Different performance measures are used to assess the behaviour, and to carry out the comparison, of classifiers in Machine Learning. Many measures have been defined on the literature, and among them, a measure inspired by Shannon’s entropy named the Confusion Entropy (CEN). In this work we introduce a new measure, MCEN, by modifying CEN to avoid its unwanted behaviour in the binary case, that disables it as a suitable performance measure in classification. We compare MCEN with CEN and other performance measures, presenting analytical results in some particularly interesting cases, as well as some heuristic computational experimentation.

We develop a predictive prognosis model to support medical experts in their clinical decision-making process in Intensive Care Units (ICUs) (a) to enhance early mortality prediction, (b) to make more efficient medical decisions about patients at higher risk, and (c) to evaluate the effectiveness of new treatments or detect changes in clinical practice. It is a machine learning hierarchical model based on Bayesian classifiers built from some recorded features of a real-world ICU cohort, to bring about the assessment of the risk of mortality, also predicting destination at ICU discharge if the patient survives, or the cause of death otherwise, constructed as an ensemble of five base Bayesian classifiers by using the average ensemble criterion with weights, and we name it the Ensemble Weighted Average (EWA). We compare EWA against other state-of-the-art machine learning predictive models. Our results show that EWA outperforms its competitors, presenting in addition the advantage over the ensemble using the majority vote criterion of allowing to associate a confidence level to the provided predictions. We also prove the convenience of locally recalibrate from data the standard model used to predict the mortality risk based on the APACHE II score, although as a predictive model it is weaker than the other.

In a network environment composed of different types of computing centers that can be divided into different layers (clod, edge layer, and others), the interconnection between them offers the possibility of peer-to-peer task offloading. For many resource-constrained devices, the computation of many types of tasks is not feasible because they cannot support such computations as they do not have enough available memory and processing capacity. In this scenario, it is worth considering transferring these tasks to resource-rich platforms, such as Edge Data Centers or remote cloud servers. For different reasons, it is more exciting and appropriate to download various tasks to specific download destinations depending on the properties and state of the environment and the nature of the functions. At the same time, establishing an optimal offloading policy, which ensures that all tasks are executed within the required latency and avoids excessive workload on specific computing centers is not easy. This study presents two alternatives to solve the offloading decision paradigm by introducing two well-known algorithms, Graph Neural Networks (GNN) and Deep Q-Network (DQN). It applies the alternatives on a well-known Edge Computing simulator called PureEdgeSim and compares them with the two default methods, Trade-Off and Round Robin. Experiments showed that variants offer a slight improvement in task success rate and workload distribution. In terms of energy efficiency, they provided similar results. Finally, the success rates of different computing centers are tested, and the lack of capacity of remote cloud servers to respond to applications in real-time is demonstrated. These novel ways of finding a download strategy in a local networking environment are unique as they emulate the state and structure of the environment innovatively, considering the quality of its connections and constant updates. The download score defined in this research is a crucial feature for determining the quality of a download path in the GNN training process and has not previously been proposed. Simultaneously, the suitability of Reinforcement Learning (RL) techniques is demonstrated due to the dynamism of the network environment, considering all the key factors that affect the decision to offload a given task, including the actual state of all devices.

This study presents a model for predicting photovoltaic power generation based on meteorological, temporal and geographical variables, without using irradiance values, which have traditionally posed challenges and difficulties for accurate predictions. Validation methods and evaluation metrics are used to analyse four different approaches that vary in the distribution of the training and test database, and whether or not location-independent modelling is performed. The coefficient of determination, $R^2$, is used to measure the proportion of variation in photovoltaic power generation that can be explained by the model’s variables, while $gC{O}_{2}eq$ represents the amount of CO${}_2$ emissions equivalent to each unit of power generation. Both are used to compare model performance and environmental impact. The results show significant differences between the locations, with substantial improvements in some cases, while in others improvements are limited. The importance of customising the predictive model for each specific location is emphasised. Furthermore, it is concluded that environmental impact studies in model production are an additional step towards the creation of more sustainable and efficient models. Likewise, this research considers both the accuracy of solar energy predictions and the environmental impact of the computational resources used in the process, thereby promoting the responsible and sustainable progress of data science.

Due to the limitations that falls have on humans, early detection of these becomes essential to avoid further damage. In many applications, various technologies are used to acquire accurate information from individuals such as wearable sensors, environmental sensors or cameras, but all of these require high computational resources in many cases, delaying the response of the entire system. The complexity of the models used to process the input data and detect these activities makes them almost impossible to complete on devices with limited resources, which are the ones that could offer an immediate response avoiding unnecessary communications between sensors and centralized computing centers. In this work, we chose to reduce the models to detect falls using images as input data. We proceeded to use image sequences as video frames, using data from two open source datasets, and we applied the Sparse Low Rank Method to reduce certain layers of the Convolutional Neural Networks that were the backbone of the models. Additionally, we chose to replace a convolutional block with Long Short Term Memory to consider the latest updates of these data sequences. The results showed that performance was maintained decently while significantly reducing the parameter size of the resulting models.