Catálogo de publicaciones - libros

FP-growth algorithm is an efficient algorithm for mining frequent patterns. It scans database only twice and does not need to generate and test the candidate sets that is quite time consuming. The efficiency of the FP-growth algorithm outperforms previously developed algorithms. But, it must recursively generate huge number of conditional FP-trees that requires much more memory and costs more time.

In this paper, we present an algorithm, CFPmine, that is inspired by several previous works. CFPmine algorithm combines several advantages of existing techniques. One is using constrained subtrees of a compact FP-tree to mine frequent pattern, so that it is doesn’t need to construct conditional FP-trees in the mining process. Second is using an array-based technique to reduce the traverse time to the CFP-tree. And an unified memeory management is also implemented in the algorithm. The experimental evaluation shows that CFPmine algorithm is a high performance algorithm. It outperforms Apriori, Eclat and FP-growth and requires less memory than FP-growth.

The ultimate aim of computer science is to assist humans with reasoning, decision-making and chores for their profession/living everyday complex problem solving. In other words, the target state of human-machine symbiosis is characterized by using a machine as a personal intellectual assistant. A principal impediment consists in the multi-disciplinary nature of the profession/living human activity. A customer, as a rule, is not a specialist in all related domains. This implies that an achievement of the posed problem is a providing a personal intellectual assistant with the multi-disciplinary knowledge. This paper deals with an approach to the problem via Web extension that contains Total Human Experience (THE Web).

Although Case-based Reasoning has been applied successfully in medical domains, case-based diagnosis handling multiple disorders is often not sufficient while multiple disorders is a daily problem in medical diagnosis and treatment. In this paper, we present an approach which integrates two case-based reasoners for diagnosing multiple faults. This multiple case-based reasoning approach has been evaluated on a medical case base taken from real world application and demonstrated to be very promising.

An intelligent knowledge-based recommendation system for multiple users in TV-Anytime environment is developed. The architecture of the multi-agent recommendation system is described. KQC (Keyword-query combination) user profile is proposed for information modeling. The priority filtering agent operates according to the similarity between the program and the KQC user profile. This knowledge-based recommendation system can provide personalized content to users based on their preferences more efficiently and more concisely.

Formal Concept Analysis (FCA), which is based on ordered lattice theory, is applied to mine association rules from web logs. The discovered knowledge (association rules) can then be used for online applications such as web recommendation and personalization. Experiments showed that FCA generated 60% fewer rules than Apriori, and the rules are comparable in quality according to three objective measures.

Oil well drilling is a complex process which frequently is leading to operational problems. The process generates huge amounts of data. In order to deal with the complexity of the problems addressed, and the large number of parameters involved, our approach extends a pure case-based reasoning method with reasoning within a model of general domain knowledge. The general knowledge makes the system less vulnerable for syntactical variations that do not reflect semantically differences, by serving as explanatory supportfor case retrieval and reuse. A tool, called TrollCreek, has been developed. It is shown how the combined reasoning method enables focused decision support for fault diagnosis and prediction of potential unwanted events in this domain.

There have been widespread applications for Multi Objective Genetic Algorithm (MOGA) on highly complicated optimization tasks in discontinuous, multi-modal, and noisy domains. Because the convergence of MOGA can be reached with the non-dominated set approximating the Pareto Optimal front, it is very important to construct the non-dominated set of MOGA efficiently. This paper proposes a new method called Dealer’s Principle to construct non-dominated sets of MOGA, and the time complexity is analyzed. Then we design a new MOGA with the Dealer’s Principle and a clustering algorithm based on the core distance of clusters to keep the diversity of solutions. We show that our algorithm is more efficient than the previous algorithms, and that it produces a wide variety of solutions. We also discuss the convergence and the diversity of our MOGA in experiments with benchmark optimization problems of three objectives.

Recently, multi-instance classification algorithm BP-MIP and multi-instance regression algorithm BP-MIR both based on neural networks have been proposed. In this paper, neural network ensemble techniques are introduced to solve multi-instance learning problems, where BP-MIP ensemble and BP-MIR ensemble are constructed respectively. Experiments on benchmark and artificial data sets show that ensembles of multi-instance neural networks are superior to single multi-instance neural networks in solving multi-instance problems.

This paper proposes a new feature selection method for text categorization. In this method, word tendency, which takes related words into consideration, is used to select best terms. Our experiments on binary classification tasks show that our method achieves better than DF and IG when the classes are semantically discriminative. Furthermore, our best performance is usually achieved in fewer features.

In this paper, we present a new support vector neural network inference system (SVNNIS) for regression estimation. The structure of the proposed SVNNIS can be obtained similar to that in the support vector regression (SVR), while the output of the SVNNIS is unbiased compared with the SVR and the weights can be updated by the recursive least square method with forgetting factor. The advantage of this system is its good generalization capability. The simulation result illustrates the effectiveness of the proposed SVNNIS.