Grigorios Tzortzis

The Global Kernel

Kernel k-means is an extension of the standard k-means clustering algorithm that identifies nonlinearly separable clusters. In order to overcome the cluster initialization problem associated with this method, we propose the global kernel k-means algorithm, a deterministic and incremental approach to kernel-based clustering. Our method adds one cluster at each stage, through a global search procedure consisting of several executions of kernel k-means from suitable initializations. This algorithm does not depend on cluster initialization, identifies nonlinearly separable clusters, and, due to its incremental nature and search procedure, locates near-optimal solutions avoiding poor local minima. Furthermore, two modifications are developed to reduce the computational cost that do not significantly affect the solution quality. The proposed methods are extended to handle weighted data points, which enables their application to graph partitioning. We experiment with several data sets and the proposed approach compares favorably to kernel k-means with random restarts.

k

Applying k-Means to minimize the sum of the intra-cluster variances is the most popular clustering approach. However, after a bad initialization, poor local optima can be easily obtained. To tackle the initialization problem of k-Means, we propose the MinMax k-Means algorithm, a method that assigns weights to the clusters relative to their variance and optimizes a weighted version of the k-Means objective. Weights are learned together with the cluster assignments, through an iterative procedure. The proposed weighting scheme limits the emergence of large variance clusters and allows high quality solutions to be systematically uncovered, irrespective of the initialization. Experiments verify the effectiveness of our approach and its robustness over bad initializations, as it compares favorably to both k-Means and other methods from the literature that consider the k-Means initialization problem. & 2014 Elsevier Ltd. All rights reserved.

-Means Algorithm for Clustering in Feature Space

Multiview clustering partitions a dataset into groups by simultaneously considering multiple representations (views) for the same instances. Hence, the information available in all views is exploited and this may substantially improve the clustering result obtained by using a single representation. Usually, in multiview algorithms all views are considered equally important, something that may lead to bad cluster assignments if a view is of poor quality. To deal with this problem, we propose a method that is built upon exemplar-based mixture models, called convex mixture models (CMMs). More specifically, we present a multiview clustering algorithm, based on training a weighted multiview CMM, that associates a weight with each view and learns these weights automatically. Our approach is computationally efficient and easy to implement, involving simple iterative computations. Experiments with several datasets confirm the advantages of assigning weights to the views and the superiority of our framework over single-view and unweighted multiview CMMs, as well as over another multiview algorithm which is based on kernel canonical correlation analysis.

The MinMax k-Means clustering algorithm

Kernel k-means is an extension of the standard k-means clustering algorithm that identifies nonlinearly separable clusters. In order to overcome the cluster initialization problem associated with this method, in this work we propose the global kernel k-means algorithm, a deterministic and incremental approach to kernel-based clustering. Our method adds one cluster at each stage through a global search procedure consisting of several executions of kernel k-means from suitable initializations. This algorithm does not depend on cluster initialization, identifies nonlinearly separable clusters and, due to its incremental nature and search procedure, locates near optimal solutions avoiding poor local minima. Furthermore a modification is proposed to reduce the computational cost that does not significantly affect the solution quality. We test the proposed methods on artificial data and also for the first time we employ kernel k-means for MRI segmentation along with a novel kernel. The proposed methods compare favorably to kernel k-means with random restarts.

Multiple View Clustering Using a Weighted Combination of Exemplar-Based Mixture Models

MOTIVATION Transient S-sulfenylation of cysteine thiols mediated by reactive oxygen species plays a critical role in pathology, physiology and cell signaling. Therefore, discovery of new S-sulfenylated sites in proteins is of great importance towards understanding how protein function is regulated upon redox conditions. RESULTS We developed PRESS (PRotEin S-Sulfenylation) web server, a server which can effectively predict the cysteine thiols of a protein that could undergo S-sulfenylation under redox conditions. We envisage that this server will boost and facilitate the discovery of new and currently unknown functions of proteins triggered upon redox conditions, signal regulation and transduction, thus uncovering the role of S-sulfenylation in human health and disease. AVAILABILITY AND IMPLEMENTATION The PRESS web server is freely available at http://press-sulfenylation.cse.uoi.gr/ CONTACTS agtzakos@gmail.com or gtzortzi@cs.uoi.gr SUPPLEMENTARY INFORMATION Supplementary data are available at Bioinformatics online.

