Online social networks, including Twitter, have expanded in both scale and diversity of content, which has created significant challenges to the average user. These challenges include finding relevant information on a topic and building social ties with like-minded individuals. The fundamental question addressed by this thesis is if an individual can leverage social network to search for information that is relevant to him or her. We propose to answer this question by developing computational algorithms that analyze a user's social network. The features of the social network we analyze include the network topology and member communications of a specific user's social network. Determining the "social value" of one's contacts is a valuable outcome of this research. The algorithms we developed were tested on Twitter, which is an extremely popular social network. Twitter was chosen due to its popularity and a majority of the communications artifacts on Twitter is publically available. In this work, the social network of a user refers to the "following relationship" social network. Our algorithm is not specific to Twitter, and is applicable to other social networks, where the network topology and communications are accessible. My approaches are as follows. For a user interested in using the system, I first determine the immediate social network of the user as well as the social contacts for each person in this network. Afterwards, I establish and extend the social network for each user. For each member of the social network, their tweet data are analyzed and represented by using a word distribution. To accomplish this, I use WordNet, a popular lexical database, to determine semantic similarity between two words. My mechanism of search combines both communication distance between two users and social relationships to determine the search results. Additionally, I developed a search interface, where a user can interactively query the system. I conducted preliminary user study to evaluate the quality and utility of my method and system against several baseline methods, including the default Twitter search. The experimental results from the user study indicate that my method is able to find relevant people and identify valuable contacts in one's social circle based on the query. The proposed system outperforms baseline methods in terms of standard information retrieval metrics.

Advances in the area of ubiquitous, pervasive and wearable computing have resulted in the development of low band-width, data rich environmental and body sensor networks, providing a reliable and non-intrusive methodology for capturing activity data from humans and the environments they inhabit. Assistive technologies that promote independent living amongst elderly and individuals with cognitive impairment are a major motivating factor for sensor-based activity recognition systems. However, the process of discerning relevant activity information from these sensor streams such as accelerometers is a non-trivial task and is an on-going research area. The difficulty stems from factors such as spatio-temporal variations in movement patterns induced by different individuals and contexts, sparse occurrence of relevant activity gestures in a continuous stream of irrelevant movements and the lack of real-world data for training learning algorithms. This work addresses these challenges in the context of wearable accelerometer-based simple activity and gesture recognition. The proposed computational framework utilizes discriminative classifiers for learning the spatio-temporal variations in movement patterns and demonstrates its effectiveness through a real-time simple activity recognition system and short duration, non- repetitive activity gesture recognition. Furthermore, it proposes adaptive discriminative threshold models trained only on relevant activity gestures for filtering irrelevant movement patterns in a continuous stream. These models are integrated into a gesture spotting network for detecting activity gestures involved in complex activities of daily living. The framework addresses the lack of real world data for training, by using auxiliary, yet related data samples for training in a transfer learning setting. Finally the problem of predicting activity tasks involved in the execution of a complex activity of daily living is described and a solution based on hierarchical Markov models is discussed and evaluated.

Natural Language Processing is a subject that combines computer science and linguistics, aiming to provide computers with the ability to understand natural language and to develop a more intuitive human-computer interaction. The research community has developed ways to translate natural language to mathematical formalisms. It has not yet been shown, however, how to automatically translate different kinds of knowledge in English to distinct formal languages. Most of the recent work presents the problem that the translation method aims to a specific formal language or is hard to generalize. In this research, I take a first step to overcome this difficulty and present two algorithms which take as input two lambda-calculus expressions G and H and compute a lambda-calculus expression F. The expression F returned by the first algorithm satisfies F@G=H and, in the case of the second algorithm, we obtain G@F=H. The lambda expressions represent the meanings of words and sentences. For each formal language that one desires to use with the algorithms, the language must be defined in terms of lambda calculus. Also, some additional concepts must be included. After doing this, given a sentence, its representation and knowing the representation of several words in the sentence, the algorithms can be used to obtain the representation of the other words in that sentence. In this work, I define two languages and show examples of their use with the algorithms. The algorithms are illustrated along with soundness and completeness proofs, the latter with respect to typed lambda-calculus formulas up to the second order. These algorithms are a core part of a natural language semantics system that translates sentences from English to formulas in different formal languages.

The fields of pattern recognition and machine learning are on a fundamental quest to design systems that can learn the way humans do. One important aspect of human intelligence that has so far not been given sufficient attention is the capability of humans to express when they are certain about a decision, or when they are not. Machine learning techniques today are not yet fully equipped to be trusted with this critical task. This work seeks to address this fundamental knowledge gap. Existing approaches that provide a measure of confidence on a prediction such as learning algorithms based on the Bayesian theory or the Probably Approximately Correct theory require strong assumptions or often produce results that are not practical or reliable. The recently developed Conformal Predictions (CP) framework - which is based on the principles of hypothesis testing, transductive inference and algorithmic randomness - provides a game-theoretic approach to the estimation of confidence with several desirable properties such as online calibration and generalizability to all classification and regression methods. This dissertation builds on the CP theory to compute reliable confidence measures that aid decision-making in real-world problems through: (i) Development of a methodology for learning a kernel function (or distance metric) for optimal and accurate conformal predictors; (ii) Validation of the calibration properties of the CP framework when applied to multi-classifier (or multi-regressor) fusion; and (iii) Development of a methodology to extend the CP framework to continuous learning, by using the framework for online active learning. These contributions are validated on four real-world problems from the domains of healthcare and assistive technologies: two classification-based applications (risk prediction in cardiac decision support and multimodal person recognition), and two regression-based applications (head pose estimation and saliency prediction in images). The results obtained show that: (i) multiple kernel learning can effectively increase efficiency in the CP framework; (ii) quantile p-value combination methods provide a viable solution for fusion in the CP framework; and (iii) eigendecomposition of p-value difference matrices can serve as effective measures for online active learning; demonstrating promise and potential in using these contributions in multimedia pattern recognition problems in real-world settings.

Prognostics and health management (PHM) is a method that permits the reliability of a system to be evaluated in its actual application conditions. This work involved developing a robust system to determine the advent of failure. Using the data from the PHM experiment, a model was developed to estimate the prognostic features and build a condition based system based on measured prognostics. To enable prognostics, a framework was developed to extract load parameters required for damage assessment from irregular time-load data. As a part of the methodology, a database engine was built to maintain and monitor the experimental data. This framework helps in significant reduction of the time-load data without compromising features that are essential for damage estimation. A failure precursor based approach was used for remaining life prognostics. The developed system has a throughput of 4MB/sec with 90% latency within 100msec. This work hence provides an overview on Prognostic framework survey, Prognostics Framework architecture and design approach with a robust system implementation.