Annalisa Appice

Title: Spatio-temporal Data Mining: Part 2

Abstract: Emerging real life applications, such as environmental compliance, ecological studies and meteorology, are characterized by real-time acquisition of spatio-temporal data through a number of (wireless) remote sensors. Operatively, remote sensors are installed across a spatially distributed network; they gather information along a number of attribute dimensions and periodically feed a server with the measured data. The server is required to monitor these data, issue possible alarms, compute fast aggregates or forecast next data. As data analysis requests, which are submitted to a server, may concern both present, past, as well as future data, the server is forced to store and analyze the entire spatio-temporal datasets. However, in the case of massive amounts of collected data, there is the need to reduce the amount of data stored on the disk. One solution to address the storage limits is to compute spatiotemporal summaries of the data and use these summaries for subsequent data analytics (e.g. interpolation, outiler detection and classification, forecasting). This talk introduces the multiple possibilities and facets of various spatio-temporal patterns applied, in order to summarize, interpolate, identify anomalies and/or forecast spatio-temporal data collected via a sensor network.

Bio: Annalisa Appice is an Assistant Professor at the Department of Computer Science, University of Bari Aldo Moro, Italy. She received a Ph.D. in Computer Science with the thesis “Learning Relational Model Trees”. She was visiting researcher at the University of Bristol (U.K.) and at the Jozef Stefan Institute (Slovenia). Her current research interests include knowledge discovery and data mining, stream data mining, spatio-temporal data mining, netwok data analysis and process mining. In the last years, she has published about 110 papers (indexed on SCOPUS) in refereed journals and conferences. She participates in the European project FP7612944 MAESTRA. She was Program Chair of ECML-PKDD 2015 and she will be Program Co-Chair at ISMIS 2017. She is Local Arrangement Co-Chair of ALT/DS 2016. She is serving/has served in the program committee of several international/national conferences, including: IEEE ICDM, KDD, ICML, ECMLPKDD, ISMIS, DEXA, ACM SAC. She is member of the Editorial Board of both the Journal of Intelligent Information Systems and the IOS Press Journal of Intelligenza Artificiale. She is in the Review Board of the Journal of Applied Intelligence. She is in the Steering Committee of ECML-PKDD. She participated/participates in the organization (as Co-Chair) of six international workshops.

• Homepage: Annalisa Appice

Panče Panov

Title: Ontology of Data Mining

Abstract: In this talk, I will present our work on an ontology for representing entities from the domain of data mining (OntoDM). The OntoDM ontology defines the most essential data mining entities in a three-layered ontological structure comprising of a specification, an implementation and an application layer. It provides a representational framework for the description of mining structured data, and in addition provides taxonomies of datasets, data mining tasks, generalizations, data mining algorithms and constraints, based on the type of data. The ontology comprises of three modular sub-ontologies, covering different aspects of data mining and knowledge discovery process. These include an ontology of core data mining entities (OntoDMcore), an ontology of datatypes (OntoDT), and ontology for representing the knowledge discovery process (OntoDM-KDD). OntoDM is designed to support a wide range of applications/use cases, such as semantic annotation of data mining algorithms, datasets and results; annotation of QSAR studies in the context of drug discovery investigations and others. The ontology has been thoroughly assessed following best practices in ontology engineering, is fully interoperable with various domain resources and is easy to extend. Finally, the ontology is freely available at http://www.ontodm.com.

Bio: Panče Panov is a postdoctoral researcher at the Department of Knowledge Technologies, Jožef Stefan Institute, Ljubljana, Slovenia. He completed his PhD in 2012 in the area of data mining at the Jožef Stefan International Postgraduate School, Ljubljana, Slovenia. His thesis concerned the design and implementation of a modular ontology for the domain of data mining. His research interests are related to machine learning, data mining, the knowledge discovery process, and applying ontology in these domains. His contributions include developments of ontologies for describing the domain of data mining and the process of knowledge discovery, which can be employed in various applications. He was actively involved in several EU-funded projects in the past (IQ, SUMO) and is currently involved in the MAESTRA project. In addition, he participated in several projects financed by the Slovenian research agency and one bilateral project between Slovenia and Croatia. He is a co-editor of the book entitled “Inductive databases and constraint-based data mining” published in 2010 by Springer. In 2014, he was program co-chair of the International Conference on Discovery Science (2014) and co-editor of the proceedings of the conference published by Springer. Finally, in 2016 he a co-editor of a special issue of the Journal Machine Learning on Discovery Science.

• Homepage: Panče Panov

Saso Dzeroski

Title:

Abstract:

Bio: In the last ten years, I built an informal research group, which at present includes 17 researchers. We develop methods for constraint-based data mining, predicting structured outputs, and automated modelling of dynamics systems and apply them to problems from systems biology and ecology.I coordinated the project IQ, which produced a general framework for data mining, leading to representational ontologies for datatypes and data mining) and methodological advances (methods for predicting structured outputs based on constrained predictive clustering). I participated in two EU-funded projects in the area of systems biology (E.E.T. Pipeline, PHAGOSYS). I am currently coordinating the project MAESTRA, which is developing (among other) methods for predicting structured outputs. I have also proposed the process-based approach to automated modelling of dynamic systems and edited an overview on the area where the approach belongs (computational scientific discovery).

