An Ensemble Generation Method Based on Instance Hardness

Abstract

In Machine Learning, ensemble methods have been receiving a great deal of attention. Techniques such as Bagging and Boosting have been successfully applied to a variety of problems. Nevertheless, such techniques are still susceptible to the effects of noise and outliers in the training data. We propose a new method for the generation of pools of classifiers based on Bagging, in which the probability of an instance being selected during the resampling process is inversely proportional to its instance hardness, which can be understood as the likelihood of an instance being misclassified, regardless of the choice of classifier. The goal of the proposed method is to remove noisy data without sacrificing the hard instances which are likely to be found on class boundaries. We evaluate the performance of the method in nineteen public data sets, and compare it to the performance of the Bagging and Random Subspace algorithms. Our experiments show that in high noise scenarios the accuracy of our method is significantly better than that of Bagging.

A white-box analysis on the writer-independent dichotomy transformation applied to offline handwritten signature verification

“…For instance, in the paper by Cruz et al (2017), the authors used IH to identify the scenarios where an ensemble with dynamic selection techniques outperform the K-NN classifier. The IH has also been used in ensemble generation methods (Walmsley et al, 2018;Souza et al, 2018a;Kabir et al, 2018).…”

An investigation into the effects of label noise on Dynamic Selection algorithms

“…Instead of a single DSEL set, filtered by ENN, we propose combining the output of several classifier ensembles, each using a bootstrapped Dynamic Selection set based on the original DSEL, using Instance Hardness as a means to determine the selection probability of each instance in the DSEL. As in (WALMSLEY et al, 2018), we chose this probabilistic filtering approach for its ability to filter out the noisy data, while still allowing for the retention of hard instances in the class borders, with non-zero probability. The results for our proposal are compared not only against the method of Cruz et al, but also against the traditional Overall Local Accuracy, Local Class Accuracy (WOODS; KEGELMEYER; BOWYER, 1997), K-Nearest Oracles-Eliminate and K-Nearest Oracles-Union (KO; SABOURIN; BRITTO, 2008) algorithms.…”

“…Nevertheless, we must first define what we mean by noise. Here, we adopt the definition used in (WALMSLEY et al, 2018), in which noise is considered as any process that changes the true value of a variable from its original value.…”

Tackling Virtual and Real Concept Drifts: An Adaptive Gaussian Mixture Model Approach

“…These points may impair the performance of the GMM over time. To overcome them, we use k-Disagreeing Neighbors (kDN) [24], defined in Eq. 7, as a noise filter.…”