Catálogo de publicaciones - libros

The Feature Selection problem involves discovering a subset of features such that a classifier built only with this subset would have better predictive accuracy than a classifier built from the entire set of features. Ensemble methods, such as Bagging and Boosting, have been shown to increase the performance of classifiers to remarkable levels but surprisingly have not been tried in other parts of the classification process. In this paper, we apply the ensemble approach to feature selection by proposing a systematic way of combining various outcomes of a feature selection algorithm. The proposed framework, named STochFS, have been shown empirically to improve the performance of well-known feature selection algorithms.

Logistic Model Trees have been shown to be very accurate and compact classifiers [8]. Their greatest disadvantage is the computational complexity of inducing the logistic regression models in the tree. We address this issue by using the AIC criterion [1] instead of cross-validation to prevent overfitting these models. In addition, a weight trimming heuristic is used which produces a significant speedup. We compare the training time and accuracy of the new induction process with the original one on various datasets and show that the training time often decreases while the classification accuracy diminishes only slightly.

In applications such as fraud and intrusion detection, it is of great interest to measure the evolving trends in the data. We consider the problem of quantifying changes between two datasets with class labels. Traditionally, changes are often measured by first estimating the probability distributions of the given data, and then computing the distance, for instance, the K-L divergence, between the estimated distributions. However, this approach is computationally infeasible for large, high dimensional datasets. The problem becomes more challenging in the streaming data environment, as the high speed makes it difficult for the learning process to keep up with the concept drifts in the data. To tackle this problem, we propose a method to quantify concept drifts using a universal model that incurs minimal learning cost. In addition, our model also provides the ability of performing classification.

Most approaches of Class Association Rule (CAR) based classification have not intensively addressed the classification of instances including numeric attributes. In this paper, a levelwise subspace clustering method deriving hyper-rectangular clusters is proposed to efficiently provide quantitative, interpretative and accurate CARs.

This paper presents an efficient algorithm for maintaining the generator representation in dynamic datasets. The generators representation is a kind of lossless, concise representation of the set of frequent itemsets. Furthermore, the algorithm utilizes a novel optimization based on generators borders for the first time in the literature. Generators borders are the borderline between frequent generators and other itemsets. New frequent generators can be generated through monitoring them. Experiments show that our algorithm is more efficient than previous solutions.

This work presents a technique that integrates the heuristics tabu search, simulated annealing, genetic algorithms and backpropagation. This approach obtained promising results in the simultaneous optimization of the artificial neural network architecture and weights.