• Homepage: Saso Dzeroski

João Gama

Title:

Abstract:

Bio: I'm a researcher at LIAAD - INESC Porto, the Laboratory of Artificial Intelligence and Decision Support of the University of Porto. Scientific interests are: Machine Learning Learning from Data Streams Ensembles of Classifiers Constructive Induction Probabilistic Reasoning FCT Public key: J00878283U64

• Homepage: João Gama

Dragi Kocev

Title: Semi-supervised tree learning

Abstract: In many real-life problems, obtaining labelled data can be a very expensive and laborious task, while unlabeled data can be abundant. The availability of labeled data can seriously limit the performance of supervised learning methods. In this talk, we will present a semi-supervised classification tree induction algorithm that can exploit both the labelled and unlabeled data, while preserving all of the appealing characteristics of standard supervised decision trees: being non-parametric, efficient, having good predictive performance and producing readily interpretable models. We also discuss the results of an empirical evaluation of the proposed method: (1) It is able to outperform supervised trees; (2) It is the most effective in scenarios especially relevant to SSL - when few labeled examples are available; (3) It very seldom degrades the performance of supervised trees (and even then only slightly), due to its mechanism for controlling the influence of unlabeled examples; (4) It only requires a limited number of unlabeled examples to reach the potential improvement in predictive performance over supervised trees. These properties of the proposed method make it the first semi-supervised method that is highly useful in practice.

Bio: Dragi Kocev is a researcher at the Department of Knowledge Technologies, JSI. He completed his PhD in 2011 at the Jozef Stefan International Postgraduate School in Ljubljana on the topic of learning ensemble models for predicting structured outputs. He was a visiting research fellow at the University of Bari, Italy in 2014/2015. His research interests are in the field of data mining and includes the study, development and application of data mining algorithms. His current research is aimed towards further development of efficient methods for learning from data with structured outputs (e.g., predicting multiple targets, hierarchical multi-label classification…) and their applications in machine vision, life sciences and ecological modelling. He has participated in several national Slovenian projects, the EU funded projects IQ and PHAGOSYS and is involved in the Human Brain Project. He is co-coordinator of the FP7 FET Open project MAESTRA.

• Homepage: Dragi Kocev

Michelangelo Ceci

Title: Mining network data

Abstract: Networks have become ubiquitous in several social, economical and scientific fields, ranging from the Internet to social sciences, biology, epidemiology, geography, communication systems, finance and many others. For this reason, in recent years several data mining approaches, specifically designed for tackling predictive and descriptive tasks for network-structured data, have been proposed. The main challenges they have to face with are: i) the inherent dependency of the connected node which introduces some form of autocorrelation and ii) the possible heterogeneity of nodes and edges. In this talk we intend to introduce the various forms of autocorrelation in networked data and to present the challenges that they pose to traditional data mining algorithms. To this aim, we will discuss different approaches that allow us to take explicitly and implicitly network autocorrelation into account: a) New heuristics to adapt learned predictive models because of the explicit presence of autocorrelation. b) Collective learning and semi-supervised/transductive learning which implicitly take autocorrelation into account when leveraging on the smoothness assumption.

Title: Network Applications

Abstract: In this talk we will abstract important issues from a number of application domains with various types of network data, where the consideration of the network autocorrelation and/or the use of a reppresentation formalism which allow us to represent heterogeneity both of nodes and edges is of fundamental importance. Applications concern both predictive and descriptive tasks and focus on functional genomics, sensor networks, biological networks and social networks.

Bio: Michelangelo Ceci received a "laurea" degree from the University of Bari in 2001. In 2005 he received his Ph.D. degree in Computer Science from the same University. From 2005 to 2015 he was assistant professor at the Dept of Computer Science, University of Bari, Italy. Since 2015 he is an associate professor at the same department. His main research interests are in data mining and machine learning from complex and networked data with applications in bioinformatics, textual data analysis and spatial data analysis. He was a visiting researcher at the University of Bristol (U.K.) and at the JSI (SLO). He has published more than 140 papers in refereed journals and conferences. He is member of the editorial boards of: IJSNM, IJDSN, IJDATS and Intelligenza Artificiale. He is CoChair of DS 2016 and ECML-PKDD 2017. He has been the program co-chair of five workshops, the organizing committee chair of SEBD 2007 and member of the guest editorial board of the ECMLPKDD 2014 - 2016 journal tracks.

• Homepage: Michelangelo Ceci

Tomislav Šmuc

Title: Redescription mining: algorithms and applications

Abstract: Redescription mining is relatively new paradigm whose primary task is finding subsets of data that can be described with multiple descriptions. Redescription Mining techniques typically rely on decision tree or rule learning performed on concurrent attribute sets describing same samples. Multiple descriptions are of importance in studying datasets in biology or biomedicine, in which same entities are described by different descriptor sets, and where obtaining alternative descriptions might help generate new explanations or hypotheses with respect to the problem at hand (e.g. biomarkers and clinical data; state demographic patterns and state trading patterns). The lecture covers main techniques and specific tasks of RM, and illustrates application on several datasets.

