Abstract
The number of Internet of Things devices generating data streams is expected to grow exponentially with the support of emergent technologies such as 5G networks. Therefore, the online processing of these data streams requires the design and development of suitable machine learning algorithms, able to learn online, as data is generated. Like their batch-learning counterparts, stream-based learning algorithms require careful hyperparameter settings. However, this problem is exacerbated in online learning settings, especially with the occurrence of concept drifts, which frequently require the reconfiguration of hyperparameters. In this article, we present SSPT, an extension of the Self Parameter Tuning (SPT) optimisation algorithm for data streams. We apply the Nelder-Mead algorithm to dynamically-sized samples, converging to optimal settings in a single pass over data while using a relatively small number of hyperparameter configurations. In addition, our proposal automatically readjusts hyperparameters when concept drift occurs. To assess the effectiveness of SSPT, the algorithm is evaluated with three different machine learning problems: recommendation, regression, and classification. Experiments with well-known data sets show that the proposed algorithm can outperform previous hyperparameter tuning efforts by human experts. Results also show that SSPT converges significantly faster and presents at least similar accuracy when compared with the previous double-pass version of the SPT algorithm.

Abstract

Abstract
The present work aims to compare several load disaggregation methods. While the supervised alternative was found to be the most competent, the semi-supervised is proved to be close in terms of potential, while the unsupervised alternative seems insufficient. By the same token, the tests with long-lasting data prove beneficial to confirm the long-term performance since no significant loss of performance is noticed with the scalar of the time-horizon. Finally, the patchwork of new parametrization and methodology fine-tuning also proves interesting for improving global performance in several methods.

Abstract
One of the most significant challenges for machine learning nowadays is the discovery of causal relationships from data. This causal discovery is commonly performed using Bayesian like algorithms. However, more recently, more and more causal discovery algorithms have appeared that do not fall into this category. In this paper, we present a new algorithm that explores global causal association rules with Uncertainty Coefficient. Our algorithm, CRPA-UC, is a global structure discovery approach that combines the advantages of association mining with causal discovery and can be applied to binary and non-binary discrete data. This approach was compared to the PC algorithm using several well-known data sets, using several metrics.

Abstract
This work aims to develop a Machine Learning framework to predict voting behaviour. Data resulted from longitudinally collected variables during the Portuguese 2019 general election campaign. Naive Bayes (NB), and Tree Augmented Naive Bayes (TAN) and three different expert models using Dynamic Bayesian Networks (DBN) predict voting behaviour systematically for each moment in time considered using past information. Even though the differences found in some performance comparisons are not statistically significant, TAN and NB outperformed DBN experts' models. The learned models outperformed one of the experts' models when predicting abstention and two when predicting right-wing parties vote. Specifically, for the right-wing parties vote, TAN and NB presented satisfactory accuracy, while the experts' models were below 50% in the third evaluation moment.

Abstract
Numerous machine learning applications involve dealing with imbalanced domains, where the learning focus is on the least frequent classes. This imbalance introduces new challenges for both the performance assessment of these models and their predictive modeling. While several performance metrics have been established as baselines in balanced domains, some cannot be applied to the imbalanced case since the use of the majority class in the metric could lead to a misleading evaluation of performance. Other metrics, such as the area under the precision-recall curve, have been demonstrated to be more appropriate for imbalance domains due to their focus on class-specific performance. There are, however, many proposed implementations for this particular metric, which could potentially lead to different conclusions depending on the one used. In this research, we carry out an experimental study to better understand these issues and aim at providing a set of recommendations by studying the impact of using different metrics and different implementations of the same metric under multiple imbalance settings.