Enhancing Missing Data Imputation Using Generative Adversarial Self-Attention Imputation Mechanism

カテゴリ: 研究会(論文単位)

論文No: CMN24032

グループ名: 【C】電子・情報・システム部門 通信研究会

発行日: 2024/03/25

タイトル(英語): Enhancing Missing Data Imputation Using Generative Adversarial Self-Attention Imputation Mechanism

著者名: Vellandurai Akhash(Hitachi India Private Limited),Sharma Ankit(Hitachi India Private Limited),Samon Thiruvengadam(Hitachi India Private Limited),Kumar Vinoth(Hitachi India Private Limited)

著者名(英語): Akhash Vellandurai(Hitachi India Private Limited),Ankit Sharma(Hitachi India Private Limited),Thiruvengadam Samon(Hitachi India Private Limited),Vinoth Kumar(Hitachi India Private Limited)

キーワード: Transportation|Smart cities|Missing Data| Imputation|Generative Adversarial Self-Attention Imputation|Mean Squared Error|Transportation|Smart cities|Missing Data| Imputation|Generative Adversarial Self-Attention Imputation|Mean Squared Error

要約(日本語): The intelligent bus transportation system is pivotal for smart city mobility, relying on accurate future demand estimation and optimized timetables. However, incomplete datasets pose challenges for data-driven improvements. This study introduces the Generative Adversarial Self Attention Imputation Network, merging Generative Adversarial Imputation Framework and Self Attention Mechanism which enhances capturing long-term dependencies. Further, loss function of generator is modified to integrate both reconstructions and discriminative loss making it adept for datasets with complex nonlinear relationships. It results in 10.374% improvement in Mean Squared Error over the next best algorithm Gaussian Process Variational Autoencoder.

要約(英語): The intelligent bus transportation system is pivotal for smart city mobility, relying on accurate future demand estimation and optimized timetables. However, incomplete datasets pose challenges for data-driven improvements. This study introduces the Generative Adversarial Self Attention Imputation Network, merging Generative Adversarial Imputation Framework and Self Attention Mechanism which enhances capturing long-term dependencies. Further, loss function of generator is modified to integrate both reconstructions and discriminative loss making it adept for datasets with complex nonlinear relationships. It results in 10.374% improvement in Mean Squared Error over the next best algorithm Gaussian Process Variational Autoencoder.

本誌: 2024年3月28日-2024年3月29日通信研究会

本誌掲載ページ: 83-88 p

原稿種別: 英語

PDFファイルサイズ: 1,011 Kバイト