No description available.

Nowadays we are living in an era of rich data, not only in the sense of the amount of data, but also in the sense of various sources and content of data. Efficient search, management, and exploitation of data have, over decades, been a major direction of database research. In this thesis, three challenging problems are proposed and studied, targeting (i) time series data, (ii) user preference data, and (iii) location-based social network data, respectively, providing efficient solutions to corresponding real-life applications. First, durability queries are studied in historical time series databases, which identify objects that have durable quality over time. For example, a sociologist may be interested in the top 10 web search terms during the period of some historical events; the police may seek for vehicles that move close to a suspect 70% of the time during a certain time, etc. Such durable top-k (DTop-k) and durable k-nearest neighbor (DkNN) queries can be viewed as natural extensions of the standard snapshot top-k and NN queries to timestamped sequences of values or locations. Although their snapshot counterparts have been studied extensively, there is little prior work that addresses this new class of durability queries. Efficient and scalable algorithms are proposed based on novel indexing techniques. Next, an efficient solution to k-nearest neighbor search over top-m lists is investigated. A top-m list is a ranking of m items, typically representing some user’s preference over these items. For example, a user may have a list of her 10 most favourite books; the result from a search engine is typically a list of webpages ranked according to their relevance to some keywords. The search problem aims at extracting k top-m lists from the database that are the “closest” to some query list where the closeness is evaluated using commonly used measures such as the Fagin’s intersection metric, Spearman’s footrule, Kendall’s tau, etc. Despite of the importance of such queries, there’s little prior work suggesting any efficient solution. In this thesis, a unified framework is proposed to answer such queries efficiently. Finally, the problem of top-N venue recommendation in location-based social networks (LBSNs) is studied, which recommends new venues to users. As an increasingly larger number of users partake in LBSNs, the recommendation problem in this setting has attracted significant attention in research and in practical applications. The detailed information about past user behavior that is traced by the LBSN differentiates the problem significantly from its traditional settings. The spatial nature in the past user behavior and also the information about the user social interaction with other users, provide a richer background to build a more accurate and expressive recommendation model. Although there have been extensive studies on recommender systems working with user-item ratings, GPS trajectories, and other types of data, there are very few approaches that exploit the unique properties of the LBSN user check-in data. In this thesis, effective and efficient algorithms that create recommendations are proposed based on such properties. / published_or_final_version / Computer Science / Doctoral / Doctor of Philosophy

The Burg spectral estimator (BSE) exhibits better peak resolution than conventional linear spectral estimators, particularly for short data records. Based on this property, the quality of the BSE is investigated with the available data record segmented and the relevant parameters or functions associated with each segment averaged. Averaging of autoregressive coefficients, reflection coefficients, or spectral density functions is used with the BSE and the corresponding performances are studied. Approximate expressions for the mean and variance of these modified Burg spectral estimators are derived. Lower bounds for the mean and variance of reflection coefficients are also deduced. Finally, the variance of the estimation errors associated with the modified power spectral density estimators is compared against the theoretical Cramer-Rao lower bound. / M.S.

Thesis (D. Tech. (Engineering: Electric)) -- Central University of technology, 2012 / Because power generation relies heavily on electricity demand, consumers are required to wisely manage their loads to consolidate the power utility‟s optimal power generation efforts. Consequently, accurate and reliable electric load forecasting systems are required. Prior to the present situation, there were various forecasting models developed primarily for electric load forecasting. Modelling short term load forecasting using artificial neural networks has recently been proposed by researchers. This project developed a model for short term load forecasting using a neural network. The concept was tested by evaluating the forecasting potential of the basic feedforward and the cascade forward neural network models. The test results showed that the cascade forward model is more efficient for this forecasting investigation. The final model is intended to be a basis for a real forecasting application. The neural model was tested using actual load data of the Bloemfontein reticulation network to predict its load for half an hour in advance. The cascade forward network demonstrates a mean absolute percentage error of less than 5% when tested using four years of utility data. In addition to reporting the summary statistics of the mean absolute percentage error, an alternate method using correlation coefficients for presenting load forecasting performance results are shown. This research proposes that a 6:1:1 cascade forward neural network can be trained with data from a month of a year and forecast the load for the same month of the following year. This research presents a new time series modeling for short term load forecasting, which can model the forecast of the half-hourly loads of weekdays, as well as of weekends and public holidays. Obtained results from extensive testing on the Bloemfontein power system network confirm the validity of the developed forecasting approach. This model can be implemented for on-line testing application to adopt a final view of its usefulness.