The primary public health goal is to increase the number of years of good health and, therefore, maintain independence and quality of life as long as possible. Healthy ageing is characterized by the avoidance of disease and disability, the maintenance of high physical and cognitive function, and sustained engagement in social and productive activities. These three components together define successful ageing. An important part of successful ageing, hence, is maximisation of physical performance. The ability to fully participate in productive and recreational activities of daily life may be af-fected when the capacity to easily perform common physical functions decreases. Health status, thus, is an important indicator of quality of life among older people. It appears that especially components of health-related fitness and functional performance, or serious, chronic conditions and diseases that directly influence the components of fitness and performance, are related to perceived health among middle-aged and older adults. Regular physical activity or exercise substantially prevents the development and progres-sion of most chronic degenerative diseases. In summary, it is evident that to increase older adults’ quality of life and fitness, we need to encourage the elderly to become more physically active and increase their fitness through training. Home environmental interventions to prevent functional decline seem to be effective and are, furthermore, preferred by elderly. Such interventions with integrated assistive technology devices have, in this context, the potential to further help in overcoming some of the barriers to start training and, thereby, maintaining physical independence for independently living elderly. Hence, the objective of this thesis is to identify how, through IT or IT-mediated persuasive soft-ware applications, we can enable independently living and healthy elderly people to perform balance and strength training plans autonomously at home and keep them motivated, in order to increase their compliance toward the plans.

The goal of multichannel audio source separation is to produce high quality separated audio signals, observing mixtures of these signals. The difficulty of tackling the problem comes from not only the source propagation through noisy and echoing environments, but also overlapped source signals. Among the different research directions pursued around this problem, the adoption of probabilistic and advanced modeling aims at exploiting the diversity of multichannel propagation, and the redundancy of source signals. Moreover, prior information about the environments or the signals is helpful to improve the quality and to accelerate the separation. In this thesis, we propose methods to increase the effectiveness of model-based audio source separation methods by exploiting prior information applying spectral and sparse modeling theories. The work is divided into two main parts. In the first part, spectral modeling based on Nonnegative Matrix Factorization is adopted to represent the source signals. The parameters of Gaussian model-based source separation are estimated in sense of Maximum-Likelihood using a Generalized Expectation-Maximization algorithm by applying supervised Nonnegative Matrix and Tensor Factorization, given spectral descriptions of the source signals. Three modalities of making the descriptions available are addressed, i.e. the descriptions are on-line trained during the separation, pre-trained and made directly available, or pre-trained and made indirectly available. In the latter, a detection method is proposed in order to identify the descriptions best representing the signals in the mixtures. In the second part, sparse modeling is adopted to represent the propagation environments. Spatial descriptions of the environments, either deterministic or probabilistic, are pre-trained and made indirectly available. A detection method is proposed in order to identify the deterministic descriptions best representing the environments. The detected descriptions are then used to perform source separation by minimizing a non-convex $l_0$-norm function. For speaker diarization where the task is to determine ``who spoke when" in real meetings, a Watson mixture model is optimized using an Expectation-Maximization algorithm in order to detect the probabilistic descriptions, best representing the environments, and to estimate the temporal activity of each source. The performance of the proposed methods is experimentally evaluated using different datasets, between simulated and live-recorded. The elaborated results show the superiority of the proposed methods over recently developed methods used as baselines.

This thesis extracts conceptual structures from multiple sources: Wordnet, Web Corpora and Wikipedia. The conceptual structures extracted from Wordnet and Web Corpora are inspired by the feature norm effort in cognitive psychology. The conceptual structure extracted from Wikipedia makes the transition between feature norm structures and theory like structures. The main contribution of this thesis can be grouped in two categories: 1. Novel methods for the extraction of conceptual structures. More precisely, there are three new methods we developed: (a) Conceptual structure extraction from Wordnet. We devise a procedure for property extraction from Wordnet using the notion of semantic neighborhood. The procedure exploits the main relations organizing the nouns, the information in glosses and the inheritance of properties principle. (b) Feature Norms like extraction from corpora. We propose a method to acquire feature norm like structures from corpora using weakly supervised methods. (c) Conceptual Structure from Wikipedia. A novel unsupervised method for the extraction of conceptual structures from Wikipedia entries of similar concepts is put forward. The main idea we follow is that similar concepts (i.e. those classied under the same node in a taxonomy) are described in a comparable way in Wikipedia. Moreover, to understand the kind of information extracted from Wikipedia we annotate this knowledge with a set of property types. 2. Evaluation. We evaluate Wordnet as a model of semantic memory and suggest the addition of new semantic relations. We also assess the properties extracted from all sources for a unified test set, in a clustering experiment.

