Abstract

Abstract
Meta-learning is increasingly used to support the recommendation of machine learning algorithms and their configurations. These recommendations are made based on meta-data, consisting of performance evaluations of algorithms and characterizations on prior datasets. These characterizations, also called meta-features, describe properties of the data which are predictive for the performance of machine learning algorithms trained on them. Unfortunately, despite being used in many studies, meta-features are not uniformly described, organized and computed, making many empirical studies irreproducible and hard to compare. This paper aims to deal with this by systematizing and standardizing data characterization measures for classification datasets used in meta-learning. Moreover, it presents an extensive list of meta-features and characterization tools, which can be used as a guide for new practitioners. By identifying particularities and subtle issues related to the characterization measures, this survey points out possible future directions that the development of meta-features for meta-learning can assume.

Abstract
Current research on imbalanced data recognises that class imbalance is aggravated by other data intrinsic characteristics, among which class overlap stands out as one of the most harmful. The combination of these two problems creates a new and difficult scenario for classification tasks and has been discussed in several research works over the past two decades. In this paper, we argue that despite some insightful information can be derived from related research, the joint-effect of class overlap and imbalance is still not fully understood, and advocate for the need to move towards a unified view of the class overlap problem in imbalanced domains. To that end, we start by performing a thorough analysis of existing literature on the joint-effect of class imbalance and overlap, elaborating on important details left undiscussed on the original papers, namely the impact of data domains with different characteristics and the behaviour of classifiers with distinct learning biases. This leads to the hypothesis that class overlap comprises multiple representations, which are important to accurately measure and analyse in order to provide a full characterisation of the problem. Accordingly, we devise two novel taxonomies, one for class overlap measures and the other for class overlap-based approaches, both resonating with the distinct representations of class overlap identified. This paper therefore presents a global and unique view on the joint-effect of class imbalance and overlap, from precursor work to recent developments in the field. It meticulously discusses some concepts taken as implicit in previous research, explores new perspectives in light of the limitations found, and presents new ideas that will hopefully inspire researchers to move towards a unified view on the problem and the development of suitable strategies for imbalanced and overlapped domains.

Abstract

Abstract
Density estimation is an important tool for data analysis. Non-parametric approaches have a reputation for offering state-of-the-art density estimates limited to few dimensions. Despite providing less accurate density estimates, histogram-based approaches remain the only alternative for datasets in high-dimensional spaces. In this paper, we present a multivariate histogram approach to estimate the density of a dataset without restrictions on the number of dimensions, containing both numerical and categorical variables (without numerical encoding) and allowing missing data (without the need to preprocess them). Results from the empirical evaluation show that it is possible to estimate the density of datasets without restrictions on dimensionality, and the method is robust to missing values and categorical variables.

Abstract
Time series forecasting is one of the most active research topics. Machine learning methods have been increasingly adopted to solve these predictive tasks. However, in a recent work, evidence was shown that these approaches systematically present a lower predictive performance relative to simple statistical methods. In this work, we counter these results. We show that these are only valid under an extremely low sample size. Using a learning curve method, our results suggest that machine learning methods improve their relative predictive performance as the sample size grows. The R code to reproduce all of our experiments is available at https://github.com/vcerqueira/MLforForecasting.