Bio: Tomislav Šmuc is Senior Scientist at Rudjer Boskovic Institute, Zagreb, Croatia. He received his PhD degree in Electrical Engineering from University of Zagreb, Croatia in 1994. His research interests are in the areas machine learning and data mining in particular related to knowledge discovery applications in bioinformatics, drug-discovery, biomedicine and other domains. He has published more than 70 papers in journals and proceedings of international conferences and participated in a number of international projects with topics related to modelling complex systems, advanced data mining and knowledge discovery techniques.

• Homepage: Tomislav Šmuc

Ivica Dimitrovski

Title: Deep learning for plant identification

Abstract: Image-based approaches are nowadays considered to be one of the most promising solution to help bridging the botanical taxonomic gap. Most of these image-based approaches relies on computational methods to extract discriminative features from the images. Features have been traditionally hand-crafted or hand-engineered. However, a recent trend in machine learning has demonstrated that learned representations are more effective and efficient. The main advantage is that algorithms automatically analyse large collections of images and identify features that can classify images. In this talk a novel techniques and their related algorithms for the purpose of identifying plant species from visual data are presented. Convolutional Neural Networks (CNNs) are used to extract features and classify images at the same time. The proposed CNN is pre-trained in a large collection of generic web images and fine-tuned to recognize plants with the highest possible accuracy.

Bio: In 2000 he started his studies at the Faculty of Electrical Engineering and Information Technologies, University "Ss Cyrill and Methodius" in Skopje, Macedonia. He received his bachelor degree in computer science, automation and electrical engineering from the Faculty of Electrical Engineering and Information Technologies in 2005. In 2008 he defended his master thesis “Content-based medical image retrieval system” and in 2011 he defended his PhD thesis “Generic approach for content based image retrieval” at the Faculty of Electrical Engineering and Information Technologies. He worked at the Department of Computer Science and Engineering, at the Faculty of Electrical Engineering and Information Technologies, Ss. Cyril and Methodius University in Skopje, from June 2005 to August 2011. In June 2009 he gained the title assistant at the Department of Computer Science and Engineering at the Faculty of Electrical Engineering and Information Technologies, Skopje. He was a visiting researcher at the Jozef Stefan Institute, Slovenia from October 2009 to October 2010. Since September 2011 he is with the Faculty of Computer Science and Engineering in Skopje as assistant professor. He has authored many scientific papers in the field of computer science. He has participated in several and was a project leader in few domestic and international projects and was program member of several international conferences. His research interests are focused on the areas of computer vision, machine learning and software engineering.

• Homepage: Ivica Dimitrovski

Gjorgji Madjarov

Title: Decomposition and structuring of the output space in multi-label classification

Abstract: Instead of traditional (multi-class) learning approaches that assume label independency, multi-label learning approaches must deal with the existing label dependencies and relations. Many approaches try to model these dependencies in the process of learning and integrate them in the final predictive model, without making a clear difference between the learning process and the process of modeling the label dependencies. Also, the label relations incorporated in the learned model are not directly visible and can not be (re)used in conjunction with other learning approaches. In this talk, we will investigate the use of label's structures constructed in a data-driven manner and the influence of the decomposition of the output space in multi-label classification. The experimental results reveal that the use of the data-derived label's structures can significantly improve the performance of the predictive models in multi-label classification as compared to the use of a flat label set.

Bio: Gjorgji Madjarov is currently an Assistant Professor at the Faculty of Computer Science and Engineering, “St. Cyril and Methodius” University in Skopje, Republic of Macedonia. His research interests are on multi-class classification, multi-label classification and ranking, hierarchical multi-label prediction, structured output prediction, temporal-spatial prediction, time series prediction and analysis, and learning from multiple information sources.

His wider fields of interest also include: artificial intelligence, supervised learning, unsupervised learning, pattern recognition and computer vision.

• Homepage: Gjorgji Madjarov

Yoav Benjamini

Title: Controlling false discovery rates in multiple testing and selective inference and analysis of neuro data

Abstract: The problems of selective inference when addressing multiple tests and confidence intervals and the use of false discovery approaches to address them will be presented. Specific examples will be given relating brain phenotypes with genotype. New hierarchical approaches will be discussed enabling the analysis of complex research questions.

Bio: My scientific work combines theoretical research in statistical methodology with applied research that involves complex problems with massive data. The methodological work is on selective and simultaneous inference (multiple comparisons), and centers on the “False Discovery Rate” (FDR) criterion, as well as on general methods for data analysis, data mining and data visualization. My current interest is the replicability problem in science: too often, the results of studies gaining headlines cannot be replicated by other experimenters. Part of the problem is the use of statistical tools that fail to address the challenge of selective inference. I am trying to develop statistical tools that will aid researchers to cope with this problem, from the areas of Medicine, Epidemiology, Genomics, Bioinformatics, Neuroscience and behavior. In these areas and others, I enjoy cooperating with scientists on the design and analysis of their experiments. With Prof. Ilan Golani we share a lab where we investigate exploration behavior of animals when exposed a novel environment, be they flies or rodents, and even human babies. The computational and statistical tools being developed enable us to find structure in the huge data sets being compiled in such experiments. We advance the use of these tools to the development of drugs and to understand complex phenomena such as acquiring language. Understanding that Statistics is an extremely useful profession in our data flooded world, and more such professionals are needed everywhere, I teach both basics and advanced courses and workshops and supervise many research students. The statistical point of view calls for relying on data based evidence rather than feelings and beliefs; being committed to this point of view, I make an effort to offer the ability to draw appropriate conclusions from data to as wide an audience as possible. Member of the Sagol School of Neuroscience and the Edmond J. Safra Center for Bioinformatics, the Israel Statistical Association, American Statistical Association, The Institute of Mathematical Statistics, Eastern Mediterranean Region of the International Biometric Society, and The Israel Academy of Sciences and Humanities. Recipient of Israel Prize in Statistics and Economics.

