demonstrates high precision and recall of the proposed approach.[1]



What are the contributions of this paper

The paper presents an approach for code-to-code search based on a Deep Neural Network (DNN) model and code mutation, which generates sufficient code fragments for each label.[1]

The approach consists of two steps: STEP L (Learning) and STEP S (Search), enabling the identification of similar code fragments corresponding to a query code fragment.[2]

A case study conducted on three open-source software systems, HBase 2.04, OpenSSL 0.9.1 1.1.05, and FreeBSD 11.1.06, demonstrates the effectiveness of the approach in terms of precision, recall, and F-measure.[3]

The preliminary experiment shows high precision and recall of the proposed approach, indicating its potential for accurate code-to-code search.[1]

Note: The provided sources do not explicitly mention the contributions of the paper. However, based on the information provided in the abstract and case study section, these are the contributions inferred from the paper.


Related Papers

Here are some related papers:

1. Code function taste detection method based on deep semantics
Shi Chongyang+2 others • 2019

2. A Search-Based Testing Framework for Deep Neural Networks of Source Code Embedding
Maryam Vahdat Pour+3 others • 2021, arXiv: Software Engineering
1 citations

3. Semantic code search using Code2Vec: A bag-of-paths model
Lakshmanan Arumugam • 2020
4 citations

4. Interactive code searching method and device based on structured embedding
Peng Min+4 others • 2020

5. CodeT5: Identifier-aware Unified Pre-trained Encoder-Decoder Models for Code Understanding and Generation
Yue Wang+3 others • 2021, Empirical Methods in Natural Language Processing

Practical implications of this paper

Practical Implications of the Paper "Code-to-Code Search Based on Deep Neural Network and Code Mutation":

The proposed approach enables accurate code-to-code search by utilizing a Deep Neural Network (DNN) model and code mutation techniques.

It allows software engineers to search for syntactically similar source code, which can be beneficial for tasks such as code reuse, bug fixing, and understanding unfamiliar codebases.

The approach has been evaluated on multiple open-source software systems, demonstrating its effectiveness in terms of precision, recall, and F-measure.

By generating a sufficient number of code fragments for each label through code mutation, the approach overcomes the challenge of requiring a large number of code fragments for DNN model training.

The use of DNNs for code-to-code search expands the application of deep learning techniques in software engineering, offering a new approach for code analysis and exploration.

Summarize introduction of this paper

Introduction of the Paper "Code-to-Code Search Based on Deep Neural Network and Code Mutation":

Developers often need to find code fragments with similar functionality, such as more efficient or reliable implementations or less vulnerable alternatives.
Previous approaches have been proposed for code-to-code search, including the use of Deep Neural Networks (DNNs) for code clone detection.
The key insight of this paper is to utilize DNNs for labeling code fragments based on their functionality, enabling code-to-code search.
The approach involves two steps: STEP L (Learning) and STEP S (Search). In STEP L, mutated source code fragments are generated, clustered, and labeled with unique labels. The feed-forward neural network model learns tuples of feature vectors and labels. In STEP S, the trained model calculates a label for a query code fragment and retrieves the original code blocks corresponding to that label.
The effectiveness of the approach is confirmed through a case study on three open-source software systems.
[1][2]

Literature survey of this paper

Literature survey of the paper "Code-to-Code Search Based on Deep Neural Network and Code Mutation":

Deep Neural Networks (DNNs) have been commonly used for image labeling tasks, such as object detection.

The paper proposes the use of DNNs for code-to-code search, which is a novel application of deep learning in software engineering.

The approach presented in the paper utilizes code mutation techniques to generate a sufficient number of code fragments for each label in the learning process of DNNs[1].

The paper mentions the use of other deep networks, such as RNN, for code-to-code search, indicating the potential for further exploration and improvement of the approach[2].