The global kernel k-means clustering algorithm

This paper proposes a novel approach for spam filtering based on the use of Deep Belief Networks (DBNs). In contrast to conventional feedfoward neural networks having one or two hidden layers, DBNs are feedforward neural networks with many hidden layers. Until recently it was not clear how to initialize the weights of deep neural networks, which resulted in poor solutions with low generalization capabilities. A greedy layer-wise unsupervised algorithm was recently proposed to tackle this problem with successful results. In this work we present a methodology for spam detection based on DBNs and evaluate its performance on three widely used datasets. We also compare our method to Support Vector Machines (SVMs) which is the state-of-the-art method for spam filtering in terms of classification performance. Our experiments indicate that using DBNs to filter spam e-mails is a viable methodology, since they achieve similar or even better performance than SVMs on all three datasets.

PRESS: protein S-sulfenylation server

During the last years, there is increasing interest in analyzing social networks and modeling their dynamics at different scales. This work focuses on predicting the future form of communities, which represent the mesoscale structure of networks, while the communities arise as a result of user interaction. We employ several structural and temporal features to represent communities, along with their past form, that are used to formulate a supervised learning task to predict whether a community will continue as currently is, shrink, grow or completely disappear. To test our methodology, we created a real-life social network dataset consisting of an excerpt of posts from the Mathematics Stack Exchange Q&A site. In the experiments, special care is taken in handling the class imbalance in the dataset and in investigating how the past evolutions of a community affect predictions.

Deep Belief Networks for Spam Filtering

Maximum margin clustering (MMC) approaches extend the large margin principle of SVM to unsupervised learning with considerable success. In this work, we utilize the ratio between the margin and the intra-cluster variance, to explicitly consider both the separation and the compactness of the clusters in the objective. Moreover, we employ multiple kernel learning (MKL) to jointly learn the kernel and a partitioning of the instances, thus overcoming the kernel selection problem of MMC. Importantly, the margin alone cannot reliably reflect the quality of the learned kernel, as it can be enlarged by a simple scaling of the kernel. In contrast, our ratio-based objective is scale invariant and also invariant to the type of norm constraints on the kernel parameters. Optimization of the objective is performed using an iterative gradient-based algorithm. Comparative clustering experiments on various datasets demonstrate the effectiveness of the proposed formulation.

Predicting the evolution of communities in social networks using structural and temporal features

Multiple kernel learning (MKL) has emerged as a powerful tool for considering multiple kernels when the appropriate representation of the data is unknown. Some of these kernels may be complementary, while others irrelevant to the learning task. In this work we present an MKL method for clustering. The intra-cluster variance objective is extended by learning a linear combination of kernels, together with the cluster labels, through an iterative procedure. Closed-form updates for the combination weights are derived, that greatly simplify the optimization. Moreover, to allow for robust kernel mixtures, a parameter that regulates the sparsity of the weights is incorporated into our framework. Experiments conducted on a collection of images reveal the effectiveness of the proposed method.

Ratio-based multiple kernel clustering

Real world social networks are highly dynamic environments consisting of numerous users and communities, rendering the tracking of their evolution a challenging problem. In this work, we propose a parallel algorithm for tracking dynamic communities between consecutive timeframes of the social network, where communities are represented as undirected graphs. Our method compares the communities based on the widely adopted Jaccard similarity measure and is implemented on top of Apache Flink, a novel framework for parallel and distributed data processing. We evaluate the benefits, in terms of execution time, that parallel processing brings to community tracking on datasets carrying different quantitative characteristics, derived from two popular social media platforms; Twitter and Mathematics Stack Exchange Q&A. Experiments show that our parallel method has the ability to calculate the similarity of communities within seconds, even for large social networks, consisting of more than 600 communities per timeframe.