Big data ethics | EPFL Graph Search

Big data ethics also known as simply data ethics refers to systemizing, defending, and recommending concepts of right and wrong conduct in relation to data, in particular personal data. Since the dawn of the Internet the sheer quantity and quality of data has dramatically increased and is continuing to do so exponentially. Big data describes this large amount of data that is so voluminous and complex that traditional data processing application software is inadequate to deal with them. Recent innovations in medical research and healthcare, such as high-throughput genome sequencing, high-resolution imaging, electronic medical patient records and a plethora of internet-connected health devices have triggered a data deluge that will reach the exabyte range in the near future. Data ethics is of increasing relevance as the quantity of data increases because of the scale of the impact. Big data ethics are different from information ethics because the focus of information ethics is more co

Automatic L2 Regularization for Multiple Generalized Additive Models

Yousra El Bachir

Multiple generalized additive models are a class of statistical regression models wherein parameters of probability distributions incorporate information through additive smooth functions of predictors. The functions are represented by basis function expansions, whose coefficients are the regression parameters. The smoothness is induced by a quadratic roughness penalty on the functionsâ curvature, which is equivalent to a weighted

L_2

regularization controlled by smoothing parameters. Regression fitting relies on maximum penalized likelihood estimation for the regression coefficients, and smoothness selection relies on maximum marginal likelihood estimation for the smoothing parameters. Owing to their nonlinearity, flexibility and interpretability, generalized additive models are widely used in statistical modeling, but despite recent advances, reliable and fast methods for automatic smoothing in massive datasets are unavailable. Existing approaches are either reliable, complex and slow, or unreliable, simpler and fast, so a compromise must be made. A bridge between these categories is needed to extend use of multiple generalized additive models to settings beyond those possible in existing software. This thesis is one step in this direction. We adopt the marginal likelihood approach to develop approximate expectation-maximization methods for automatic smoothing, which avoid evaluation of expensive and unstable terms. This results in simpler algorithms that do not sacrifice reliability and achieve state-of-the-art accuracy and computational efficiency. We extend the proposed approach to big-data settings and produce the first reliable, high-performance and distributed-memory algorithm for fitting massive multiple generalized additive models. Furthermore, we develop the underlying generic software libraries and make them accessible to the open-source community.

EPFL2019

Designing peer-to-peer overlays

Sarunas Girdzijauskas

The Small-World phenomenon, well known under the phrase "six degrees of separation", has been for a long time under the spotlight of investigation. The fact that our social network is closely-knitted and that any two people are linked by a short chain of acquaintances was confirmed by the experimental psychologist Stanley Milgram in the sixties. However, it was only after the seminal work of Jon Kleinberg in 2000 that it was understood not only why such networks exist, but also why it is possible to efficiently navigate in these networks. This proved to be a highly relevant discovery for peer-to-peer systems, since they share many fundamental similarities with the social networks; in particular the fact that the peer-to-peer routing solely relies on local decisions, without the possibility to invoke global knowledge. In this thesis we show how peer-to-peer system designs that are inspired by Small-World principles can address and solve many important problems, such as balancing the peer load, reducing high maintenance cost, or efficiently disseminating data in large-scale systems. We present three peer-to-peer approaches, namely Oscar, Gravity, and Fuzzynet, whose concepts stem from the design of navigable Small-World networks. Firstly, we introduce a novel theoretical model for building peer-to-peer systems which supports skewed node distributions and still preserves all desired properties of Kleinberg's Small-World networks. With such a model we set a reference base for the design of data-oriented peer-to-peer systems which are characterized by non-uniform distribution of keys as well as skewed query or access patterns. Based on this theoretical model we introduce Oscar, an overlay which uses a novel scalable network sampling technique for network construction, for which we provide a rigorous theoretical analysis. The simulations of our system validate the developed theory and evaluate Oscar's performance under typical conditions encountered in real-life large-scale networked systems, including participant heterogeneity, faults, as well as skewed and dynamic load-distributions. Furthermore, we show how by utilizing Small-World properties it is possible to reduce the maintenance cost of most structured overlays by discarding a core network connectivity element – the ring invariant. We argue that reliance on the ring structure is a serious impediment for real life deployment and scalability of structured overlays. We propose an overlay called Fuzzynet, which does not rely on the ring invariant, yet has all the functionalities of structured overlays. Fuzzynet takes the idea of lazy overlay maintenance further by eliminating the need for any explicit connectivity and data maintenance operations, relying merely on the actions performed when new Fuzzynet peers join the network. We show that with a sufficient amount of neighbors, even under high churn, data can be retrieved in Fuzzynet with high probability. Finally, we show how peer-to-peer systems based on the Small-World design and with the capability of supporting non-uniform key distributions can be successfully employed for large-scale data dissemination tasks. We introduce Gravity, a publish/subscribe system capable of building efficient dissemination structures, inducing only minimal dissemination relay overhead. This is achieved through Gravity's property to permit non-uniform peer key distributions which allows the subscribers to be clustered close to each other in the key space where data dissemination is cheap. An extensive experimental study confirms the effectiveness of our system under realistic subscription patterns and shows that Gravity surpasses existing approaches in efficiency by a large margin. With the peer-to-peer systems presented in this thesis we fill an important gap in the family of structured overlays, bringing into life practical systems, which can play a crucial role in enabling data-oriented applications distributed over wide-area networks.

EPFL2009