Published papers

Currently available tools for semantic clone detection have serious type limitations related to objects, in particular related to instantiation and comparison. The trend in the recent research has focused on improving the performance rather than removing these constraints. This paper proposes a test-based approach for semantic clone detection in Java which overcomes these problems. The test-based approach consists in running two methods with the same input and comparing their outputs. For object instantiation, the approach uses EvoSuite, an automatic JUnit test generator, to create the tests with correct instantiations of all the needed classes, which are then modified to make them suitable for semantic clone detection. The output comparison uses the DeepHash function to obtain numerical values for objects, considering the values of each instance variable. The approach was evaluated with SemanticCloneBench. The results show that the proposed approach performs very well in terms of precision, achieving 94.4%, which is about 8% higher than the reported precision of SLACC, a recent semantic clone detector. Furthermore, out of 228 methods used for the evaluation, the approach was not able to perform the detection for only ten methods.

With the development of the game industry, the maturity of online game sales platforms, and the increasing complexity of game software itself, game companies need to analyze massive amounts of user reviews to understand the hidden defects of the game and the direction of future iterations. Manually reading game reviews is a labor-intensive and time-consuming task, as the number of reviews can go up to several thousand per day. Automatically classifying these game reviews will help alleviate this issue, but traditional classifiers will need a large number of labeled instances for training. In this paper, we propose and implement an approach that combines transfer learning in natural language processing (BERT), unsupervised learning, and active learning to classify game reviews using only a small amount of labeled instances. We found that our approach obtains 88.8% classification accuracy with only 100 labeled training instances. Our implementation can be extended to handle different types of new games by using a small amount of extra labeled instances and manual work.

There have been various studies on the automation of mobile app testing. Typical methods for automated testing of mobile apps are based on random search and on building state transition models. But there are problems in terms of the efficiency of search and accuracy of model building. This paper focuses on applying reinforcement learning to testing of mobile apps, especially issues such as explosion of the number of states, fixed rewards for transitions, and difficulty in convergence of learning. We focus on state definition, reward function, and a learning method to solve these problems. Specifically, we define states using discrete values of UI (User Interface) information on the screen, def ine adynamic reward function, and perform periodic learning by using the transition history. The proposed method is implemented and evaluated. Evaluation results show that our proposed approach shows 1.21 times higher coverage than an existing tool using reinforcement learning.

Many approaches have been proposed to detect code clones, which are basically similar code fragments. Most approaches are based on textual similarity. These approaches cannot detect semantic code clones, which are clones that have the same functionality but implemented with different syntax. Two functions can be considered to have the same functionality, when the output is the same given the same input. In order to appropriately generate inputs, we propose applying symbolic execution to semantic code clone detection. These functions are executed to obtain outputs, which are compared to determine if function pairs are clones. Our approach also does not limit output to return values; we also handle arrays and pointers as output, as the execution of the function may cause changes in their values. Furthermore, we classify types to enable cases where the types of inputs and/or outputs are not exactly the same. We evaluate our approach on SemanticCloneBench.

Software defects can lead to severe issues in software systems, such as software errors, security vulnerabilities, and decreased software performance. Early prediction of software defects can prevent these problems, reduce development costs, and enhance system reliability. However, existing methods often focus on manually crafted code features and overlook the rich semantic and contextual information in program code. In this paper, we propose a novel approach that integrates JavaBERT-based embeddings with a CNN-BiLSTM model for software defect prediction. Our model considers code context and captures code patterns and dependencies throughout the code, thereby improving prediction performance. We incorporate Optuna to find optimal hyperparameters. We conducted experiments on the PROMISE dataset, which demonstrated that our approach outperforms baseline models, particularly in leveraging code semantics to enhance defect prediction performance.

Recent years have seen an increase of decision-making software based on Machine Learning (ML). Multiple cases have been reported where such software are discriminatory based on attributes such as race and gender. Thus, ML-based decision making software need to be tested for discrimination, or fairness testing. One popular approach to fairness testing is to find discriminatory data items by first conducting a global search, and then searching locally near the found discriminatory data items. Aequitas, CGFT, and KOSEI are three examples taking this approach. However, there are issues in terms of stability and efficiency. We thus propose an approach called Robust Search Using T-way Testing (RSUTT), which is based on CGFT for global search and KOSEI for local search. Experiments showed that RSUTT performs more efficiently compared to Aequitas, CGFT, and KOSEI.

