Research
I am interested in Machine Learning and Pattern Recognition, especially simultaneous clustering and kernel learning. Clustering is an essential and very frequently performed task in pattern recognition and data mining. It can aid in a variety of tasks related to understanding and exploring the structure of large and high dimensional data. The goal of cluster analysis is to find natural groupings in a set of objects such that objects in the same cluster are as similar as possible and objects in different clusters are as dissimilar as possible. For some real world applications, the distribution of the features can be more challenging. In fact, the different categories may not be linearly separable in the original feature space. My research focuses on learning the underlining information and using it to learn a similarity measure that allows to better segregate the different categories of the data. To this end, I have worked on designing efficient clustering and learning algorithms capable of handling the overlapping boundaries among categories associated with large scale data sets. I have explored highly diverse machine learning aspects, including graphical and relational models, fuzzy clustering, spectral and diffusion mapping, distance measure and kernel learning. Furthermore, I explored machine learning techniques in the context of hyperspectral image analysis. I designed an endmember detection and spectral unmixing algorithms that simultaneously learn the underling fuzzy multiple convex geometry models and find a robust estimate of the model's endmembers and abundances.
Unsupervised kernel Learning
I designed and developed a new fuzzy relational clustering technique with Local Scaling Parameter Learning (LSPL). This approach learns the underlying cluster dependent dissimilarity measure while finding compact clusters in the given data set. The learned measure is a Gaussian similarity function defined with respect to each cluster that allows to control the scaling of the clusters and thus, improve the final partition. One objective function is minimized for both the optimal partition and for the cluster dependent scaling parameter. This optimization is done iteratively by dynamically updating the partition and the scaling parameter in each iteration. This makes the proposed algorithm simple and fast. Moreover, as the data is assumed to be available in a relational form, this approach is applicable even when only the degree to which pairs of objects in the data are related is available. It is also more practical when similar objects cannot be represented by a single prototype.
Semi-supervised kernel Learning
I designed and developed a new semi-supervised fuzzy clustering technique with adaptive local distance measure (SURF- LDML). This algorithm uses pair-wise relations between points. This makes it more suitable to identify clusters with arbitrary shapes that cannot be represented by a single prototype. SURF-LDML learns the underlying cluster-dependent dissimilarity measure while finding compact clusters. The learned distance controls the shape of the clusters and is a full ranked Mahalanobis measure. SURF-LDML minimizes a joint objective function that integrates penalty and reward cost functions. These cost functions are constructed using side information in the form of a small set of constraints on which instances should or should not reside in the same cluster. The penalty and reward terms are soft and depend on the pair-wise distance as well as the fuzzy membership of the points in the clusters. The distance dependency allows the adaptation of the distance measure to satisfy as many constraints as possible. The membership dependency makes the reward (penalty) term for satisfying (violating) a constraint stronger for pairs of points that are at the core of the cluster than points at the outskirt of the cluster. We compare the performance of SURF-LDML to other relational clustering methods. We show that the learned local distance provides better mapping and partitioning of the data.
Hyperspectral image analysis
Spectral analysis
I designed and developed a hyperspectral endmember detection and spectral unmixing algorithm that finds multiple sets of endmembers. This algorithm, the Piece-wise Convex Multiple Model Endmember Detection (P-COMMEND) algorithm, models a hyperspectral image using a piece-wise convex representation. Each hyperspectral pixel is represented as a weighted sum of convex combinations of endmembers. By using a piece-wise convex representation, non convex hyperspectral data are more accurately characterized. For example, the well-known Indian Pines hyperspectral image is used as an example of a piece-wise convex collection of pixels. The convex regions, weights, endmembers and abundances are found using an iterative fuzzy clustering method.
Simultaneous spacial and spectral analysis
I designed and developed a new spectral mixture analysis approach. Unlike most available approaches that only use the spectral information, this approach uses the spectral and spatial information available in the hyperspectral data. Moreover, it does not assume a global convex geometry model that encompasses all the data but rather multiple local convex models. Both the multiple model boundaries and the model’s endmembers and abundances are fuzzy. This allows points to belong to multiple groups with different membership degrees. Our approach is based on minimizing a joint objective function to simultaneously learn the underling fuzzy multiple convex geometry models and find a robust estimate of the model’s endmembers and abundances.
Referred Conference Papers
Ouiem Bchir and Hichem Frigui. Fuzzy relational kernel clustering with local scaling parameter learning. IEEE MLSP 2010 Kittila, Finland, August 2010.
Ouiem Bchir, Hichem Frigui, Alina Zare and Paul Gader. Multiple Model Endmember Detection Based On Spectral And Spatial Information. IEEE WHISPERS 2010, Reykjavik, Iceland , June 2010.
Alina Zare, Ouiem Bchir , Hichem Frigui and Paul Gader . Spatially-Smooth Piece-Wise Convex Endmember Detection. IEEE WHISPERS 2010, Reykjavik, Iceland, June 2010.
Alina Zare, Ouiem Bchir , Hichem Frigui and Paul Gader. A Comparison Of Deterministic And Probabilistic Approaches To Endmember Representation. IEEE WHISPERS 2010, Reykjavik, Iceland, June 2010.
Referred Journal Papers
Alina Zare, Ouiem Bchir , Hichem Frigui and Paul Gader. Piece-wise Convex Multiple Model Endmember Detection. IEEE TGRS 2010. (under revision).
Ouiem Bchir and Hichem Frigui. Fuzzy clustering with learnable cluster dependent kernels. IEEE TFS. (under revision)