The thesis is a contribution to case-based reasoning (CBR). It rests on the observation that experts who express their knowledge in cases are inclined to state significant amounts of it only as informal arguments, which can be called “folk arguments”. These arguments are undervalued in CBR, sometimes not used at all. This work provides a means of capturing and representing such information in basically the natural form in which an expert states it, as an enhancement for traditional case-based knowledge. The aim is to exploit collections of facts and folk expert arguments in order to improve the quality of results obtained by CBR computations. In the novel CBR framework proposed, reasoning templates from knowledge engineering methodologies are offered as a systematic means of collecting and representing these arguments in cases – the “ArgCases framework”. Exploitation of procedures of numerical taxonomy in the investigation of case similarities and organisation of the case bases where facts and folk arguments are included then leads to a “Taxonomic ArgCases framework”. These contributions are validated in two applications where expert behaviour is primarily about reasoning on cases: allocation of frequencies for reliable reception of shortwave radio broadcasting, and the authentication (dating) of paintings. In both applications, case bases are constructed from information about expert analysis of problems, as captured from past records and with the help of new reasoning templates. Through proposal and inspection of taxonomies involving cases, it is shown how collections of factual and folk argumentation characteristics can be indexed in order to support the answering of alternative forms of query in CBR. The contributions of the thesis demonstrate how both numerical taxonomy and reasoning templates can be exploited within that area of artificial intelligence. In addition to these, its major contribution is to make a place for “folk arguments” within CBR.

Knowledge discovery in databases is the field of computer science that, combines different computational, statistical and mathematical techniques, in order to create systems that support the process of finding new knowledge from databases. This thesis investigates adaptability and reusability techniques, namely abstraction and structure tactics, applicable to software solutions in the field of Knowledge Discovery in Databases. The research is driven by two business problems in operational loss specifically fraud and system failure. Artificial Intelligence (AI)1 techniques are increasingly being applied to complex and dynamic business process. Business processes that require analytical processing of large volumes of data, highly changing and complex domain knowledge driven data analysis, knowledge management and knowledge discovery are examples of problems that would typically be addressed using AI techniques. To control the business, data and software complexity, the industry has responded with a wide variety of products that have specific software architectures, user environments and include the latest AI trends. Specific fields of research like knowledge discovery in databases (KDD) [1][2] have been created in order to address the different challenges of supporting the discovery of new knowledge from raw data using AI related techniques (e.g. data mining). Regardless of all this academic and commercial effort, solutions for specific business processes are suffering from adaptability, flexibility and reusability limitations. The solutions‟ software architecture and user interfacing environments are not adaptable and flexible enough to cope with business process changes or the need to reuse accumulated knowledge (i.e. prior analyses). Consequently the life time of some of these solutions is reduced severely or increasing efforts are required to keep them running. This research is driven by a specific business domain and it is conducted in two phases. The first phase focuses on a single intelligent and analytical system solution and aims to capture specific problem domain requirements that drive the definition of a business domain specific KDD reference architecture. Through a case study a detailed analysis of the semantics of fraud detection is done and the elements, components and services of an intelligent and analytics fraud detection system are investigated. The second phase takes the architectural observations from phase I, to the more generic and wide KDD challenges, defines an operational loss domain model, a reference architecture and tests its reuse in a different type of operational loss business problem. Software related KDD challenges are revised and addressed in the reference architecture. A second application is analysed through a second case study and it is used to test the architecture and refine it. This application is in the domain of detection and prevention of operational loss due to data related system failure, The software architectures defined in the different phases of this research are analyzed using the Architecture Trade off Analysis Method (ATAM)2 [3] in order to evaluate risks and compare their adaptability, flexibility and reusability properties. This thesis has the following contributions: It constitutes one of the first investigations of adaptability and reusability in business domain specific KDD software architecture from an abstraction and structure viewpoint. It defines the TRANSLATIONAL architectural style for high data volume and intensive data analysis systems that supports the balancing of flexibility, reusability and performance. Using the TRANSLATIONAL architectural style, it defines and implements OL-KDA, a reference architecture that can be applied to problems in operational loss, namely fraud and data related system failure, and supports the complexity and dynamicity challenges. Developed and implemented a method for supporting data, dataflow and rules in KDD pre-processing and post-processing tasks. It defines a data manipulation and maintenance model that favours performance and adaptability in specific KDD tasks. Two substantial case studies where developed and analysed in order to understand and subsequently test the defined techniques and reference architecture in business domains. 1 AI: Artificial Intelligence techniques are used in computer science to mimic or use aspects of human behavior within information systems. 2 ATAM: Architecture analysis method that focuses on analyzing quality attributes and use cases.