コーディング規約違反検出のための静的解析ツールでは，違反に対して警告すべき優先度が設定されていることがある．本研究では，多数のOSSプロジェクトを解析することによって，違反に対する優先度が実際に開発者にとって有益であるかどうかを明らかにする．

ソフトウェア開発において，開発者が守るべきと考えるようなコーディングスタイルに関する基準(コーディング規約)が存在する．しかし，実際の開発プロジェクトで必ずしも遵守されているとは限らない．我々は，OSSの1,000 プロジェクトのリポジトリを対象に，C++向けコーディング規約準拠検査ツールCpplintに実装された 87 種類のコーディング規約に対するコーディング規約違反の発生と，発生した違反へのリファクタリング等による対処の状況を調査した．その結果，多くのプロジェクトで開発者が「あたりまえ」に守っており，あまり違反が発生せずに発生しても対処される傾向が強い規約と，逆に多くのプロジェクトで開発者にあまり意識されず違反が多く発生し，発生した違反に対しても無視される傾向が強い規約が存在することを確認した．また，遵守の傾向が開発期間内で変化するかどうかを調査したところ，IDEにより自動修正が行われる一部の規約については，プロジェクトの特定期間で遵守の傾向が変化するプロジェクト数が，他の規約とは有意に異なり多いことを確認した．それらの規約については，開発者により守るべきという共通認識が形成されていない状況において，IDEの自動修正により開発者の意図と関係なく結果的に遵守している可能性があることを示唆している．

Android applications are event driven. Developers implement event handlers by inheriting and then coding callback methods that are defined in the Android API. If developers forget to implement relevant callback methods, the application may behave in an unexpected manner. In this paper, we propose an approach that detects methods that are currently unimplemented, but should be implemented. Detected methods are shown to developers for their consideration. We focus on callback methods that are frequently imple-mented, as well as pairs of methods that are frequently implemented together. Our approach first extracts such methods, method pairs and their frequencies from existing projects. Then, the extracted information is used to detect methods that are unimplemented but should be implemented, which are then shown to the developers. We implemented our approach as a plug-in for Android Studio. Our evaluation showed that our approach was able to detect more than half of the unimplemented methods that should be implemented within the top 15% of all unimplemented callback methods.

Much work has been done on automating regression testing for applications. But most of them focus on test execution. Little work has been done on automatically determining if a test case passes or fails. This decision is often made by comparing the results of executing test cases on a base version of the application and post-modification version of the application. If the two results match, the test case passes, otherwise fails. However, to the best of our knowledge, there is no regression testing method for automatically deciding pass/fail of dynamic Web applications which use JavaScript or CSS. We propose a method that automatically decides if a dynamic Web application passes a regression test case. The basic idea is to obtain a screenshot each time the GUI of the Web application (i.e. Web page) changes its state, and then compare each corresponding screenshot to see if they match. The evaluation results showed that the accuracy rate of our approach is high and our approach can be considered as fast enough for practical use.

© 2020, The Author(s). Various computing subdisciplines, such as computer science and software engineering, each have their own curricular guidelines. They can be very difficult to understand and compare for people such as prospective students, industry personnel, and even faculty members. This is compounded by a lack of information surrounding undergraduate computing curricular topics via visual methods. This paper describes two experimental activities where the objective is to explore the possibility of obtaining quantitative data sets necessary for visualization, one based on competencies and the other based on knowledge areas. Both activities were based on surveys. The results from the first activity showed that a consensus interpretation could be obtained for the knowledge, skills, and dispositions implied by the competency descriptions, although not as strongly for dispositions. The second activity resulted in a table of knowledge areas with minimum and maximum weights for six computing subdisciplines. Finally, this paper also shows two examples of how users can explore the various curricular guidelines through visualization.