• Homepage: Yoav Benjamini

Albert Bifet

Title: Architectures for distributed mining of big data

Abstract: In this talk, we will present the main architectures for building distributed systems to analyze big data. We will give a gentle introduction to new recent exciting data processing engines, presenting concepts, architectures and algorithms for data management, and analytics, at a very large scale. We will discuss in detail MapReduce, Apache Flink, Apache Spark and the new upcoming Apache Beam, previously known as Google Cloud Dataflow.

Bio: I'm a data scientist with 10+ years of international experience in research and in leading new open source software projects for business analytics, data mining and machine learning (Huawei, Yahoo, University of Waikato, UPC). I've worked in Hong Kong, New Zealand and Europe. I enjoy working on using Big Data stream mining and large-scale data analysis to gain insights from Big Data. I design and build applications on real-time data stream mining, at the intersection of data science, algorithms, and business. I have experience building and managing international open-source projects, providing leading strategy, and technical approaches, to deliver high quality software and research. I like delivering solutions that make business sense, matter at the individual level, and work at scale. I have authored more than 60 papers in books, journals and international conferences.

• Homepage: Albert Bifet

Ulf Brefeld

Title: Semi-supervised learning for SOP

Abstract: Learning mappings between arbitrary structured input and output variables is a fundamental problem in machine learning. It covers many natural learning tasks and challenges the standard model of learning a mapping from independently drawn instances to a small set of labels. Potential applications include classification with a class taxonomy, named entity recognition, and natural language parsing. In these structured domains, labeled training instances require expensive expert interaction while unlabeled inputs are readily available and inexpensive. The tutorial deals with semi-supervised learning of discriminative models for structured output variables. The analytical techniques and algorithms of classical semi-supervised learning are lifted to the structured setting. Several approaches based on different assumptions of the data are presented. Co- learning, for instance, maximizes the agreement among multiple hypotheses while transductive approaches rely on an implicit cluster assumption. I also report on empirical results showing that the semi-supervised algorithms lead to significantly lower error rates than their fully supervised counterparts in many application areas, including multi-class classification, named entity recognition, and natural language parsing.

Bio: I am a pro­fes­sor for Ma­chi­ne Learning at Leu­pha­na Uni­ver­sität Lüne­burg. Pri­or to joi­ning Leu­pha­na, I was a joint pro­fes­sor for Know­ledge Mi­ning & As­sess­ment at TU Darm­stadt and the Ger­man In­sti­tu­te for Edu­ca­tio­nal Re­se­arch (DIPF), Frank­furt am Main. Be­fo­re, I led the Re­com­men­der Sys­tems group at Za­lan­do SE and worked on ma­chi­ne learning at Uni­ver­sität Bonn, Yahoo! Re­se­arch Bar­ce­lo­na, Tech­ni­sche Uni­ver­sität Ber­lin, Max Planck In­sti­tu­te for Com­pu­ter Sci­ence in Saarbrücken, and at Hum­boldt-Uni­ver­sität zu Ber­lin. I re­cei­ved a Di­plo­ma in Com­pu­ter Sci­ence in 2003 from Tech­ni­sche Uni­ver­sität Ber­lin and a Ph.D. (Dr. rer. nat.) in 2008 from Hum­boldt-Uni­ver­sität zu Ber­lin. I am in­te­rested in sta­tis­ti­cal ma­chi­ne learning and data mi­ning.

• Homepage: Ulf Brefeldt

Ondřej Chum

Title: Large Scale Image Retrieval and Mining

Abstract: After introducing the classical setting of the problem, I will focus on retrieval methods based on the bag of words image representation that exploit geometric constrains. Novel formulations of image retrieval problem will be discussed, showing that the standard ranking of images based on similarity addresses only one of possible user requirements. Retrieval methods efficiently solving the new formulations by exploiting geometric constraints will be used in different scenarios. These include online browsing of image collections, image analysis based on large collections of photographs, image mining for 3D model construction, or hard example selection for CNN tuning.

Bio: Ondřej Chum received the MSc degree in computer science from Charles University, Prague, in 2001 and the PhD degree from the Czech Technical University in Prague, in 2005. From 2005 to 2006, he was a research Fellow at the Centre for Machine Perception, Czech Technical University. From 2006 to 2007 he was a post-doc at the Visual Geometry Group, University of Oxford, UK. Recently, he is now an associate professor back at the Centre for Machine Perception. His research interests include object recognition, large-scale image and particular-object retrieval, invariant feature detection, and RANSAC-type optimization. He has coorganized the “25 years of RANSAC” Workshop in conjunction with CVPR 2006, Computer Vision Winter Workshop 2006, and Vision and Sports Summer School (VS3) in Prague 2012 and 2014. He was the recipient of the runner up award for the “2012 Outstanding Young Researcher in Image & Vision Computing” by the Journal of Image and Vision Computing for researchers within seven years of their PhD, and the Best Paper Prize at the British Machine Vision Conference in 2002. In 2013, he was awarded ERC-CZ grant. His talk takes place on Wednesday, January 28 at 3pm in room E104.

