TOP

PhD Accepted Papers

More details on PhD session website.

Accepted for Oral presentation

  • Be In The Know: Connecting News Articles to Relevant Twitter Conversations (Bichen Shi, Georgiana Ifrim, Neil Hurley)

    Authors

    Bichen Shi, Insight Centre, University College Dublin
    Georgiana Ifrim, University College Dublin, Ireland
    Neil Hurley, Insight Centre, University College Dublin

    Abstract

    In this paper we propose a framework for tracking and automatically connecting news articles to Twitter conversations as captured by Twitter hash- tags. For example, such a system could alert journalists about news that get a lot of Twitter reaction, so they can investigate those conversations for new developments in the story, promote their article to a set of interested consumers, or discover general sentiment towards the story. Mapping articles to hashtags is nevertheless challenging, due to different language style of articles versus tweets, the streaming aspect, and user behavior when marking tweet-terms as hashtags. We track the IrishTimes RSS-feed and a focused Twitter stream over a two months period, and present a system that assigns hashtags to each article, based on its Twitter echo. We propose a machine learning approach for classifying article- hashtag pairs. Our empirical study shows that our system delivers high precision for this task.

  • Recommender-based Multiple Classifier System (Yury Kashnitsky, Dmitry Ignatov, Sergei Kuznetsov)

    Authors

    Yury Kashnitsky, Higher School of Economics
    Dmitry Ignatov, Higher School of Economics
    Sergei Kuznetsov, Higher School of Economics

    Abstract

    The paper makes a brief introduction into multiple classifier systems and describes a particular algorithm which improves classification accuracy by making a recommendation of an algorithm to an object. This recommendation is done under a hypothesis that a classifier is likely to predict the label of the object correctly if it has correctly classified its neighbors. The process of assigning a classifier to each object involves here the apparatus of Formal Concept Analysis. We explain the principle of the algorithm on a toy example and describe experiments with real-world datasets.

  • Inference of Switched Biochemical Reaction Networks Using Sparse Bayesian Learning (Wei Pan, Ye Yuan, Aivar Sootla, Guy-Bart Stan)

    Authors

    Wei Pan, Imperial College London
    Ye Yuan, University of Cambridge
    Aivar Sootla, Imperial College London
    Guy-Bart Stan, Imperial College London

    Abstract

    This paper proposes an algorithm to identify biochemical reaction networks with time-varying kinetics. We formulate the problem as a nonconvex optimisation problem casted in a sparse Bayesian learning framework. The nonconvex problem can be efficiently solved using Convex-Concave programming. We test the effectiveness of the method on a simulated example of DNA circuit realising a switched chaotic Lorenz oscillator.

  • Clustering Boolean Tensors (Saskia Metzler, Pauli Miettinen)

    Authors

    Saskia Metzler, MPI for Informatics
    Pauli Miettinen, Max-Planck-Institut für Informatik

    Abstract

    Tensor factorizations are computationally hard problems, and in particular, often significantly harder than their matrix counterparts. In case of Boolean tensor factorizations — where the input tensor and all the factors are required to be binary and we use Boolean algebra — much of that hardness comes from the possibility of overlapping components. Yet, in many applications we are perfectly happy to partition at least one of the modes. In this paper we investigate what consequences does this partitioning have on the computational complexity of the Boolean tensor factorizations and present a new algorithm for the resulting clustering problem. While future work aims at further tuning our algorithm for Boolean tensor clustering, it already now can obtain better results than algorithms solving different relaxations of the problem.

  • Search for User-related Features in Matrix Factorization-based Recommender Systems (Marharyta Aleksandrova, Armelle Brun, Anne Boyer, Oleg Chertov)

    Authors

    Marharyta Aleksandrova, Université de Lorraine
    Armelle Brun, Université de lorraine – Loria
    Anne Boyer, Université de lorraine – Loria
    Oleg Chertov

    Abstract

    Matrix factorization (MF) is one of the most powerful approaches used in the frame of recommender systems. It aims to model the preferences of users about items through a reduced set of latent features. One main drawback of MF is the difficulty to interpret the automatically formed features. Following the intuition that the relation between users and items can be expressed through a reduced set of users, referred to as representative users, we propose a simple modification of a traditional MF algorithm, that forms a set of features corresponding to these representative users. On one state of the art dataset, we show that proposed representative users-based non-negative matrix factorization (RU-NMF) discovers interpretable features and does not significantly decrease the accuracy on test with 10 and 15 features.

