José A. Sáez

Late survival of Neanderthals at the southernmost extreme of Europe

The late survival of archaic hominin populations and their long contemporaneity with modern humans is now clear for southeast Asia. In Europe the extinction of the Neanderthals, firmly associated with Mousterian technology, has received much attention, and evidence of their survival after 35 kyr bp has recently been put in doubt. Here we present data, based on a high-resolution record of human occupation from Gorham’s Cave, Gibraltar, that establish the survival of a population of Neanderthals to 28 kyr bp. These Neanderthals survived in the southernmost point of Europe, within a particular physiographic context, and are the last currently recorded anywhere. Our results show that the Neanderthals survived in isolated refuges well after the arrival of modern humans in Europe.

A Survey of Discretization Techniques: Taxonomy and Empirical Analysis in Supervised Learning

Discretization is an essential preprocessing technique used in many knowledge discovery and data mining tasks. Its main goal is to transform a set of continuous attributes into discrete ones, by associating categorical values to intervals and thus transforming quantitative data into qualitative data. In this manner, symbolic data mining algorithms can be applied over continuous data and the representation of information is simplified, making it more concise and specific. The literature provides numerous proposals of discretization and some attempts to categorize them into a taxonomy can be found. However, in previous papers, there is a lack of consensus in the definition of the properties and no formal categorization has been established yet, which may be confusing for practitioners. Furthermore, only a small set of discretizers have been widely considered, while many other methods have gone unnoticed. With the intention of alleviating these problems, this paper provides a survey of discretization methods proposed in the literature from a theoretical and empirical perspective. From the theoretical perspective, we develop a taxonomy based on the main properties pointed out in previous research, unifying the notation and including all the known methods up to date. Empirically, we conduct an experimental study in supervised classification involving the most representative and newest discretizers, different types of classifiers, and a large number of data sets. The results of their performances measured in terms of accuracy, number of intervals, and inconsistency have been verified by means of nonparametric statistical tests. Additionally, a set of discretizers are highlighted as the best performing ones.

Study on the Impact of Partition-Induced Dataset Shift on

Cross-validation is a very commonly employed technique used to evaluate classifier performance. However, it can potentially introduce dataset shift, a harmful factor that is often not taken into account and can result in inaccurate performance estimation. This paper analyzes the prevalence and impact of partition-induced covariate shift on different k-fold cross-validation schemes. From the experimental results obtained, we conclude that the degree of partition-induced covariate shift depends on the cross-validation scheme considered. In this way, worse schemes may harm the correctness of a single-classifier performance estimation and also increase the needed number of repetitions of cross-validation to reach a stable performance estimation.

k

Noisy and borderline examples in imbalanced datasets harm classifier performance.Our proposal reduces the noise and makes the class boundaries more regular.Our proposal performs better than other re-sampling methods in this scenario.Ensemble-based noise filters work well when dealing with noise.The iterative noise elimination is a good approach to deal with noisy datasets. Classification datasets often have an unequal class distribution among their examples. This problem is known as imbalanced classification. The Synthetic Minority Over-sampling Technique (SMOTE) is one of the most well-know data pre-processing methods to cope with it and to balance the different number of examples of each class. However, as recent works claim, class imbalance is not a problem in itself and performance degradation is also associated with other factors related to the distribution of the data. One of these is the presence of noisy and borderline examples, the latter lying in the areas surrounding class boundaries. Certain intrinsic limitations of SMOTE can aggravate the problem produced by these types of examples and current generalizations of SMOTE are not correctly adapted to their treatment.This paper proposes the extension of SMOTE through a new element, an iterative ensemble-based noise filter called Iterative-Partitioning Filter (IPF), which can overcome the problems produced by noisy and borderline examples in imbalanced datasets. This extension results in SMOTE-IPF. The properties of this proposal are discussed in a comprehensive experimental study. It is compared against a basic SMOTE and its most well-known generalizations. The experiments are carried out both on a set of synthetic datasets with different levels of noise and shapes of borderline examples as well as real-world datasets. Furthermore, the impact of introducing additional different types and levels of noise into these real-world data is studied. The results show that the new proposal performs better than existing SMOTE generalizations for all these different scenarios. The analysis of these results also helps to identify the characteristics of IPF which differentiate it from other filtering approaches.

-Fold Cross-Validation

The presence of noise in data is a common problem that produces several negative consequences in classification problems. In multi-class problems, these consequences are aggravated in terms of accuracy, building time, and complexity of the classifiers. In these cases, an interesting approach to reduce the effect of noise is to decompose the problem into several binary subproblems, reducing the complexity and, consequently, dividing the effects caused by noise into each of these subproblems. This paper analyzes the usage of decomposition strategies, and more specifically the One-vs-One scheme, to deal with noisy multi-class datasets. In order to investigate whether the decomposition is able to reduce the effect of noise or not, a large number of datasets are created introducing different levels and types of noise, as suggested in the literature. Several well-known classification algorithms, with or without decomposition, are trained on them in order to check when decomposition is advantageous. The results obtained show that methods using the One-vs-One strategy lead to better performances and more robust classifiers when dealing with noisy data, especially with the most disruptive noise schemes.

SMOTE-IPF

Classifier performance, particularly of instance-based learners such as k-nearest neighbors, is affected by the presence of noisy data. Noise filters are traditionally employed to remove these corrupted data and improve the classification performance. However, their efficacy depends on the properties of the data, which can be analyzed by what are known as data complexity measures. This paper studies the relation between the complexity metrics of a dataset and the efficacy of several noise filters to improve the performance of the nearest neighbor classifier. A methodology is proposed to extract a rule set based on data complexity measures that enables one to predict in advance whether the use of noise filters will be statistically profitable. The results obtained show that noise filtering efficacy is to a great extent dependent on the characteristics of the data analyzed by the measures. The validation process carried out shows that the final rule set provided is fairly accurate in predicting the efficacy of noise filters before their application and it produces an improvement with respect to the indiscriminate usage of noise filters.