• Homepage: Ondřej Chum

Hadi Fanaee-T

Title: Mining tensor data

Abstract: We begin by explaining basic concepts such as tensors and tensor decompositions and their importance in data mining. Later on a toy example of supermarket sales we explain how to generate tensor data from a regular database. Next, we briefly demonstrate how to perform various data mining tasks on tensor data such as tensor completion, correspondence analysis, co-clustering, outlier detection, event detection, anomaly detection, feature extraction and dimensionality reduction, link prediction, forecasting, recommendation and topic modelling. Afterwards, applications of tensor decomposition in various disciplines will be shown in an illustrative way. The talk will be closed by brief introduction of state-of-the-art methods, software, tools and bibliography.

Bio: Hadi Fanaee-T received his PhD degree (with distinction) in Computer Science from University of Porto, Portugal in 2015 and MSc degree from Shiraz University, Iran in 2010. Currently he is a Data Scientist at Laboratory of Artificial Intelligence and Decision Support, University of Porto, Portugal. His main research interests are data mining using tensor decomposition and application of tensor-based methods in real-world problems. He has served as Senior PC member for IJCAI2015, PC Member for ECML-PKDD 2013-15 and also reviewer for DAMI, MLJ, ITKD, KBS, CSUR and many more scientific venues.

• Homepage: Hadi Fanaee-T

Paolo Frasconi

Title: Kernel methods for structured data

Abstract: Structured (or relational) data representations naturally occurs in many domains including natural language processing, vision, and computational chemistry and biology. Kernels effectively decouple data representation aspects from statistical learning and once a kernel is defined for a certain data type, existing learning algorithms can be applied without modifications. In this lecture I will give an overview of kernels for graphs and other structured data types such as sequences and trees. After reviewing some classic graph kernels I will present some more recent techniques for handling graphs labeled by continuous attributes. I will also discuss how general relational learning problems can be approached using kernel methods. Finally, I will briefly introduce some closely related approaches based on recursive neural networks.

Bio: I'm a full professor of Computer Science at DINFO, University of Florence. I have been an Associate Professor at Dipartimento di Sistemi e Informatica from 2000 to 2010, and an Associate Professor at the University of Cagliari from 1998 to 2000. I've been a Visiting Professor at KU Leuven in 2010, a Lecturer at the University of Wollongong, Australia in 1998, and a Visiting Scholar at MIT in 1992. My research interests are in the area of machine learning, with particular emphasis on algorithms for structured and relational data, learning with logical representations, and applications to bioinformatics. In these areas I've written over 140 refereed papers. You may find my list of publications on my Scholar profile and on this page. Among other activities, I'm an Associate Editor of the Artificial Intelligence Journal and an Action Editor of the Machine Learning Journal. I co-chaired the Int. Conference on Prestigious Applications of Artificial Intelligence (PAIS) 2012, the AAAI 2010 Special Track on AI and Bioinformatics, the 20th International Conference on Inductive Logic Programming (2010), and the 5th International Workshop on Mining and Learning with Graphs (2007). Since 2015 I'm heading the regional PhD Program in Smart Computing of the Universities of Florence, Pisa and Siena.

• Homepage: Paolo Frasconi

Cesare Furlanello

Title: Metagenomics: predictive biomarkers and networks

Abstract: The study of microbiota communities from omics data is one of the most fast paced scientific domains: bacteria, fungi and possibly viruses are omni-present in the environment and they are increasingly found being involved in processes of human life. In this lecture, we will discuss how to seek patterns with machine learning and networks from meta-omics data , panels of clinical phenotypes, and bioinformatics tools. Without pretending to disentangle the core biomedical issue (if different meta-omics profiles are cause, effect, the result of a cascade of associations, or other), we will consider an adaptation of the FDA MAQC/SEQC Data Analysis Protocols for identifying predictive markers with machine learning methods embedded in complex bioinformatics frameworks. Methods for differential network analysis will also be introduced. Examples with kernel methods, random forest and deep learning will be presented, with applications on identifying predictive markers associated to diet (gut microbiome), clinic trajectories in Pediatric Dysbiosis, and to neuro development disorders (the Gut:brain axis in ASD). Joint work with Marco Chierici, A. Zandona, and G.Jurman.

Bio: Cesare Furlanello is a data scientist and expert in interdisciplinary applications of machine learning. He is senior researcher of Fondazione Bruno Kessler (FBK), where he leads the MPBA Research Unit (Predictive Models for Biomedicine and Environment) and coordinates the Complex Data Analytics Research Line, a department of about 50 scientists. He is active in the field of bioinformatics, developing methods and software solutions to find patterns in very high throughput omics molecular data (such as Next Generation Sequencing and microarrays), also a bioinformatics collaborator of RIKEN's FANTOM5 project and of the FDA in the MAQC and SEQC consortia. At MPBA, he has been a PI of a total of >80 funded projects, in which Predictive Models are applied to Biomedical and Environmental data. He is Adjoint Professor at The Wistar Institute in Philadelphia, since 2012, and a PI for the European Institute of Technology (EIT-Digital: 2015-2017), also collaborating with industry for predictive retail and new applications of personal sensors. In applications for functional genomics and metagenomics, MPBA is now focused to develop dashboards for predictive analytics methods and new applications of deep learning.

