Finding communities in networks is crucial to understand their structure and function, as well as to identify the role of the nodes and uncover hidden relationships between nodes. In this talk I will briefly introduce the problem and then focus on algorithms based on optimization. I will discuss the limits of global optimization approaches and the potential advantages of local techniques. I will assess the delicate issue of testing the performance of methods. Finally I will be presenting evidence of a discrepancy between annotated and structural clusters, that may enforce major changes in current community finding algorithms and more generally in the basic notion of community.

User generated text is an element of Big Data that corresponds to all five V problems: volume, velocity, variety, veracity, and value. Natural Language Processing involves a collection of tools for processing text, transforming unstructured text into structured items to which Big Data methods of statistical processing can be applied. In this talk, we will give an overview of the size of each V problem for text, of the availability and shortcomings of tools for attacking the problem, and of the basic approaches being used to extract structure from text. Finally, we will describe how all the textual problems will soon be exacerbated by wearable computers capturing speech and written text which significantly augment the amount of text data to be processed, creating personal Big Data and bring about the need for personal semantics.

Large-scale classification of textual data into a large number of target classes (as the ones encoutered in Directory Mozilla or the Wikipedia category system) has been the focus of several studies, from researchers and developers in industry and academia alike. Several approaches have been proposed for deploying classifiers in such large-scale taxonomies, from flat to fully hierarchical ones. In this presentation, we will first introduce a theoretical analysis of the problem that exhibits the different parameters controlling the accuracy of each appraoch, and then propose several methods for enhanced classification in web-scale taxonomies. We will finally study the distribution of classes in such taxonomies, and in doing so show the importance of "small" classes, and will develop an approach for improving the classification on these small classes.

Events in an online social network can be categorized roughly into endogenous events, where users just respond to the actions of their neighbors within the network, or exogenous events, where users take actions due to drives external to the network. How much external drive should be provided to each user, such that the network activity can be steered towards a target state? In this paper, we model social events using multivariate Hawkes processes, which can capture both endogenous and exogenous event intensities, and derive a time dependent linear relation between the intensity of exogenous events and the overall network activity. Exploiting this connection, we develop a convex optimization framework for determining the required level of external drive in order for the network to reach a desired activity level. We experimented with event data gathered from Twitter, and show that our method can steer the activity of the network more accurately than alternatives.

Graphs are an abstraction widely adopted to represent networks of data or users. As a consequence, many machine learning approaches have been developed to deal with prediction tasks on graphs. In this talk, however, situations will be exhibited where standard graphs seem to be an inadequate representation. New tools that overcome these shortcomings will be introduced to model networked information. Associated machine learning algorithms will be presented along with their relationships to graph-based learning.

The goal of data science is the design of automated methods to analyze massive and complex data in order to extract useful information. The emergence of this unique research domain has certainly an element of hype (cf. big data). Nevertheless, it reflects on the profound changes triggered by the appearance of large and complex data in science, and it is also partially motivated by the undeniable convergence of traditionally distinct communities in statistics, machine learning, and signal processing.

In this talk we will survey some of the latest developments in data science research through the optics of potential applications in high-energy physics. We will then briefly describe three ongoing projects, mainly focusing on the HiggsML data challenge we organized with the ATLAS collaboration for optimizing the discovery significance in particle search.

The explosion of interest in big data has come hand in hand with new distributed system architectures facilitating the management of big data and the massively-parallel execution of complex queries and analytics tasks. In this talk I will present a thesis arguing for a much needed new research direction, based on three recent research results from my group, relating to temporal analytics queries, rank and join queries, and missing value imputations.

There is an increasingly pressing need, by several applications in diverse domains, for developing techniques able to index and mine very large collections of sequences, or data series. Examples of such applications come from biology, astronomy, entomology, the web, and other domains. It is not unusual for these applications to involve numbers of data series in the order of hundreds of millions to billions, which are often times not analyzed in their full detail due to their sheer size.

In this talk, we describe recent efforts in designing techniques for indexing and mining truly massive collections of data series that will enable scientists to easily analyze their data. We show that the main bottleneck in mining such massive datasets is the time taken to build the index, and we thus introduce solutions to this problem. Furthermore, we discuss novel techniques that adaptively create data series indexes, allowing users to correctly answer queries before the indexing task is finished. We also show how our methods allow mining on datasets that would otherwise be completely untenable, including the first published experiments using one billion data series.

Finally, we present our vision for the future in big sequence management research.

One hundred hours of video are uploaded to YouTube each minute. Over one hundred thousand hours of video are watched in the same time. To help viewers discover the content that will interest them, YouTube organizes its videos and channels by what they are about. The size of the YouTube corpus makes this a challenge. In this talk, I will describe the work we do in YouTube's Paris team to understand and classify our huge repository of videos, what we're able to do at scale, and what remains to be done.

Managing big data and running supercomputing jobs used to be for only well-funded research organizations and large corporations, but not any longer. AWS has democratized supercomputing and big data for the masses! AWS can provide you with the 64th fastest supercomputer in the world, on-demand and pay as you go. Hear from Stephan Hadinger, Sr Manager Solutions Architecture, to learn how our customers are using big data and high performance computing to change the world. Not only is AWS technology available to everyone, but it is self-service and cheaper than ever before, featuring innovative technology and flexible pricing models – our AWS cloud computing platform has disrupted big data and HPC. Learn from customer successes, as Stephan shares real-world case studies describing the specific big data and high performance computing challenges being solved on AWS. We will also share all the information you need to get you started quickly.

This session is recommended for anyone wanting to get started with big data in the AWS Cloud or want to get an update on AWS's recent big data/HPC innovations.

The Insurance Industry has always been a data-driven Industry; We all use “small” data (i.e a rather limited number of data points) to assess risks, to understand or anticipate customer needs, to improve customer satisfaction and retention. Big Data simply describes the fact that applying algorithms to increasingly large amounts of data can solve concrete business problems that had suboptimal solutions with smaller data sets.

Many features, signals, patterns can indeed only be observed by collecting and processing massive amounts of data – this was difficult, or even impossible a couple of years ago but the advent of technologies and scientific progress will progressively make it possible. Challenges are obviously on building capabilities and expertise but probably even more in the cultural shift – this is not an evolution but a true revolution in the way we approach business problems – innovation spirit, “exploration”, test and learn mode, acceptance of failure - and I believe that all leaders in the Industry will have to put Big Data at the very top of their agenda and align their priorities and organization accordingly. In that respect, I believe that the upcoming revolution is a huge challenge but also a once-in-a-life opportunity for those who will be able to become truly Customer-centric Data-Driven Companies.