Mohamed Maher Ben Ismail

Unsupervised clustering and feature weighting based on Generalized Dirichlet mixture modeling

Abstract We propose a possibilistic approach for Generalized Dirichlet mixture parameter estimation, data clustering, and feature weighting. The proposed algorithm, called Robust and Unsupervised Learning of Finite Generalized Dirichlet Mixture Models (RULe_GDM), exploits a property of the Generalized Dirichlet distributions that transforms the data to make the features independent and follow Beta distributions. Then, it learns optimal relevance weights for each feature within each cluster. This property makes RULe_GDM suitable for noisy and high-dimensional feature spaces. In addition, RULe_GDM associates two types of memberships with each data sample. The first one is the posterior probability and indicates how well a sample fits each estimated distribution. The second membership represents the degree of typicality and is used to identify and discard noise points and outliers. RULe_GDM minimizes one objective function which combines learning the two membership functions, distribution parameters, and the relevance weights for each feature within each distribution. We also extend our algorithm to find the optimal number of clusters in an unsupervised and efficient way by exploiting some properties of the possibilistic membership function. The performance of RULe_GDM is illustrated and compared to similar algorithms. We use synthetic data to illustrate its robustness to noisy and high dimensional features. We also compare our approach to other relevant algorithms using several standard data sets.

Endoscopy video summarization based on unsupervised learning and feature discrimination

We propose a novel endoscopy video summarization approach based on unsupervised learning and feature discrimination. The proposed learning approach partitions the collection of video frames into homogeneous categories based on their visual and temporal descriptors. Also, it generates possibilistic memberships in order to represent the degree of typicality of each video frame within every category, and reduce the influence of noise frames on the learning process. The algorithm learns iteratively the optimal relevance weight for each feature subset within each cluster. Moreover, it finds the optimal number of clusters in an unsupervised and efficient way by exploiting some properties of the possibilistic membership function. The endoscopy video summary consists of the most typical frames in all clusters after discarding noise frames. We compare the performance of the proposed algorithm with state-of-the-art learning approaches. We show that the possibilistic approach is more robust. The endoscopy videos collection includes more than 90k video frames.

Possibilistic Clustering Based on Robust Modeling of Finite Generalized Dirichlet Mixture

We propose a novel possibilistic clustering algorithm based on robust modelling of the Generalized Dirichlet (GD) finite mixture. The algorithm generates two types of membership degrees. The first one is a posterior probability that indicates the degree to which the point fits the estimated distribution. The second membership represents the degree of “typicality” and is used to indentify and discard noise points. The algorithm minimizes one objective function to optimize GD mixture parameters and possibilistic membership values. This optimization is done iteratively by dynamically updating the Dirichlet mixture parameters and the membership values in each iteration. We compare the performance of the proposed algorithm with an EM based approach. We show that the possibilistic approach is more robust.

Generic Evaluation Metrics for Hyperspectral Data Unmixing

AbstractWe propose novel generic performance metric for hyperspectral unmixing techniques. This relative metric compares two abundance matrices. The first one represents the unmixing result. The second matrix can be either another unmixing result or the ground truth of the hyperspectral scene. This metric starts by computing coincidence matrices corresponding to the two abundance matrices, then the comparison is carried out by computing statistics of the number of pairs of data points that have high abundances with respect to the same endmember for the first unmixing approach, but have large abundance differences with respect to the same endmember for the second unmixing technique, or large differences in both. The main advantage of this metric approach is that there is no need to pair the endmembers of the two unmixing approaches. Rather, it assumes that the pixels, which are considered as different/same material by one unmixing approach should also be considered different/same material by the other. Our...

Insult Detection in Social Network Comments Using Possibilistic Based Fusion Approach

This paper aims to propose a novel approach to automatically detect verbal offense in social network comments. It relies on a local approach that adapts the fusion method to different regions of the feature space in order to classify comments from social networks as insult or not. The proposed algorithm is formulated mathematically through the minimization of some objective function. It combines context identification and multi-algorithm fusion criteria into a joint objective function. This optimization is intended to produce contexts as compact clusters in subspaces of the high-dimensional feature space via possibilistic unsupervised learning and feature weighting. Our initial experiments have indicated that the proposed fusion approach outperforms individual classifiers and the global fusion method. Also, in order to validate the obtained results, we compared the performance of the proposed approach with related fusion methods.

