Co-embedding of Structurally Missing Data by Locally Linear Alignment
This paper proposes a “co-embedding” method to embed the row and column vectors of an observation matrix data whose large portion is structurally missing into low-dimensional latent spaces simultaneously. A remarkable characteristic of this method is that the co-embedding is efficiently obtained via eigendecomposition of a matrix, unlike the conventional methods which require iterative estimation of missing values and suffer from local optima. Besides, we extend the unsupervised co-embedding method to a semi-supervised version, which is reduced to a system of linear equations.In an experimental study, we apply the proposed method to two kinds of tasks – (1) Structure from Motion (SFM) and (2) Simultaneous Localization and Mapping (SLAM).
KeywordsDimensionality Reduction Singular Value Decomposition Latent Vector Wireless Device Latent Semantic Indexing
Unable to display preview. Download preview PDF.
- 2.Pan, J., Yang, Q.: Co-localization from labeled and unlabeled data using graph laplacian. In: Proceedings of IJCAI 2007, pp. 2166–2171 (2007)Google Scholar
- 4.Roweis, S.: Em algorithms for pca and spca. In: in Advances in Neural Information Processing Systems, pp. 626–632 (1998)Google Scholar
- 6.Sindhwani, V., Bucak, S.S., Hu, J., Mojsilovic, A.: One-class matrix completion with lowdensity factorizations. In: Proceedings of the 2010 IEEE International Conference on Data Mining, pp. 1055–1060 (2010)Google Scholar
- 8.Thrun, S., Burgard, W., Fox, D.: Probabilistic Robotics. MIT Press (2005)Google Scholar
- 9.Verbeek, J., Roweis, S.T., Vlassis, N.: Non-linear cca and pca by alignment of local models. In: Procs. of NIPS (2004)Google Scholar