Dynamic constraints such as Separation of Duty (SoD) prevent the possibility of frauds and enable flexible protection of sensitive resources appearing in active contexts. They are enforced in various ways depending on the access control model and the application. Role based access control (RBAC) employs restrictions on the activation of roles and the exercise of permissions by individuals for enforcing the constraints. However, whether a constraint specification correctly enforces a given dynamic policy is an open research question. This is mainly due to the nature of the dynamic constraint enforcement: a constraint satisfied at a state can be violated at a future state as a result of the event sequences occurred in between. Moreover, the support of dynamic enforcement usually imposes low level extensions to the implementation, which in return requires another level of verification. In the approaches that tackle this problem at run-time, efficiency is a key concern. In this dissertation, we present two approaches for analyzing and enforcing dynamic constraints. The first is employed on-line and is based on software testing features available in software model checkers. The relevant components of an access control system are modeled as a software and the execution of this software mimics the RBAC run-time. A software model checker is used to check some properties that represent constraint specifications and the actual authorization policies encoded in eXtensible Access Control Language (XACML). We demonstrate our approach by using an open source software model checker, Java Path Finder (JPF), and its sub-projects for dierent testing scenarios. In this first approach, efficiency is not the main concern but coverage is. The second approach relies on a propositional satisability (SAT) based run-time procedure to replace the conventional policy evaluation in RBAC systems. Efficiency and flexibility are the prominent features of this approach. Efficiency is obtained by dividing the steps involved in policy evaluation into on-line and off-line. On-line steps correspond to request answering in conventional policy evaluation and have to be done at run-time. On-line steps can be performed as pre-processing or post-processing of the on-line steps and have no effect on policy evaluation performance. We experimentally show that our approach is efficient and scales well in realistic scenarios. The final chapter of the thesis presents an extensive study of XACML policy evaluation performance. Policy evaluation corresponds to a function, Eval(Policy,Request), that takes a policy and a request as input, and produces an access control decision. Our experimental results show that the Eval function can create a bottleneck in application domains where the number of policies and rules is large. We present a list of optimization techniques that can speed up the evaluation performance.

In the past years we have witnessed Sentiment Analytics becoming increasingly popular topic in Information Retrieval, which has established itself as a promising direction of research. With the rapid growth of the user-generated content represented in blogs, forums, social networks and micro-blogs, it became a useful tool for social studies, market analysis and reputation management, since it made possible capturing sentiments and opinions at a large scale and with the ever-growing precision. Sentiment Analytics came a long way from product review mining to full-fledged multi-dimensional analysis of social sentiment, exposing people attitude towards any topic aggregated along different dimensions, such as time and demographics. The novelty of our work is that it approaches Sentiment Analytics from the perspective of Data Mining, addressing some important problems which fall out of the scope of Opinion Mining. We develop a framework for Large Scale Aggregated Sentiment Analytics, which allows to capture and quantify important changes in aggregated sentiments or in their dynamics, evaluate demographical aspects of these changes, and explain the underlying events and mechanisms which drive them. The first component of our framework is Contradiction Analysis, which studies diverse opinions and their interaction, and allows tracking the quality of aggregated sentiment or detecting interesting sentiment differences. Targeting large scale applications, we develop a sentiment contradiction measure based on the statistical properties of sentiment and allowing efficient computation from aggregated sentiments. Another important component of our framework addresses the problem of monitoring and explaining temporal sentiment variations. Along this direction, we propose novel time series correlation methods tailored specifically for large scale analysis of sentiments aggregated over users demographics. Our methods help to identify interesting correlation patterns between demographic groups and thus better understand the demographical aspect of sentiment dynamics. We bring another interesting dimension to the problem of sentiment evolution by studying the joint dynamics of sentiments and news, uncovering the importance of news events and assessing their impact on sentiments. We propose a novel and universal way of modeling different media and their dynamics, which aims to describe the information propagation in news- and social media. Finally, we propose and evaluate an updateable method of sentiment aggregation and retrieval, which preserves important properties of aggregated sentiments and also supports scalability and performance requirements of our applications.

