Visualizing Intrinsic Space for Spatial Data via Input Regularized Gaussian Process Latent Variable Models

• Item type: SIG Technical Reports(1)
• 公開日: 2016-07-18
• タイトル: Visualizing Intrinsic Space for Spatial Data via Input Regularized Gaussian Process Latent Variable Models
• タイトル / 言語: en
• タイトル: Visualizing Intrinsic Space for Spatial Data via Input Regularized Gaussian Process Latent Variable Models
• 言語: eng
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: NTT Communication Science Laboratories
• 著者所属: NTT Communication Science Laboratories
• 著者所属(英): en
• 著者所属(英): NTT Communication Science Laboratories
• 著者所属(英): en
• 著者所属(英): NTT Communication Science Laboratories
• 著者名: Tomoharu, Iwata; Naonori, Ueda
• 著者名(英): Tomoharu, Iwata; Naonori, Ueda
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: We propose the input-regularized Gaussian process latent variable model for visualizing a latent intrinsic input space that improves interpolation performance in regression tasks. The proposed model assumes that a latent location is associated with each observed input location, and the covariance function is determined by distance between the latent locations. The latent locations are estimated so that the output covariance of the given data is appropriately captured by the latent locations while preserving the neighbor relationships between the observed input space and the latent space by input regularization. When the input regularization is omitted, the proposed model reduces to the Gaussian process latent variable model. When the input regularization is strong enough to perfectly preserve the neighbor relationships, the proposed model becomes Gaussian process regression. The degree of the regularization is controlled by a hyperparameter, which can be automatically selected by cross-validation using the given data. We demonstrate the effectiveness of the proposed model with real-world spatial data sets in terms of interpolation performance of multiple output values.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: We propose the input-regularized Gaussian process latent variable model for visualizing a latent intrinsic input space that improves interpolation performance in regression tasks. The proposed model assumes that a latent location is associated with each observed input location, and the covariance function is determined by distance between the latent locations. The latent locations are estimated so that the output covariance of the given data is appropriately captured by the latent locations while preserving the neighbor relationships between the observed input space and the latent space by input regularization. When the input regularization is omitted, the proposed model reduces to the Gaussian process latent variable model. When the input regularization is strong enough to perfectly preserve the neighbor relationships, the proposed model becomes Gaussian process regression. The degree of the regularization is controlled by a hyperparameter, which can be automatically selected by cross-validation using the given data. We demonstrate the effectiveness of the proposed model with real-world spatial data sets in terms of interpolation performance of multiple output values.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN10505667
• 書誌情報: 研究報告数理モデル化と問題解決（MPS）; 巻 2016-MPS-109, 号 5, p. 1-4, 発行日 2016-07-18
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8833
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会