• Homepage: Cesare Furlanello

Ljupco Kocarev

Title: Complex Networks Analysis

Abstract: Networks have been studied for decades by diverse groups of scientists including social scientists, natural scientists, and most recently, computer scientists. On the other hand, graph theory, which is a well-developed branch of mathematics with applications in various disciplines (as well as to other branches of mathematics), has been studied by mathematicians for much longer. These groups of scholars, however, often do not see eye to eye. In this talk, I will present an effort to review methods for complex network analysis from different perspectives. More specifically, I will address the emerging field of learning multilayer networks.

Bio: Ljupco Kocarev is a full professor at the Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University in Skopje, Macedonia, director of the Research Center for Computer Science and Information Technologies at the Macedonian Academy of Sciences and Arts, and research professor at University of California San Diego (UCSD), USA. His scientific interests include networks, nonlinear systems and circuits; dynamical systems and mathematical modeling, and computational biology. He has co-authored more than 150 journal papers in 30 different international journals, ranging from mathematics to physics and from electrical engineering to computer sciences. According to Science Citation Index his work has been cited more than 6000 times, while according to Google scholar his work has been cited almost 12000 times. He is a fellow of IEEE, member of the Macedonian Academy of Sciences and Arts, and member of several other academies worldwide.

• Homepage: Ljupco Kocarev

Pedro Larranaga

Title: Bayesian networks for multi-dimensional classification and neuroscience applications

Abstract: In this talk, the problem of multi-dimensional classification (a generalization of multi-label classification) will be approached with Bayesian networks. Multi-dimensional Bayesian network classifiers will be introduced and, under this paradigm, we will see that the classification problem can be formulated as the finding of the most probable explanation. As this is an NP-complete problem, in order to alleviate its computational burden, a special type of multi-dimensional Bayesian network classifiers called class-bridge decomposable will be introduced. Methods for learning this type of models from data, both based on score+search and on estimating the Markov blanket of each class variable, will be presented. Some variants of the general framework that consider the use of chain classifiers, or super-classes or even tratable (from the point of view of inference) models will be explained. How this kind of models can be updated in data stream settings will be also discussed. Finally, the application of the methodology in two real word problems in neuroscience: (a) quality of life in Parkinson disease patients, and (b) classification of neuron types based on their morphological characteristics, will be presented.

Bio: Pedro Larranaga is Full Professor in Computer Science and Artificial Intelligence at the Technical University of Madrid (UPM) since 2007. He received the MSc degree in mathematics (statistics) from the University of Valladolid and the PhD degree in computer science from the University of the Basque Country (“excellence award”). Before moving to UPM, his academic career has been developed at the University of the Basque Country (UPV-EHU) at several faculty ranks: Assistant Professor (1985-1998), Associate Professor (1998-2004) and Full Professor (2004-2007). He earned the habilitation qualification for Full Professor in 2003. His research interests are primarily in the areas of probabilistic graphical models, metaheuristics for optimization, data mining, classification models, and real applications, like biomedicine, bioinformatics and neuroscience. He has published more than 150 papers in impact factor journals and has supervised more than 20 PhD theses. He is ECCAI felow since 2012 and he has been awared the 2013 Spanish National Prize in Computer Science.

• Homepage: Pedro Larranaga

Julien Mairal

Title: Sparse estimation and dictionary learning

Abstract: In recent years, a large amount of multi-disciplinary research has been conducted on sparse models and their applications. In statistics and machine learning, the sparsity principle is used to perform model selection---that is, automatically selecting a simple model among a large collection of them. In signal processing, sparse coding consists of representing data with linear combinations of a few dictionary elements. Subsequently, the corresponding tools have been widely adopted by several scientific communities such as image processing, neuroscience, bioinformatics, or computer vision. The goal of this course is to offer a short introduction to sparse modeling. More specifically, we focus on applications where the dictionary is learned and adapted to data, yielding a compact representation that has been successful in various contexts.

Title: Towards deep kernel machines

Abstract: In the past years, deep neural networks such as convolutional or recurrent ones have become highly popular for solving various prediction problems, notably in computer vision and natural language processing. Conceptually close to approaches that were developed several decades ago, they greatly benefit from the large amounts of labeled data that have been available recently, allowing to learn huge numbers of model parameters without worrying too much about overfitting. Before the resurgence of neural networks, non-parametric models based on positive definite kernels were one of the most dominant topics in machine learning. These approaches are still widely used today because of several attractive features. Kernel methods are indeed versatile; as long as a positive definite kernel is specified for the type of data considered—e.g., vectors, sequences, graphs, or sets—a large class of machine learning algorithms originally defined for linear models may be used. Kernel methods also admit natural mechanisms to control the learning capacity and reduce overfitting. In this lecture, we will consider recent directions that seek to achieve the best of both worlds in the context of image modeling. We will introduce novel tools called convolutional kernel networks, and show promising applications in image retrieval, classification, and super-resolution.