This thesis primarily investigates lossy compression of an inverted index. Two approaches of lossy compression are studied in detail, i.e. (i) term frequency quantization, and (ii) document pruning. In addition, a technique for document pruning, i.e. the entropy-based method, is applied to re-rank retrieved documents as query-independent knowledge. Based on the quantization theory, we examine how the number of quantization levels for coding the term frequencies affects retrieval performance. Three methods are then proposed for the purpose of reducing the quantization distortion, including (i) a non-uniform quantizer; (ii) an iterative technique; and (iii) term-specific quantizers. Experiments based on standard TREC test sets demonstrate that nearly no degradation of retrieval performance can be achieved by allocating only 2 or 3 bits for the quantized term frequency values. This is comparable to lossless coding techniques such as unary, γ and δ-codes. Furthermore, if lossless coding is applied to the quantized term frequency values, then around 26% (or 12%) savings can be achieved over lossless coding alone, with less than 2.5% (or no measurable) degradation in retrieval performance. Prior work on index pruning considered posting pruning and term pruning. In this thesis, an alternative pruning approach, i.e. document pruning, is investigated, in which unimportant documents are removed from the document collection. Four algorithms for scoring document importance are described, two of which are dependent on the score function of the retrieval system, while the other two are independent of the retrieval system. Experimental results suggest that document pruning is comparable to existing pruning approaches, such as posting pruning. Note that document pruning affects the global statistics of the indexed collection. We therefore examine whether retrieval performance is superior based on statistics derived from the full or the pruned collection. Our results indicate that keeping statistics derived from the full collection performs slightly better. Document pruning scores documents and then discards those that fall outside a threshold. An alternative is to re-rank documents based on these scores. The entropy-based score, which is independent of the retrieval system, provides a query-independent knowledge of document specificity, analogous to PageRank. We investigate the utility of document specificity in the context of Intranet search, where hypertext information is sparse or absent. Our results are comparable to the previous algorithm that induced a graph link structure based on the measure of similarity between documents. However, a further analysis indicates that our method is superior on computational complexity.

This thesis investigates the forensic detection of re-quantization and re-sampling that often occurs when multimedia content has been tampered with. The detection is based on statistical classification techniques, including Fisher linear discriminant (FLD) and support vector machine (SVM). Successively compressing an image with different quality factors is a typical example of uniform re-quantization. In the first part of the thesis, we investigate the forensic detection of uniform re-quantization for images. Three features are introduced based on the observation that uniform re-quantization (i) introduces discontinuities in the signal histogram and (ii) induces periodic artifacts. After validating the discriminative potential of these features with synthetic Gaussian signals, we propose a system to detect JPEG re-compression. Both linear (i.e. FLD) and non-linear (i.e. SVM) classifications are examined for comparative purposes. Experimental results clearly demonstrate the ability of the proposed features to detect JPEG re-compression, as well as their competitiveness compared to prior approaches to achieve the same goal. Successively compressing a speech signal with different speech encodings is a typical example of non-uniform re-quantization. In the second part of the thesis, we investigate the forensic detection of non-uniform re-quantization for speech signals. Two detection algorithms, based on the non-periodic histogram artifacts present in the time-domain and DFT-domain respectively, are compared with each other. Comparative experiments indicate that both detection algorithms produce reliable results with high area under the curve (AUC) values for a set of different experimental scenarios. In general, the time-domain detection performs slightly better than the DFT-domain detection. However, the latter is superior in the less dimensionality of input vectors to the FLD classifier being used. Re-sizing an image is a typical example of re-sampling. In the third part of the thesis, we investigate the forensic detection of re-sampling for images. A new method is proposed to detect re-sampled imagery. The method is based on examining the normalized energy density present within windows of varying size in the second derivative of the image in the frequency domain, and exploiting this characteristic to derive a 19-dimensional feature vector that is used to train a SVM classifier. Comparison with prior work reveals that the proposed algorithm performs similarly for re-sampling rates greater than 1, and is superior to prior work for re-sampling rates less than 1. Experiments are performed for both bilinear and bicubic interpolations and qualitatively similar results are observed for each. Results are also provided for the detection of re-sampled imagery after noise corruption and JPEG compression. As expected, some degradation in performance is observed as the noise increases or the JPEG quality factor declines.

Computer performance is suffering greatly from diminishing returns as the increasing cost of implementing complex hardware optimizations and of increasing clock frequency no longer yields the gains in computational ability and power efficiency consumers demand. Notable products including a generation of Intel Pentium 4 processors have been cancelled as a result. This sudden hiccup in an historically predictable performance road map has inspired research and industrial communities to investigate architectures, some rather unorthodox, that complete work more quickly and more efficiently. One such computer architecture under development, Fleet, exposes fine-grain instruction level concurrency, addresses the growing costs of on-chip communications, and promotes simplicity in the underlying hardware design. This one-instruction computer transports data using simple move operations. The globally-asynchronous architecture promotes high modularity allowing specialized configurations of the architecture to be generated quickly with low hardware and software complexity. The Armada architecture presented in this thesis expands on Fleet by introducing constructs that exploit thread-level concurrency. The proposals herein aim to increase the performance efficiency of Fleet and other communicationcentric architectures. Trade-offs between software and hardware complexity and between the static and dynamic division of labor are investigated through the implementation and study of an Armada microarchitecture and an Armada compiler created for this research. This thesis explores the merits and pitfalls of this unique architecture as the basis for general-purpose computers for the future.

No description available.

No description available.

No description available.

No description available.

No description available.