Technology is evolving towards a higher degree of automation and connectivity, with concepts of Pervasive Computing, Smart Factories, Cyber-Physical Systems (CPS) and the Internet of Things (IoT)promising to integrate countless communicating devices into objects around us at home, in the streets, and on industrial sites. These embedded devices are often very small in size, autonomously-powered, and have restricted computational and communicational capabilities. Low-power and Lossy Networks (LLNs) are multi-hop, typically wireless, self-organising networks aimed at interconnecting hundreds or thousands of such embedded devices. They inherit many techniques from wireless sensor networks, though going beyond their original task of collecting sensor readings. New applications comprising actuators, control loops, user interface devices and requiring connectivity of every ``smart thing'' with the Internet, pose new challenges to the network protocol stacks. These stacks should not only efficiently support data collection from numerous low-power sensors, but provide scalable data forwarding in the opposite direction, making every single device in the LLN addressable and reachable from a central controller or from the Internet. This type of forwarding is needed to send commands to wireless actuators in the LLN or to enable request-response communication between a low-power device and a remote server. Control loops additionally require real-time guarantees from the communication system. We demonstrate in this thesis that reconciling the battery lifetime with high reliability and low latency is still a challenge for existing protocols even at the scale of few hundreds of network nodes. Moreover, current techniques have a significant performance gap between their data collection and actuation forwarding components on memory-constrained platforms. This gap limits the applicability of the stacks, as the overall performance is determined by the weaker component. Motivated by two real-life applications, we first study novel techniques that eliminate the performance gap in the standard IPv6 stack for LLNs, making the actuation traffic forwarding as performant as the data collection one in networks that are five times larger than what the original standard stack is able to support. Second, we demonstrate that the reliability of packet delivery in the standard-compliant solution is limited in practice at around 99% while its routing overhead causes significant inefficiency in energy consumption. Therefore, we change focus to a forwarding mechanism based on the principle of synchronous transmissions, made popular by Glossy. It is a recent and, thus, non-standard technique, known for excellent reliability, speed and energy efficiency of the flooding-based data dissemination service it provides. This service is a perfect match for actuation, but a similarly efficient data collection protocol did not exist. To close this gap, we design Crystal, a novel data collection protocol based on the same core principle of synchronous transmissions. We show that, depending on the application, Crystal reaches per-mille or even parts-per-million radio duty cycle. It does that with a packet loss rate lower than 10e-5 under external Wi-Fi interference of a noisy office building, and provides a much higher reliability and energy efficiency than the state of the art under even stronger interference generated by JamLab. We thoroughly evaluate the proposed solutions both in realistic simulations and two large-scale testbeds. We follow a principled approach based on understanding of the environment and the properties of the network topologies. The latter are acquired by our connectivity assessment tool Trident, which itself is one of the contributions of this thesis. Through these contributions, this thesis pushes forward the applicability of LLNs, by improving their scalability, reliability, latency, energy efficiency and interference resilience, both in the context of an existing standard and in a clean-slate design. Further, by achieving this superior performance via network stacks that natively support both collection and actuation traffic, this thesis provides a stepping stone for applications that strongly rely on both, notably including the low-power wireless control applications.