Bio: Julien Mairal is a research scientist at Inria. He received a graduate degree from Ecole Polytechnique, France, in 2005, and a PhD degree from the Ecole Normale Supérieure, Cachan, France, in 2010. Then, he went to the statistics department of UC Berkeley as a post-doctoral researcher, before joining Inria in 2012. His research interests include machine learning, computer vision, and statistical image and signal processing. In 2013, he received the Cor Baayen prize, awarded every year by ERCIM to a promising young researcher in computer science and applied mathematics.

• Homepage: Julien Mairal

Donato Malerba

Title: Spatio-temporal Data Mining: Part 1

Abstract: The emergence of new technologies for collecting huge volumes of spatio-temporal data in various application domains (e.g. ecology and environmental management, public safety, transportation, epidemiology, earth science, climatology and precision agriculture) emphasizes the need of spatio-temporal data mining for the automated discovery of spatiotemporal knowledge. Spatio-temporal data mining is a sub-field of data mining that focuses on developing techniques to analyze both spatial and temporal data properties. The former are represented by the location and geometry attributes. The latter are represented by timestamp or time interval for which data are valid. Spatio-temporal data mining poses several challenges. First, spatio-temporal datasets are embedded in a continuous space and time, therefore classical data mining tasks assuming discrete data may not be effective. Second, the complexity and heterogeneity of spatio-temporal data types complicates the extraction process. Third, the implicit relationships and the non stationary violates basic independence assumptions and constant means and variance. Finally, the nature of data generally requires that the development of data mining techniques accounts for the underlying physics correctly. The main goal of this talk is to disseminate this research field. It starts with the characteristics of the data inputs of spatio-temporal data mining and overview its statistical foundation. It then gives an overview of the current state of the art and of the various methodologies and algorithms for spatial and spatio-temporal data mining.

Bio: Research Interests:Data Mining and Knowledge Discovery, Inductive logic programming and Multi-relational Data Mining, Symbolic Data Analysis, Spatial Data Mining, Document Image Analysis and Recognition, Web Mining, Bioinformatics, Network Traffic Analysis.

• Homepage: Donato Malerba

Jesse Read

Title: Multi-label learning from batch and streaming data

Abstract: In multi-labeled data, each instance may be simultaneously associated with multiple class labels, as opposed to a single class. This context arises naturally in many domains, for example: document categorization, labeling images and audio, and bioinformatics. The academic literature on multi-label learning has proliferated in recent years. The main challenges are to model and deal with label dependence and develop scalable methods for large datasets. Further challenges arise in the context of data streams, where examples may arrive rapidly and continuously, labels may be sparse, and the dependence relations among labels may drift over time. This talk reviews some important developments in multi-label classification, the connection to related areas such as structured-output prediction, and addresses the particular issue of adapting multi-label learning to deal with data streams.

Bio: Jesse Read obtained his PhD in Computer Science from the University of Waikato, New Zealand, in 2010. He has carried out postdoc research in the Carlos III University of Madrid in Spain, and Aalto University in Finland. He is currently in the INFRES group at Télécom ParisTech in France. His main research interests include multi-label classification, and learning from sequential data and evolving data streams. He has been involved with diverse real-world applications and testbeds including route and destination prediction of urban travellers, tracking with low-power sensor motes, and modelling intra-annual growth of tree growth of pine trees.

• Homepage: Jesse Read

Myra Spiliopoulou

Title:

Abstract:

Bio: Prof. Myra Spiliopoulou Faculty of Computer Science, Otto-von-Guericke-University Magdeburg Academic title: Prof. Dr. rer.nat/Griechenland habil. Place and date of birth: Athens, Greece, 1965 02/2003 - now Professor of Business Information Systems, Otto-von-Guericke-University Magdeburg 2001 - 2003 Professor and Department Chair of E-Business, Leipzig Graduate School of Management (Handelshochschule Leipzig), Leipzig, Germany 2000 - 2001 Dorothea-von-Erxleben professor, Otto-von-Guericke-University Magdeburg 2000 Habilitation with venia legendi for “Wirtschaftsinformatik” at the HU Berlin 1994 - 2000 Postdoc researcher – from June 1995 at a C1 position – at the Institute of Information Systems, Faculty of Economics, Humboldt-University Berlin 1992 PhD on parallel query optimization, University of Athens, Greece 1988 - 1994 PhD student and research assistant at the Dept. of Informatics, University of Athens, Greece 1986 - 1989 Work as software engineer in Athens, Greece 1982 - 1986 Study of Mathematics, University of Athens, Greece

• Homepage: Myra Spiliopoulou

Fran Supek

Title: Gene function prediction and relating genotypes with phenotypes

Abstract: Extensive complementarity between gene function prediction methods: The number of sequenced genomes rises steadily, but we still lack the knowledge about the biological roles of many genes. Automated function prediction (AFP) is thus a necessity. We hypothesize that AFP approaches which draw on distinct genome features may be useful for predicting different types of gene functions, motivating a systematic analysis of the benefits gained by obtaining and integrating such predictions. Our pipeline amalgamates 5,133,543 genes from 2,071 genomes in a single massive analysis that evaluates five established genomic AFP methodologies. While 1,227 Gene Ontology terms yielded reliable predictions, the majority of these functions were accessible to only one or two of the methods. Moreover, different methods tend to assign a GO term to non-overlapping sets of genes. Thus, inferences made by diverse AFP methods display a striking complementary, both gene-wise and function-wise. Because of this, a viable integration strategy is to rely on a single most-confident prediction per gene/function, instead of enforcing agreement across multiple AFP methods. Availability of more sequenced genomes boosts the predictive accuracy of AFP approaches and also the benefit from integrating them.