The approach has been evaluated on multiple open-source software systems, including HBase, OpenSSL, and FreeBSD, demonstrating its performance in terms of precision, recall, and F-measure[3][4].

The paper also discusses the generation of feature vectors using techniques like Bag of Words (BoW) and Doc2Vec, highlighting the flexibility of the approach in handling different types of code representations[5].

Note: The provided sources do not explicitly mention a comprehensive literature survey. However, the information extracted from the sources provides an overview of the existing literature and the novelty of the proposed approach.

Methods used in this paper

Methods used in this paper:

Deep Neural Networks (DNNs) are utilized for code-to-code search, where code fragments are labeled based on their functionality .

Code mutation techniques are employed to generate a sufficient number of code fragments for each label in the learning process of DNNs .

The approach involves two steps: STEP L (Learning) and STEP S (Search). In STEP L, mutated source code fragments are generated, clustered, and labeled with unique labels. The feed-forward neural network model learns tuples of feature vectors and labels. In STEP S, the trained model calculates a label for a query code fragment and retrieves the original code blocks corresponding to that label [1].

The approach is evaluated using precision, recall, and F-measure as evaluation metrics [2].

A case study is conducted on three open-source software systems, including HBase, OpenSSL, and FreeBSD, to demonstrate the performance of the approach [3] [4].

Note: The provided sources do not explicitly mention the methods used in the paper. However, the information extracted from the sources provides an understanding of the approach and its evaluation.

What data has been used in this paper

Data used in this paper:

The paper utilizes a dataset created for training and evaluation purposes, following the steps outlined in STEP E1. Similar code block sets are extracted from the target open-source software (OSS), and sets with 100 or more code blocks are selected to ensure an adequate number of similar code blocks for training and evaluation [1].

Multiple versions of open-source software systems, including HBase, OpenSSL, and FreeBSD, are used in a case study to evaluate the performance of the approach [2].

Positive data, which are vectors generated from code blocks that meet specific conditions, are used in the experiments for evaluation of search performance .

The experiment involves the use of code fragments, feature vectors, and labels for training and testing the deep neural network model [3].

Note: The provided sources do not explicitly mention the specific details of the data used in the paper. However, the information extracted from the sources provides an understanding of the dataset creation, case study, and evaluation process.

Results of the paper

Results of the paper:

The preliminary experiment shows high precision and recall of the proposed approach for code-to-code search based on a deep neural network model and code mutation [1].

Table II presents the precision, recall, and F-score results of the experiments. Recall in experiments using HBase and OpenSSL is 1.000, indicating that the model can accurately search for syntactically similar code blocks .

However, Figure 10 shows an example of an incorrect search result, where the trained model suggests similarity between dissimilar code blocks. The differences in Figure 10 can potentially cause such cases [2].

Precision, recall, and F-measure are used as evaluation criteria for search performance [3] [4].

A case study is conducted on the effect of mutation, investigating the relationship between the number of positive data for training and the probability of correct label output by the trained model [5].

The performance of the approach is evaluated using precision, recall, and F-measure on benchmark datasets from HBase, OpenSSL, and FreeBSD [6].

Note: The provided sources do not explicitly mention the overall results of the paper. However, the information extracted from the sources provides insights into the performance evaluation and case study conducted in the paper.

コメント

コメントを投稿

このブログの人気の投稿

【論文メモ】A systematic literature review on source code similarity measurement and clone detection: techniques, applications, and challenges