The present thesis deals with the fundamental machine learning issues of increasing the accuracy of learning systems and their computational performances. The key concept which is exploited throughout the thesis, is the tunable trade-off between local and global approaches to learning, integrating the effective setting of Instance Based Learning with the sound foundations of Statistical Learning Theory. Four are the main contributions of the thesis in this context: (i) a theoretical analysis and empirical evaluation of the Local SVM approach, (ii) a family of operators on kernels to obtain Quasi-Local kernels, (iii) the framework of Local Kernel Machines, and (iv) a local maximal margin approach to noise reduction. In our analysis of Local SVM, we derive a new learning bound starting from the theory of Local Learning Algorithms, and we showed that Local SVM statistically significantly overcomes the classification accuracy of SVM in a number of scenarios. The novel family of operators on kernels integrates local feature-space information into any kernel obtaining Quasi-Local kernels, mixing the effect of the input kernel with a kernel which is local in the feature space of the input one. With Local Kernel Machine we show that locality can be exploited to obtain fast and scalable kernel machines, whereas existing fast approximated SVM solvers try to globally smooth the decision functions. Fast Local Kernel SVM (FaLK-SVM) trains a set of local SVMs on redundant neighbourhoods in the training set selecting at testing time the most appropriate model for each query point. Theoretically supported by a recent result relating consistency and localizability, our approach divides the separation function in solutions of local optimization problems that can be handled very efficiently. For this novel approach, we derive a fast local model selection strategy, theoretical learning bounds and favourable complexity bounds. Local Kernel Machines can also be applied to the problem of detecting and removing noisy examples from datasets in order to enhance the generalization ability of Instance Based Learning and for data-cleansing. The local maximal margin principle provides a more robust alternative to the majority rule on which almost all the existing noise reduction techniques are based and a scalable version of the approach extends the feasibility of the noise reduction task to large datasets such as genomic-scale biological data. Extensive evaluations of the proposed techniques are carried out on more than 100 datasets with up to 3 millions examples, and statistically significantly showed that Quasi-Local kernels are more accurate than the corresponding input kernels using SVM, and that Local Kernel Machines can improve the generalization ability of accurate and approximated SVM solvers and of traditional noise-reduction techniques with much faster training and testing times and better scalability performances.

Information Extraction (IE) aims at mapping texts into fixed structure representing the key information. A typical IE system will try to answer the questions like who are present in the text, what events happen and when these events happen. The task is making possible significant advances in applications that require deep understanding capabilities such as question-answering engines, dialogue systems, or the semantic web. Due to the huge effort and time consumation of developping extraction systems by domain experts, our approach focuses on machine learning methods that can accurately infer an extraction model by training on a dataset. The goal of this research is to design and implement models with improved performance by learning the combination of different algorithms or by inventing novel structures that are able to exploit kinds of evidence that have not been explored in the literature. A basic component of an IE system is named entity recognition (NER) whose purpose is to locate objects that can be referred by names, belonging to a predefined set of categories. We approach this task by proposing a novel reranking framework that employs two learning phases to pick the best candidate. The task is considered as sequence labelling with Conditional Random Fields (CRFs) is selected as the baseline algorithm. Our research employs novel kernels based on structured and unstructured features for reranking the N-best hypotheses from the CRFs baseline. The former features are generated by a polynomial kernel encoding entity features whereas tree kernels are used to model dependencies amongst tagged candidate examples. Relation Extraction (RE) is concerned with finding relationships between pairs of entities in texts. State-of-the-art relation extraction model is based on convolution kernel over the constituent parse tree. In our research, we employ dependency parses from dependency parsing in addition to phrase-structure parses from constituent parsing. We define several variations of dependency parses to inject additional information into the trees. Additionally, we provide an extensive ablation over various types of kernels by combining the tree, sequence, and polynomial kernels. These novel kernels are able to exploit learned correlations between phrase-structure parses and grammatical relations. A large amounts of wide-coverage semantic knowledge today exists in large repositories of unstructured or semi-structured text documents. The increased availability of online collaborative resources has attracted the attention of much work in the Artificial Intelligence (AI) community. Nevertheless, the ability to extract it using statistical machine learning techniques is hindered by well-known problems such as heavy supervision and scalability. These drawbacks can be alleviated by applying a form of weakly supervision, specifically named distant supervision (DS), to automatically derive explicit facts from the semi-structured part of Wikipedia. To learn relational facts from Wikipedia without any labeled example or hand-crafted pattern, we employ DS where the relation providers are external repositories, e.g., YAGO (a huge semantic knowledge base), and the training instances are gathered from Freebase (a huge semantic database). These allow for potentially obtaining larger training data and many more relations, defined in different sources. We apply state-of-the-art models for ACE RE, that are sentence level RE (SRLE), to Wikipedia. Based on a mapping table of relations from YAGO to ACE (according to their semantic definitions), we design a joint RE model of DS/ACE and tested it on ACE annotations (thus according to expert linguistic annotators). Moreover, we experiment with end-to-end systems for real-world RE applications. Consequently, our RE system is applicable to any document/sentence, i.e. another major improvement on previous work, which, to our knowledge, does not show experiments on end-to-end SLRE.

