Helsinki Institute for Information Technology

About: Helsinki Institute for Information Technology is a facility organization based out in Espoo, Finland. It is known for research contribution in the topics: Population & Bayesian network. The organization has 630 authors who have published 1962 publications receiving 63426 citations.

Safe and Complete Contig Assembly Through Omnitigs

Abstract: Contig assembly is the first stage that most assemblers solve when reconstructing a genome from a set of reads. Its output consists of contigs—a set of strings that are promised to appear in any genome that could have generated the reads. From the introduction of contigs 20 years ago, assemblers have tried to obtain longer and longer contigs, but the following question remains: given a genome graph G (e.g., a de Bruijn, or a string graph), what are all the strings that can be safely reported from G as contigs? In this article, we answer this question using a model in which the genome is a circular covering walk. We also give a polynomial-time algorithm to find such strings, which we call omnitigs. Our experiments show that omnitigs are 66%–82% longer on average than the popular unitigs, and 29% of dbSNP locations have more neighbors in omnitigs than in unitigs.

Hierarchical architectures in structured peer-to-peer overlay networks

Abstract: Distributed Hash Tables (DHT) are presently used in several large-scale systems in the Internet and envisaged as a key mechanism to provide identifier-locator separation for mobile hosts in Future Internet. Such P2P-based systems become increasingly complex serving popular social networking, resource sharing applications, and Internet-scale infrastructures. Hierarchy is a standard mechanism for coping with heterogeneity and scalability in distributed systems. To address the shortcomings of flat DHT designs, many hierarchical P2P designs have been proposed over recent years. The last generation is hierarchical DHTs (HDHTs) where nodes are organized onto layers and groups. This article discusses hierarchical architectures applied in structured P2P overlay networks, focusing on HDHT designs. We introduce a framework consisting of conceptual models of network hierarchy, multi-layer hierarchical DHT architectures, principles affecting the design choices, and cost models for system tradeoff analysis, performance evaluation, and scalability estimation. Based on the framework we provide a taxonomy and survey more than 20 hierarchical HDHT proposals.

Quotient Normalized Maximum Likelihood Criterion for Learning Bayesian Network Structures

Abstract: We introduce an information theoretic criterion for Bayesian network structure learning which we call quotient normalized maximum likelihood (qNML). In contrast to the closely related factorized normalized maximum likelihood criterion, qNML satisfies the property of score equivalence. It is also decomposable and completely free of adjustable hyperparameters. For practical computations, we identify a remarkably accurate approximation proposed earlier by Szpankowski and Weinberger. Experiments on both simulated and real data demonstrate that the new criterion leads to parsimonious models with good predictive accuracy.

Bayesian exponential family projections for coupled data sources

Abstract: Exponential family extensions of principal component analysis (EPCA) have received a considerable amount of attention in recent years, demonstrating the growing need for basic modeling tools that do not assume the squared loss or Gaussian distribution. We extend the EPCA model toolbox by presenting the first exponential family multi-view learning methods of the partial least squares and canonical correlation analysis, based on a unified representation of EPCA as matrix factorization of the natural parameters of exponential family. The models are based on a new family of priors that are generally usable for all such factorizations. We also introduce new inference strategies, and demonstrate how the methods outperform earlier ones when the Gaussianity assumption does not hold.

Live and learn from mistakes: A lightweight system for document classification

Abstract: We present a Life-Long Learning from Mistakes (3LM) algorithm for document classification, which could be used in various scenarios such as spam filtering, blog classification, and web resource categorization. We extend the ideas of online clustering and batch-mode centroid-based classification to online learning with negative feedback. The 3LM is a competitive learning algorithm, which avoids over-smoothing, characteristic of the centroid-based classifiers, by using a different class representative, which we call clusterhead. The clusterheads competing for vector-space dominance are drawn toward misclassified documents, eventually bringing the model to a ''balanced state'' for a fixed distribution of documents. Subsequently, the clusterheads oscillate between the misclassified documents, heuristically minimizing the rate of misclassifications, an NP-complete problem. Further, the 3LM algorithm prevents over-fitting by ''leashing'' the clusterheads to their respective centroids. A clusterhead provably converges if its class can be separated by a hyper-plane from all other classes. Lifelong learning with fixed learning rate allows 3LM to adapt to possibly changing distribution of the data and continually learn and unlearn document classes. We report on our experiments, which demonstrate high accuracy of document classification on Reuters21578, OHSUMED, and TREC07p-spam datasets. The 3LM algorithm did not show over-fitting, while consistently outperforming centroid-based, Naive Bayes, C4.5, AdaBoost, kNN, and SVM whose accuracy had been reported on the same three corpora.