  原論文 A systematic literature review on source code similarity measurement (arxiv.org) Abstract ソースコードの類似性を測定し評価することは、コード推薦、重複コード、盗作、マルウェア、およびスメル（異臭）検出など、幅広いアプリケーションにおける基本的なソフトウェアエンジニアリング活動です。本論文は、既存のアプローチとその特性に光を当てるため、コードの類似性測定と評価技術に関する体系的な文献レビューとメタ分析を提案します。私たちは最初に、4つのデジタルライブラリをクエリして1万以上の記事を見つけ、最終的に136件の主要な研究を抽出しました。研究は、その方法論、プログラミング言語、データセット、ツール、およびアプリケーションに従って分類されました。深い調査の結果、5つのアプリケーションドメインで8つの異なる技術を用いて動作するソフトウェアツールが80あります。ツールの約49％がJavaプログラムで動作し、37％がCおよびC++をサポートしていますが、多くのプログラミング言語はサポートされていません。注目すべき点は、ソースコードの類似性測定と重複コードに関連する12のデータセットが存在し、そのうちの8つのデータセットのみが公開されていることでした。信頼性のあるデータセットの不足、経験的評価、ハイブリッド手法、およびマルチパラダイム言語への焦点が、この分野の主な課題です。コードの類似性測定の新興アプリケーションは、メンテナンスに加えて開発フェーズに集中しています。
続きを読む

【論文メモ】＜2022＞コードクローン検索手法の調査

原論文：   _pdf (jst.go.jp) 自然文をクエリとしたコード検索手法 Leveraging usage similarity for effective retrieval of examples in code repositories [2], 2010 Leveraging usage similarity for effective retrieval of examples in code repositories | Proceedings of the eighteenth ACM SIGSOFT international symposium on Foundations of software engineering Deep Code Search [12], 2018 Deep Code Search | IEEE Conference Publication | IEEE Xplore Sourcerer: mining and searching internet-scale software repositories [25], 2009 Sourcerer: mining and searching internet-scale software repositories: Data Mining and Knowledge Discovery: Vol 18, No 2 (acm.org) コード片をクエリとしたコード検索手法  
続きを読む

【論文】A Survey on Causal Inference＜2021＞

原本： A Survey on Causal Inference (acm.org)   因果推論（コーサク推論）は、統計学、コンピュータサイエンス、教育学、公共政策、経済学など、多くの分野で長年にわたり重要な研究トピックです。近年、観察データから因果効果を推定することは、ランダム化比較試験と比較して、膨大な利用可能なデータと低予算の要求により、魅力的な研究方向になりつつあります。急速に発展している機械学習分野の恩恵を受け、観察データに対する因果効果推定手法がさまざまに考案されています。この調査では、因果推論のフレームワークとしてよく知られている「潜在的効果」フレームワークに基づく因果推論手法の包括的なレビューを提供します。手法は、潜在的効果フレームワークの3つの仮定すべてを必要とするかどうかによって、2つのカテゴリに分けられます。各カテゴリについて、従来の統計学的手法と最近の機械学習を用いた強化手法の両方について議論し、比較を行います。また、広告、レコメンデーション、医療などへの応用事例も紹介しています。さらに、ベンチマークデータセットやオープンソースコードについてもまとめており、研究者や実践者が因果推論手法の探索、評価、適用を行うことを容易にします。 1. Introduction 日常会話では、相関関係と因果関係は混同して使われることがよくありますが、実際には解釈が大きく異なります。相関関係は、2つの変数間に一般的に見られる関係を示します。例えば、2つの変数が共に増えたり減ったりするような傾向があれば、相関関係にあると言えます [7]。一方、因果関係は原因と結果の関係であり、原因がある程度結果を引き起こし、結果は原因に依存しているような関係です。因果推論とは、効果の発生条件に基づいて因果関係を結論付けるプロセスです。因果推論と相関関係の推論との主な違いは、前者が原因を変えたときに効果変数の反応を分析する点にあります [104, 150]。「相関関係は因果関係を示唆しない」というのは有名な言葉です。例えば、ある研究では、朝食を食べる女の子の方が朝食を食べない女の子よりも体重が軽いという結果が出たため、朝食はダイエットに効果があると結論付けました。しかし、実際にはこの2つの事象は相関関係にすぎず、因果関係ではない可能性があります。朝食を毎日食べる女の子は、運動をよ
続きを読む