The ranking aggregation problem is that to establishing a new aggregate ranking given a set of rankings of a finite set of items. This problem is met in various applications, such as the combination of user preferences, the combination of lists of documents retrieved by search engines and the combination of ranked gene lists. In the literature, the ranking aggregation problem has been solved as an optimization of some distance between the rankings overlooking the existence of a true ranking. In this thesis we address the ranking aggregation problem assuming the existence of a true ranking on the set of items: the goal is to estimate an unknown, true ranking given a set of input rankings provided by experts with different approximation quality. We propose a novel solution called Belief Ranking Estimator (BRE) that takes into account two aspects still unexplored in ranking combination: the approximation quality of the experts and for the first time the uncertainty related to each item position in the ranking. BRE estimates in an unsupervised way the true ranking given a set of rankings that are diverse quality estimations of the unknown true ranking. The uncertainty on the items's position in each ranking is modeled within the Belief Function Theory framework, that allows for the combination of subjective knowledge in a non Bayesian way. This innovative application of belief functions to rankings, allows us to encode different sources of a priori knowledge about the correctness of the ranking positions and also to weigh the reliability of the experts involved in the combination. We assessed the performance of our solution on synthetic and real data against state-of-the-art methods. The tests comprise the aggregation of total and partial rankings in different empirical settings aimed at representing the different quality of the input rankings with respect to the true ranking. The results show that BRE provides an effective solution when the input rankings are heterogeneous in terms of approximation quality with respect to the unknown true ranking.

This thesis is concerned with the problem of automatic extraction of harmonic and rhythmic information from music audio signals using statistical framework and advanced signal processing methods. Among different research directions, automatic extraction of chords and key has always been of a great interest to Music Information Retrieval (MIR) community. Chord progressions and key information can serve as a robust mid-level representation for a variety of MIR tasks. We propose statistical approaches to automatic extraction of chord progressions using Hidden Markov Models (HMM) based framework. General ideas we rely on have already proved to be effective in speech recognition. We propose novel probabilistic approaches that include acoustic modeling layer and language modeling layer. We investigate the usage of standard N-grams and Factored Language Models (FLM) for automatic chord recognition. Another central topic of this work is the feature extraction techniques. We develop a set of new features that belong to chroma family. A set of novel chroma features that is based on the application of Pseudo-Quadrature Mirror Filter (PQMF) bank is introduced. We show the advantage of using Time-Frequency Reassignment (TFR) technique to derive better acoustic features. Tempo estimation and beat structure extraction are amongst the most challenging tasks in MIR community. We develop a novel method for beat/downbeat estimation from audio. It is based on the same statistical approach that consists of two hierarchical levels: acoustic modeling and beat sequence modeling. We propose the definition of a very specific beat duration model that exploits an HMM structure without self-transitions. A new feature set that utilizes the advantages of harmonic-impulsive component separation technique is introduced. The proposed methods are compared to numerous state-of-the-art approaches by participation in the MIREX competition, which is the best impartial assessment of MIR systems nowadays.