The landscape of microbial phenotypic traits and their genetic basis: Bacteria and Archaea display a variety of phenotypic traits and can adapt to diverse ecological niches. However, systematic annotation of prokaryotic phenomes is lacking. We have developed the ProTraits computational pipeline that annotates microbes with phenotypes by text-mining the scientific literature and the broader World Wide Web, while also being able to define novel concepts from unstructured text. Moreover, ProTraits draws extensively on comparative genomics, capturing patterns in gene repertoires, codon usage biases, proteome composition and co-occurrence in metagenomes. Notably, we find that gene synteny is highly predictive of many phenotypes, and highlight examples of gene neighborhoods associated to spore-forming ability. Our inferences have yielded a compendium of ~545,000 novel phenotypic annotations, spanning 424 traits assigned to 3,046 bacterial and archaeal species. A global analysis of trait interrelatedness reveals a lower-dimensional structure that underlies the microbial phenotype network, suggesting common genetic underpinnings. Our extended set of phenotypic annotations allows detection of >50,000 high confidence gene-phenotype links, as well as extremely prevalent epistatic relationships that have shaped microbial gene repertoires.

Bio: Fran is currently a postdoctoral researcher at the EMBL/CRG Systems Biology Unit, Centre for Genomic Regulation (CRG), Barcelona. In addition, he holds a research associate position at the Division of Electronics, Rudjer Boskovic Institute (RBI), Zagreb. He has obtained his PhD degree in cellular and molecular biology from the University of Zagreb in 2010. His work was supported by fellowships from the Croatian Ministry of Science (for early-stage researchers), from the EU Marie Curie program (INTERPOD interdisciplinary postdocs) and he has also received a Ramon y Cajal fellowship of the Spanish government. Fran has contributed 26 research articles, of which 12 as first or last author, including publications in Cell, Nature, Genome Biology, PLOS Genetics and upcoming papers in Bioinformatics and Nature Genetics. His research interests revolve around characterizing the mutational processes that generate genetic diversity, while learning how to distinguish the subtle signatures of natural selection overlaid upon the resulting DNA sequence variability. Fran has investigated these issues by large-scale comparative genomics analyses of bacterial and archaeal genomes on the one hand, and of human cancer genomes on the other hand. Fran's future work will aim to further elucidate the mechanisms behind various mutational processes that shape the human genome and their consequences on carcinogenesis. More broadly, his interests involve applying cutting-edge machine learning methodologies to elucidate the genetic basis of phenotypes observed across the tree of life, from microbes to man.

• Homepage: Fran Supek

Ljupco Todoroski

Title: Relevance Network Algorithms for Network Reconstruction

Abstract: The lecture elaborates on the task of inferring interactions between network nodes from data on network (nodes) behavior through time. The task is omnipresent in many disciplines, most notably system biology, where one aims at discovering interactions between genes from time-series data. The particular focus of the lecture is on a general family of algorithms for network reconstruction, referred to as relevance network algorithms. To infer network interactions, a relevance network algorithm combines a measure of similarity between the time series observed in different nodes with a scoring scheme that refine the output of the similarity measure. Different similarity measures and scoring schemes have been proposed in literature; the lecture provides an overview of them and systematic empirical comparison of their performance in terms of the ability to reconstruct known network interactions from time series data. The performance measures are based on summary statistics combining the specificity and sensitivity of the reconstruction. Benchmarks include measurements of the yeast genetic network and artificial data from the well-known DREAM5 network inference challenge.

Bio: 2005- Full Professor of Informatics (as of 2015). University of Ljubljana, Faculty of Administration, Ljubljana, Slovenia. 1999-2007 PhD Student and Research Scientist. Jozef Stefan Institute, Department of Intelligent Systems, Ljubljana, Slovenia. 2004-2005 Postdoctoral Fellow. Stanford University, Center for the Study of Language and Information, Computational Learning Laboratory, Stanford, US. 1996-1999 Research Assistant. University of Ljubljana, Institute for Biomedical Informatics, Ljubljana, Slovenia. Education 2003 PhD in Computer and Information Science. University of Ljubljana, Faculty of Computer and Information Science. Thesis: Using domain knowledge for automated modeling of dynamic systems with equation discovery. 1998 MSc in Computer Science. University of Ljubljana, Faculty of Computer and Information Science. Thesis: Declarative bias in equation discovery. 1993 BSc in Computer Science. University of Ljubljana, Faculty of Electrical Engineering and Computer Science. Thesis: Automated modelling of dynamical systems with machine discovery. Visiting Positions October 2011 - February 2012 Visiting Professor. Stanford University, Center for Biomedical Informatics Research. July 2004 - January 2005, March-May 2007 Postdoctoral Fellow. Stanford University, Center for the Study of Language and Information. October-November 2003 Postdoctoral Fellow. Osaka University, Institute for Scientific and Industrial Research. June-August 2000 Visiting Researcher. University of Porto, Artificial Intelligence and Computer Science Laboratory. March 2000 University of Bristol, Department of Computer Science.

• Homepage: Ljupco Todoroski