  • Heterogeneous Bayes Filters with Sparse Bayesian Models: Application to state estimation in robotics (Alexandre Ravet, Simon Lacroix)

    Authors

    Alexandre Ravet, LAAS CNRS
    Simon Lacroix, LAAS-CNRS

    Abstract

    This study introduces a new augmented Bayes filter model for time-varying, context-dependent emission noise. The expected application, robust state estimation for a robot, motivates the use of the Relevance Vector Machine to model the emission noise, because it provides sparsity and fast inference capabilities. Besides the introduction of this new model, the paper also aims at comparing the final filter performance when discriminative training is used instead of generative training. The theoretical foundations for training and using inference over the model are proposed.

  • Optimistic Active Learning for Classification (Timothé Collet, Olivier Pietquin)

    Authors

    Timothé Collet, Supelec
    Olivier Pietquin, Université Lille1

    Abstract

    In this paper, we propose to reformulate the active learning problem occurring in classification as a sequential decision making problem. We particularly focus on the problem of dynamically allocating a fixed budget of samples. This raises the problem of the trade off between exploration and exploitation which is traditionally addressed in the framework of the multi-armed bandits theory. Based on previous work on bandit theory applied to active learning for regression, we introduce two novel algorithms for solving the online allocation of the budget in a classification problem. Experiments on a generic classification problem demonstrate that these new algorithms compare positively to state-of-the-art methods.

  • On Improving Operational Planning and Control in Public Transportation Networks using Streaming Data: A Machin3 Learning Approach (Luis Moreira-Matias, Joao Moreira, Joao Gama, Michel Ferreira)

    Authors

    Luis Moreira-Matias, Telecommunications Institute
    Joao Moreira, LIAAD-INESC TEC
    Joao Gama, University of Porto
    Michel Ferreira, U. Porto

    Abstract

    Nowadays, the transportation vehicles are equipped with intelligent sensors. Together, they form collaborative networks which are broadcasting real-time data about the mobility patterns in an urban area. Online intelligent transportation systems for taxi dispatching, time-saving route finding or automatic vehicle location are already exploring such information on the taxi/buses transport industries. In this PhD spotlight paper, the authors presented two ML applications focused on improving Public Transportation (PT) operations: 1) Bus Bunching (BB) Online Detection and the 2) Taxi-Passenger Demand Prediction. By doing so, we intend to give an overview on the type of approaches applicable to these type of problems at a glance. Our frameworks are straightforward by employing online learning frameworks able to use both historical and real-time data to update the inference models. The results are promising.

Accepted for Poster presentation

  • Generalizing, Optimizing, and Decoding Support Vector Machine Classification (Mario Michael Krell, Sirko Straube, Hendrik Wöhrle, Frank Kirchner)

    Authors

    Mario Michael Krell, University of Bremen
    Sirko Straube, University of Bremen
    Hendrik Wöhrle, DFKI Robotics Innovation Center
    Frank Kirchner, DFKI Robotics Innovation Center, University of Bremen

    Abstract

    A major challenge in the classification of spatiotemporal data, that requires the combination of several processing steps, is the selection of algorithms for preprocessing and classification. Here, we present three steps to face this problem. First, we propose the signal processing and classification environment pySPACE which enables the systematic evaluation and comparison of algorithms. Second, we introduce a generalized model for Support Vector Machine variants which generates both unary and online classifiers. This model improves the understanding of relationships between the variants which facilitates the choice of the classifier. Third, we introduce an approach called backtransformation which enables a visualization of the complete processing chain in the the input data space and thereby allows for a joint interpretation of preprocessing and classification to decode the decision process. Finally, the benefit of combining all three approaches is shown in an application. on handwritten digit classification.

  • Heterogeneous Dataflow Hardware Accelerators for Machine Learning on Reconfigurable Hardware (Hendrik Wöhrle, Johannes Teiwes, Mario Krell, Anett Seeland, Elsa Kirchner, Frank Kirchner)

    Authors

    Hendrik Wöhrle, DFKI Robotics Innovation Center
    Johannes Teiwes, University of Bremen
    Mario Krell, University of Bremen
    Anett Seeland, DFKI GmbH Robotics Innovation Center
    Elsa Kirchner, DFKI GmbH Robotics Innovation Center
    Frank Kirchner, DFKI Robotics Innovation Center, University of Bremen

    Abstract

    The trend in robotics is to develop increasingly more intelligent systems. This leads directly to a considerable demand for more and more computational power. However, there are several technical limitations, like restrictions regarding power consumption and physical size, that make the use of powerful generic processors unfeasible. One possibility to overcome this problem is the usage of specialized hardware accelerators, which are designed for typical tasks in robotics and machine learning. In this paper, we propose an approach for the rapid development of hardware accelerators that are based on the heterogeneous dataflow computing paradigm. The developed techniques are collected in a framework to provide a simple access to them. We discuss different application areas and show first results in the field of biosignal analysis that can be used for rehabilitation robotics.

