Catálogo de publicaciones - libros

Local methods have significant advantages when the probability measure defined on the space of symbolic objects for each class is very complex, but can still be described by a collection of less complex local approximations. We propose a technique of local bagging of decision stumps. We performed a comparison with other well known combining methods using the same base learner, on standard benchmark datasets and the accuracy of the proposed technique was greater in most cases.

Resistance spot welding is an important and widely used method for joining metal objects. In this paper, various classification methods for identifying welding processes are evaluated. Using process identification, a similar process for a new welding experiment can be found among the previously run processes, and the process parameters leading to high-quality welding joints can be applied. With this approach, good welding results can be obtained right from the beginning, and the time needed for the set-up of a new process can be substantially reduced. In addition, previous quality control methods can also be used for the new process. Different classifiers are tested with several data sets consisting of statistical and geometrical features extracted from current and voltage signals recorded during welding. The best feature set - classifier combination for the data used in this study is selected. Finally, it is concluded that welding processes can be identified almost perfectly by certain features.

Support Vector Machines constitute a powerful Machine Learning technique originally designed for the solution of 2-class problems. In multiclass applications, many works divide the whole problem in multiple binary subtasks, whose results are then combined. This paper introduces a new framework for multiclass Support Vector Machines generation from binary predictors. Minimum Spanning Trees are used in the obtainment of a hierarchy of binary classifiers composing the multiclass solution. Different criteria were tested in the tree design and the results obtained evidence the efficiency of the proposed approach, which is able to produce good hierarchical multiclass solutions in polynomial time.

Ship detection in high frequency ground wave radar can be approached by one-class classifier where ship echoes are regarded as abnormal situations to typical ocean clutter. In this paper we consider the problems of feature extraction and representation problems. We first study characters of ocean clutter and ship echo, and find that initial frequency and chirp rate are two proper features to tell difference between ship echoes and ocean clutters. However to lower the probability of misjudging, we represent data examples in a combined similarity-dissimilarity space other than using these two features directly. A hypersphere with minimal volume is adopted to bound training examples, from which an efficient one-class classifier is established upon limited number of typical examples. The comparison result to a one-class classifier based on original feature representation is given.

Understandability of decision trees depends on two key factors: the size of the trees and the complexity of their node functions. Most of the attempts to improve the behavior of decision trees have been focused only on reducing their sizes by building the trees on complex features. These features are usually linear or non-linear functions of all the original attributes. In this paper, reduced complexity features are proposed as a way to reduce the size of decision trees while keeping understandable functions at their nodes. The proposed approach is tested on a robot grasping application where the goal is to obtain a system able to classify grasps as valid or invalid and also on three datasets from the UCI repository.

The aim of this paper is to explore the effectiveness of Bayesian classifiers in intrusion detection (ID). Specifically, we provide an experimental study that focuses on comparing the accuracy of different classification models showing that the Bayesian classification approach is reasonably effective and efficient in predicting attacks and in exploiting the knowledge required by a computational intelligent ID process.

Association rules discovery is a fundamental task in spatial data mining where data are naturally described at multiple levels of granularity. ARES is a spatial data mining system that takes advantage from this taxonomic knowledge on spatial data to mine multi-level spatial association rules. A large amount of rules is typically discovered even from small set of spatial data. In this paper we present a graph-based visualization that supports data miners in the analysis of multi-level spatial association rules discovered by ARES and takes advantage from hierarchies describing the same spatial object at multiple levels of granularity. An application on real-world spatial data is reported. Results show that the use of the proposed visualization technique is beneficial.

We propose a selection of knowledge technologies to support decisions of the management of public health care in Slovenia, and present a specific application in one region (Celje). First, we exploit data mining and statistical techniques to analyse databases that are regularly collected for the national Institute of Public Health. Next, we study organizational aspects of public health resources in the Celje region with the objective to identify the areas that are atypical in terms of availability and accessibility of the public health services for the population. The most important step is the detection of outliers and the analysis of the causes for availability and accessibility deviations. The results can be used for high-level health-care planning and decision-making.

Subsequence matching is one of the most important issues in the field of data mining. The existing subsequence matching algorithms use windows of the fixed size to construct only one index. The algorithms have a problem that their performance gets worse as the difference between the query sequence length and the window size increases. In this paper, we propose a new subsequence matching method based on index interpolation, which is a technique that constructs the indexes for multiple window sizes and chooses an index most appropriate for a given query sequence for subsequence matching.We first examine the performance change due to the window size effect through preliminary experiments, and devise a cost function for subsequence matching that reflects the distribution of query sequence lengths in the view point of physical database design. Next, we propose a new subsequence matching method to improve search performance, and present an algorithm based on the cost function to construct the multiple indexes to maximize the performance. Finally, we verify the superiority of the proposed method through a series of experiments using the real and the synthetic data sequences.

Conceptual models play an important part in a variety of domains, especially in meteorological applications. This paper proposes a novel conceptual modeling approach based on a two-phase spatial data mining and knowledge discovery method, aiming to model the concepts of the evolvement trends of Mesoscale Convective Clouds (MCCs) over the Tibetan Plateau with derivation rules and environmental physical models. Experimental results show that the proposed conceptual model to much extent simplifies and improves the weather forecasting techniques on heavy rainfalls and floods in South China.