Graph Embedding and Extensions: A General Framework for Dimensionality Reduction IEEE TRANSSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE Shuicheng Yan, Dong Xu, Benyu Zhang, Hong-Jiang Zhang, Qiang Yang, Stephen Lin Presented by meconin Outline • • • • • Introduction Graph Embedding (GE) Marginal Fisher Analysis (MFA) Experiments Conclusion and Future Work Introduction • Dimensionality Reduction – Linear • PCA, LDA, are the two most popular due to simplicity and effectiveness • LPP, preserves local relationships in the data set, and uncovers its essential manifold structure Introduction • Dimensionality Reduction – For nonlinear methods, ISOMAP, LLE, Laplacian Eigenmap are three algorithms have been developed recently – Kernel trick: • linear methods → nonlinear ones • performing linear operations on higher or even infinite dimensional by kernel mapping function Introduction • Dimensionality Reduction – Tensor based algorithms • 2DPCA, 2DLDA, DATER Introduction • Graph Embedding is a general framework for dimensionality reduction – With it’s linearization, kernelization, and tensorization, we have a unified view for understanding DR algorithms – The above-mentioned algorithms can all be reformulated with in it Introduction • This paper show that GE can be used as a platform for developing new DR algorithms – Marginal Fisher Analysis (MFA) • Overcome the limitations of LDA Introduction • LDA (Linear Discriminant Analysis) – Find the linear combination of features best separate classes of objects – Number of available projection directions is lower than class number – Based upon interclass and intraclass scatters, optimal only when the data of each class is approximately Gaussian distributed Introduction • MFA advantage: (compare with LDA) – The number of available projection directions is much larger – No assumption on the data distribution, more general for discriminant analysis – The interclass margin can better characterize the separability of different classes Graph Embedding • For classification problem, the sample set is represented as a matrix X = [x1, x2, …, xN], xi Rm • In practice, the feature dimension m is often very high, thus it’s necessary to transform the data to a low-dimensional one yi = F(xi), for all i Graph Embedding Graph Embedding • Different motivations of DR algorithms, their objectives are similar – to derive lower dimensional representation • Can we reformulate them within a unifying framework? Whether the framework assists design new algorithms? Graph Embedding • Give a possible answer – Represent each vertex of a graph as a low-dimensional vector that preserves similarities between the vertex pairs – The similarity matrix of the graph characterizes certain statistical or geometric properties of the data set Graph Embedding • G = { X, W } be an undirected weighted graph with vertex set X and similarity matrix W RNN • The diagonal matrix D and the Laplacian matrix L of a graph G are defined as L = D W, Dii = W , i j i Wij j i ij Graph Embedding • Graph embedding of G is an algorithm to find low-dimensional vector representations relationships among the vertices of G • • B is the constraint matrix, and d is a constant, for avoid trivial solution Graph Embedding • For larger similarity between samples xi and xj, the distance between yi and yj should be smaller to minimize the objective function • To offer mappings for data points throughout the entire feature space – Linearization, Kernelization, Tensorization Graph Embedding • Linearization Assuming y = XTw • Kernelization : x F, assuming Graph Embedding • The solutions are obtained by solving the generalized eigenvalue decomposition problem • F. Chung, “Spectral Graph Theory,” Regional Conf. Series in Math.,no. 92, 1997 Graph Embedding • Tensor – the extracted feature from an object may contain higher-order structure – Ex: • an image is a second-order tensor • sequential data such as video sequences is a third-order tensor Graph Embedding • Tensor – In n dimensional space, nr directions, r is the rank(order) of a tensor – For tensor A, B Rm m …m the inner product 1 2 n Graph Embedding • Tensor – For a matrix U Rm m’ , B = A k U k – k Graph Embedding • The objective funtion: • In many case, there is no closedform solution, but we can obtain the local optimum by fixing the projection vector General Framework for DR • The differences of DR algorithms: – the computation of the similarity matrix of the graph – the selection of the constraint matrix General Framework for DR General Framework for DR • PCA – seeks projection directions with maximal variances – it finds and removes the projection direction with minimal variance General Framework for DR • KPCA – applies the kernel trick on PCA, hence it is a kernelization of graph embedding • 2DPCA is a simplified second-order tensorization of PCA and only optimizes one projection direction General Framework for DR • LDA – searches for the directions that are most effective for discrimination by minimizing the ratio between the intraclass and interclass scatters General Framework for DR • LDA General Framework for DR • LDA – follows the linearization of graph embedding – the intrinsic graph connects all the pairs with same class labels – the weights are in inverse proportion to the sample size of the corresponding class General Framework for DR • The intrinsic graph of PCA is used as the penalty graph of LDA PCA LDA General Framework for DR • KDA is the kernel extension of LDA • 2DLDA is the second-order tensorization of LDA • DATER is the tensorization of LDA in arbitrary order General Framework for DR • • • • LLP ISOMAP LLE Laplacian Eigenmap (LE) Related Works • Kernel Interpretation – Ham et al. – KPCA, ISOMAP, LLE, LE share a common KPCA formulation with different kernel definitions – Kernel matrix v.s Laplacian matrix from similarity matrix – Only unsupervised v.s more general Related Works • Out-of-Sample Extension – Brand – Mentioned the concept of graph embedding – Brand’s work can be considered as a special case of our graph embedding Related Works • Laplacian Eigenmap – Work with only a single graph, i.e., the intrinsic graph, and cannot be used to explain algorithms such as ISOMAP, LLE, and LDA – Some works use a Gaussian function to compute the nonnegative similarity matrix Marginal Fisher Analysis • Marginal Fisher Analysis Marginal Fisher Analysis • Intraclass compactness (intrinsic graph) Marginal Fisher Analysis • Interclass separability (penalty graph) The first step of MFA The second step of MFA Marginal Fisher Analysis • Intraclass compactness (intrinsic graph) Marginal Fisher Analysis • Interclass separability (penalty graph) The third step of MFA First of Four steps of MFA LDA v.s MFA 1. The available projection directions are much greater than that of LDA 2. There is no assumption on the data distribution of each class 3. The interclass margin in MFA can better characterize the separability of different classes than the interclass variance in LDA Kernel MFA • • The distance between two samples • For a new data point x, its projection to the derived optimal direction Tensor MFA Experiments • Face Recognition – XM2VTS, CMU PIE, ORL • A Non-Gaussian Case Experiments • XM2VTS, PIE-1, PIE-2, ORL Experiments Experiments Experiments Experiments Experiments