  • Evaluating Collaborative Filtering: Methods within a Binary Purchase Setting (Stijn Geuens, Kristof Coussement, Koen De Bock)

    Authors

    Stijn Geuens, IESEG School of Management
    Kristof Coussement, IESEG School of Management
    Koen De Bock, IESEG School of Management

    Abstract

    The study of recommender systems based upon online user behavior constitutes an under-investigated area. The objective of this study is to evaluate configuration options of memory-based collaborative filtering (CF) for generating recommendations based upon online binary purchase data with different characteristics. First different algorithm configurations will be identified. More specifically, three important algorithm parameters are investigated: the data reduction technique, the user- versus item-based CF and the similarity measure. Preliminary results on data from a large European apparel E-tailor show that especially the used reduction technique and the CF-method influence the accuracy results and computation time of an algorithm. In a second phase, yet to be executed, extended experiments will be set up to gain more insight into the influence of input data characteristics on the relative success of the CF configuration options. In particular, three input characteristics, sparsity level, item purchase distribution and item/user ratio are manipulated to analyze the impact on the algorithm’s best configuration.

  • An opinion mining Partial Least Square Path Modeling for football betting (Mohamed EL HAMDAOUI, Jean-Valère Cossu)

    Authors

    Mohamed EL HAMDAOUI, LIA UAPV
    Jean-Valère Cossu, LIA

    Abstract

    In the last few years, football betting had known a large expansion in the world, using different ways to try to guess and predict the unknown in the sport. Every time, people try to prognosticate the results of matches using probabilistic, statistic and other methods to get the maximum benefits, especially with the emerging of betting websites. In this paper, we present an alternative approach, to state of the art probabilistic models, based on Partial Least Square Path Modeling (PLS-PM). We first show that the simple PLS model containing only statistical resources about each team are efficient to predict the team ranking at d+1 and this gives a state of the art prediction of match outcomes. We then take advantage of PLS ability of integrating complex and heterogeneous data to reach a practical model by including textual data, taken from tweets related to teams, that we previously classify by polarity using robust sentiment analysis in multiple languages. Another learning of our experiment is the role of the inner model in PLS when used for prediction purpose. Unlikely Bayesian networks, the latent variable used in the prediction need to be deeply inside the inner model and not considered as marginal outcomes, this to allow back and forth retro-propagation from multiple types of data. The main purpose of our work is to show that PLS-PM can be surprisingly efficient in predicting tournament outcomes for which temporal statistics and social network data are available if inference is based on central inner latent variables.

  • Multivariate Normal Distribution Based Multi-Armed Bandits Pareto Algorithm (Saba Yahyaa, Madalina Drugan, Bernard Manderick)

    Authors

    Saba Yahyaa, Vrije Universiteit Brussel
    Madalina Drugan, VUB
    Bernard Manderick, Vrije Universiteit Brussel

    Abstract

    In the stochastic multivariate multi-armed bandit, arms generate a vector of stochastic normal rewards, one per objective, instead of a single scalar reward. As a result, there is not only one optimal arm, but there is a set of optimal arms (Pareto front) using Pareto dominance relation. The goal of an agent is to trade-off between exploration and exploitation. Exploration means finding the Pareto front and exploitation means selecting fairly or evenly the optimal arms. We propose annealing-Pareto algorithm that trades-off between exploration and exploitation by using a decaying parameter εt in combination with Pareto dominance relation. We compare experimentally Pareto-KG, Pareto-UCB1 and annealing-Pareto on multi-objective normal distributions and we conclude that the annealing-Pareto is the best performing algorithm. Keywords: Multi armed bandit problem, multi objective optimization, annealing algorithm, exploration/exploitation.