Mixture analysis based on spectral summarization

We introduce a new spectral mixture analysis approach. The proposed Mixture Analysis based on Spectral Summarization (MASS) uses all the wavelengths of the hyperspectral image and assumes a convex geometry model in order to estimate the endmembers and their corresponding abundances. MASS unmixing technique is based on the information provided by the summarization of the hyperspectral image. The summarization is performed through a fuzzy partitioning of the hyperspectral scene.

Verbal offense detection in social network comments using novel fusion approach

We propose a framework for automatic verbal offense detection in social network comments. The proposed approach adapts a possibilistic based fusion method to different regions of the feature space in order to classify social network comments as insult or not. The proposed algorithm is formulated mathematically through the minimization of some objective functions. It combines context identification and multi-algorithm fusion criteria into a joint objective function. The optimization is intended to produce contexts as compact clusters in subspaces of the high-dimensional feature space via possibilistic unsupervised learning and feature discrimination. The clustering component associates a degree of typicality with each data sample in order to identify and reduce the influence of noise points and outliers. Also, the approach provides optimal fusion parameters for each context. Our initial experiments on synthetic datasets and standard SMS datasets indicate that the proposed fusion approach outperforms individual classifiers. Finally, the proposed system is assessed using real collection of social network comments, and compared to state-of-the-art fusion technique.

Survey on number of endmembers estimation techniques for hyperspectral data unmixing

Hyperspectral imagery is a main tool of remote sensing applications. As a signal is transmitted towards a given scene, reflected and scattered again by interacting with the various components of the atmosphere and the surface, the reflectance spectra analysis allows recognition and/or quantification of the materials. Hyperspectral image is three-dimensional data cube that containing the values of the radiation that has been collected over an area in a wide range of wavelengths. This hyperspectral data serves to identify the scene composition, and includes applications such as chemical analysis, plant and mineral recognition, and urban mapping. Each reflected signal to form the hyperspectral data can be a mixture of different materials (endmembers). Spectral unmixing is decomposing the hyperspectral image into pure spectral signatures of the materials in the scene, and proportioning every material at pixel location. Despite the recent advances in hyperspectral technology, spectral unmixing, which consists in finding a set of spectrally pure components (endmembers) and their associated fractions coverage for each pixel (abundances) in hyperspectral data, remains a challenging research field. Also, most the state-of-the-art approaches assume that the number of endmembers is known a priori. Some approaches have been proposed to estimate the number of endmembers. These approaches are first applied to the hyperspectral data in order to learn the number of endmembers. Then, the unmixing is performed given the learned parameter. In this paper, we review the number of endmembers estimation techniques used for hyperspectral data unmixing.

Empirical Comparison of Visual Descriptors for Multiple Bleeding Spots Recognition in Wireless Capsule Endoscopy Video

Wireless Capsule Endoscopy (WCE) is the latest technology able to screen intestinal anomalies at early stage. Although its convenience to the patient and its effectiveness to show small intestinal details, the physician diagnosis remains not straight forward and time consuming. Thus, a computer aid diagnosis would be helpful. In this paper, we focus on The Multiple Bleeding Spots (MBS) anomaly. We propose to conduct an empirical evaluation of four feature descriptors in a the challenging problem of MBS recognition on WCE video using the SVM classifier. The performance of the four descriptors is based on the assessment of the performance of the output of the SVM classifier.

SEMI-SUPERVISED FUZZY CLUSTERING WITH LEARNABLE CLUSTER DEPENDENT KERNELS

Many machine learning applications rely on learning distance functions with side information. Most of these distance metric learning approaches learns a Mahalanobis distance. While these approaches may work well when data is in low dimensionality, they become computationally expensive or even infeasible for high dimensional data. In this paper, we propose a novel method of learning nonlinear distance functions with side information while clustering the data. The new semi-supervised clustering approach is called Semi-Supervised Fuzzy clustering with Learnable Cluster dependent Kernels (SS-FLeCK). The proposed algorithm learns the underlying cluster-dependent dissimilarity measure while finding compact clusters in the given data set. The learned dissimilarity is based on a Gaussian kernel function with cluster dependent parameters. This objective function integrates penalty and reward cost functions. These cost functions are weighted by fuzzy membership degrees. Moreover, they use side-information in the form of a small set of constraints on which instances should or should not reside in the same cluster. The proposed algorithm uses only the pairwise relation between the feature vectors. This makes it applicable when similar objects cannot be represented by a single prototype. Using synthetic and real data sets, we show that SS-FLeCK outperforms several other algorithms.