Analyzing the presence of noise in multi-class problems: alleviating its influence with the One-vs-One decomposition

Canonical machine learning algorithms assume that the number of objects in the considered classes are roughly similar. However, in many real-life situations the distribution of examples is skewed since the examples of some of the classes appear much more frequently. This poses a difficulty to learning algorithms, as they will be biased towards the majority classes. In recent years many solutions have been proposed to tackle imbalanced classification, yet they mainly concentrate on binary scenarios. Multi-class imbalanced problems are far more difficult as the relationships between the classes are no longer straightforward. Additionally, one should analyze not only the imbalance ratio but also the characteristics of the objects within each class. In this paper we present a study on oversampling for multi-class imbalanced datasets that focuses on the analysis of the class characteristics. We detect subsets of specific examples in each class and fix the oversampling for each of them independently. Thus, we are able to use information about the class structure and boost the more difficult and important objects. We carry an extensive experimental analysis, which is backed-up with statistical analysis, in order to check when the preprocessing of some types of examples within a class may improve the indiscriminate preprocessing of all the examples in all the classes. The results obtained show that oversampling concrete types of examples may lead to a significant improvement over standard multi-class preprocessing that do not consider the importance of example types. Graphical abstractDisplay Omitted HighlightsA thorough analysis of oversampling for handling multi-class imbalanced datasets.Proposition to detect underlying structures and example types in considered classes.Smart oversampling based on extracted knowledge about imbalance distribution types.In-depth insight into the importance of selecting proper examples for oversampling.Guidelines that allow to design efficient classifiers for multi-class imbalanced data.

Predicting noise filtering efficacy with data complexity measures for nearest neighbor classification

Traditional classifier learning algorithms build a unique classifier from the training data. Noisy data may deteriorate the performance of this classifier depending on the degree of sensitiveness to data corruptions of the learning method. In the literature, it is widely claimed that building several classifiers from noisy training data and combining their predictions is an interesting method of overcoming the individual problems produced by noise in each classifier. This statement is usually not supported by thorough empirical studies considering problems with different types and levels of noise. Furthermore, in noisy environments, the noise robustness of the methods can be more important than the performance results themselves and, therefore, it must be carefully studied. This paper aims to reach conclusions on such aspects focusing on the analysis of the behavior, in terms of performance and robustness, of several Multiple Classifier Systems against their individual classifiers when these are trained with noisy data. In order to accomplish this study, several classification algorithms, of varying noise robustness, will be chosen and compared with respect to their combination on a large collection of noisy datasets. The results obtained show that the success of the Multiple Classifier Systems trained with noisy data depends on the individual classifiers chosen, the decisions combination method and the type and level of noise present in the dataset, but also on the way of creating diversity to build the final system. In most of the cases, they are able to outperform all their single classification algorithms in terms of global performance, even though their robustness results will depend on the way of introducing diversity into the Multiple Classifier System.

Analyzing the oversampling of different classes and types of examples in multi-class imbalanced datasets

Semi-supervised classification methods have received much attention as suitable tools to tackle training sets with large amounts of unlabeled data and a small quantity of labeled data. Several semi-supervised learning models have been proposed with different assumptions about the characteristics of the input data. Among them, the self-training process has emerged as a simple and effective technique, which does not require any specific hypotheses about the training data. Despite its effectiveness, the self-training algorithm usually make erroneous predictions, mainly at the initial stages, if noisy examples are labeled and incorporated into the training set. Noise filters are commonly used to remove corrupted data in standard classification. In 2005, Li and Zhou proposed the addition of a statistical filter to the self-training process. Nevertheless, in this approach, filtering methods have to deal with a reduced number of labeled instances and the erroneous predictions it may induce. In this work, we analyze the integration of a wide variety of noise filters into the self-training process to distinguish the most relevant features of filters. We will focus on the nearest neighbor rule as a base classifier and ten different noise filters. We provide an extensive analysis of the performance of these filters considering different ratios of labeled data. The results are contrasted with nonparametric statistical tests that allow us to identify relevant filters, and their main characteristics, in the field of semi-supervised learning.

Tackling the problem of classification with noisy data using Multiple Classifier Systems: Analysis of the performance and robustness

Noise is common in any real-world data set and may adversely affect classifiers built under the effect of such type of disturbance. Some of these classifiers are widely recognized for their good performance when dealing with imperfect data. However, the noise robustness of the classifiers is an important issue in noisy environments and it must be carefully studied. Both performance and robustness are two independent concepts that are usually considered separately, but the conclusions reached with one of these metrics do not necessarily imply the same conclusions with the other. Therefore, involving both concepts seems to be crucial in order determine the expected behavior of the classifiers against noise. Existing measures fail to properly integrate these two concepts, and they are also not well suited to compare different techniques over the same data. This paper proposes a new measure to establish the expected behavior of a classifier with noisy data trying to minimize the problems of considering performance and robustness individually: the Equalized Loss of Accuracy (ELA). The advantages of ELA against other robustness metrics are studied and all of them are also compared. Both the analysis of the distinct measures and the empirical results show that ELA is able to overcome the aforementioned problems that the rest of the robustness metrics may produce, having a better behavior when comparing different classifiers over the same data set.