  • Managing Ventilation Systems for Improving User Comfort in Smart Buildings using Reinforcement Learning Agents (Jiawei ZHU, Fabrice LAURI, Abderrafiaa KOUKAM, Vincent Hilaire)

    Authors

    Jiawei ZHU, IRTES-SeT
    Fabrice LAURI, IRTES-SeT
    Abderrafiaa KOUKAM, IRTES-SeT
    Vincent Hilaire

    Abstract

    With the fast development of information technology and increasingly prominent environmental problems, building comfort and energy management become the major tasks for an intelligent residential building system. This paper identifies the system requirements of Smart Buildings, analyzes the problems that need to be solved and how Reinforcement Learning is suitable for dealing with them. It also proposes to represent parts of Smart Buildings as Cyber-Physical Systems. Although the global goal is to model and manage a complex and whole system of a Smart Building, since the work is in progress, in this paper we mainly focus on how Reinforcement Learning technique is good at controlling subsystems, specifically the Ventilation System. The experimental results show the advantages of our system compared with the widely used baselines: On/Off control and PI control approaches.

  • Robust Optimization using Machine Learning for Uncertainty Sets (Theja Tulabandhula)

    Author

    Theja Tulabandhula, MIT

    Abstract

    Our goal is to build robust optimization problems that make decisions about the future, and where complex data from the past are used to model uncertainty. In robust optimization (RO) generally, the goal is to create a policy for decision-making that is robust to our uncertainty about the future. In particular, we want our policy to best handle the the worst possible situation that could arise, out of an uncertainty set of possible situations. Classically, the uncertainty set is simply chosen by the user, or it might be estimated in overly simplistic ways with strong assumptions; whereas in this work, we learn the uncertainty set from complex data from the past. The past data are drawn randomly from an (unknown) possibly complicated high-dimensional distribution. We propose a new uncertainty set design and show how tools from statistical learning theory can be employed to provide probabilistic guarantees on the robustness of the policy.

  • Parallel Learning Algorithm for Large-Scale Regression with Additive Models (Valeriy Khakhutskyy, Markus Hegland)

    Authors

    Valeriy Khakhutskyy, Technische Universität München
    Markus Hegland, Australian National University

    Abstract

    We present a novel parallel algorithm for training additive regression models. The approach relates to a number of other methods including the backfitting algorithm and alternating direction method of multipliers. However, we extend the scope of possible applications to include any ANOVA-type decomposition or an ensemble of random subspace projection models. We derive a distributed parallel algorithm for this method with good weak and strong scaling characteristics. The experimental results illustrate the convergence and scaling properties of the algorithm on real and synthetic data.

  • A Framework for Pattern Classifier Selection and Fusion (Fabio Faria, Anderson Rocha, Ricardo Torres)

    Authors

    Fabio Faria, University of Campinas
    Anderson Rocha, University of Campinas
    Ricardo Torres, University of Campinas

    Abstract

    In this work, we propose a framework for classifier selection and fusion. Our method seeks to combine image characterization and learning methods by means of a meta-learning approach responsible for assessing which methods contribute more towards the solution of a given problem. The framework uses three different strategies of classifier selection that pinpoint the less correlated, yet effective, classifiers through a series of diversity measure analysis. The experiments show that the proposed approaches yield comparable results to well-known algorithms from the literature on many different applications but using less learning and description methods as well as not incurring in the curse of dimensionality and normalization problems common to some fusion techniques. Furthermore, our approach yields effective classification results using very reduced training sets.

  • Affinity Analysis between Researchers using Text Mining and Differential Analysis of Graphs (Luis Trigo, Pavel Brazdil)

    Author

    Luis Trigo, INESC-TEC
    Pavel Brazdil, INESC-TEC

    Abstract

    Finding people with similar skills within a domain may provide an important support for managing research centers. The academic production, albeit in an unstructured format, is easily accessible. Thus, we resort to sources on the web – academic and bibliographic databases – to uncover the affinities among researchers. What interests us most are affinities that are not yet evidenced by co-authorship. Besides, of interest are also other outputs of the method in the form of subgroups and the researchers that play an important role in them.

  • NASSAU: Description Length Minimization for Boolean Matrix Factorization (Sanjar Karaev)

    Author

    Sanjar Karaev, MPI for informatics

    Abstract

    Boolean Matrix Factorization (BMF) is an improtant tool in data mining that in many cases allows to increase interpretability for bi- nary data. In BMF one decomposes a given binary matrix into a Boolean product of binary factors such that some cost function is minimized. In this work we consider the description length of the data as the cost, which has been proven effective in uncovering true structure of the data and removing the noise. The argument is that structured data is easier to compress than the noise, and hence simpler models should be favored. We introduce a new BMF algorithm, that we call NASSAU, which traces back its history and correct its earlier mistakes. As turned our in our experiments, this approach performs reasoably well for both real-world and synthetic data.

X