Abstract
Critical health events represent a relevant cause of mortality in intensive care units of hospitals, and their timely prediction has been gaining increasing attention. This problem is an instance of the more general predictive task of early anomaly detection in time series data. One of the most common approaches to solve this problem is to use standard classification methods. In this paper we propose a novel method that uses a layered learning architecture to solve early anomaly detection problems. One key contribution of our work is the idea of pre-conditional events, which denote arbitrary but computable relaxed versions of the event of interest. We leverage this idea to break the original problem into two layers, which we hypothesize are easier to solve. Focusing on critical health episodes, the results suggest that the proposed approach is advantageous relative to state of the art approaches for early anomaly detection. Although we focus on a particular case study, the proposed method is generalizable to other domains. © Springer Nature Switzerland AG 2019.

Abstract
Social networks have an evolving characteristic due to the continuous interaction between users, with nodes associating and disassociating with each other as time flies. The analysis of such networks is especially challenging, because it needs to be performed with an online approach, under the one-pass constraint of data streams. Such evolving behavior leads to changes in the network topology that can be investigated under different perspectives. In this work we focus on the analysis of nodes position evolution—a node-centric perspective. Our goal is to spot change-points in an evolving network at which a node deviates from its normal behavior. Therefore, we propose a change detection model for processing evolving network streams which employs three different aggregating mechanisms for tracking the evolution of centrality metrics of a node. Our model is space and time efficient with memory less mechanisms and in other mechanisms at most we require the network of current time step T only. Additionally, we also compare the influence on different centralities’ fluctuations by the dynamics of real-world preferences. Consecutively, we apply our model in the user preference change detection task, reaching competitive levels of accuracy on Twitter network. © 2019, Springer International Publishing AG, part of Springer Nature.

Abstract
The literature on gait analysis in Vascular Parkinsonism (VaP), addressing issues such as variability, foot clearance patterns, and the effect of levodopa, is scarce. This study investigates whether spatiotemporal, foot clearance and stride-to-stride variability analysis can discriminate VaP, and responsiveness to levodopa. Fifteen healthy subjects, 15 Idiopathic Parkinson's Disease (IPD) patients and 15 VaP patients, were assessed in two phases: before (Off-state), and one hour after (On-state) the acute administration of a suprathreshold (1.5 times the usual) levodopa dose. Participants were asked to walk a 30-meter continuous course at a self-selected walking speed while wearing foot-worn inertial sensors. For each gait variable, mean, coefficient of variation (CV), and standard deviations SDI and SD2 obtained by Poincare analysis were calculated. General linear models (GLMs) were used to identify group differences. Patients were subject to neuropsychological evaluation (MoCA test) and Brain MRI. VaP patients presented lower mean stride velocity, stride length, lift-off and strike angle, and height of maximum toe (later swing) (p < .05), and higher %gait cycle in double support, with only the latter unresponsive to levodopa. VaP patients also presented higher CV, significantly reduced after levodopa. Yet, all VaP versus IPD differences lost significance when accounting for mean stride length as a covariate. In conclusion, VaP patients presented a unique gait with reduced degrees of foot clearance, probably correlated to vascular lesioning in dopaminergic/non-dopaminergic cortical and subcortical non-dopaminergic networks, still amenable to benefit from levodopa. The dependency of gait and foot clearance and variability deficits from stride length deserves future clarification.

Abstract

Abstract
In this paper, we present the BeamSouL sequence miner that finds sequences of logical atoms. This algorithm uses a levelwise hybrid search strategy to find a subset of contrasting logical sequences available in a SeqLog database. The hybrid search strategy runs an exhaustive search, in the first phase, followed by a beam search strategy. In the beam search phase, the algorithm uses the confidence metric to select the top k sequential patterns that will be specialized in the next level. Moreover, we develop a first-order logic classification framework that uses predicate invention technique to include the BeamSouL findings in the learning process. We evaluate the performance of our proposals using four multi-relational databases. The results are promising, and the BeamSouL algorithm can be more than one order of magnitude faster than the baseline and can find long and highly discriminative contrasting sequences.

Abstract
Portfolio optimization in stock markets has been investigated by many researchers. It looks for a subset of assets able to maintain a good trade-off control between risk and return. Several algorithms have been proposed to portfolio management. These algorithms use known return and correlation data to build subset of recommended assets. Dynamic stock correlation networks, whose vertices represent stocks and edges represent the correlation between them, can also be used as input by these algorithms. This study proposes the definition of constants of the classical mean-variance analysis using machine learning and weighted link prediction in stock networks (method named as MLink). To assess the performance of MLink, experiments were performed using real data from the Brazilian Stock Exchange. In these experiments, MLink was compared with mean-variance analysis (MVA), a popular method to portfolio optimization. According to the experimental results, using weighted link prediction in stock networks as input considerably increases the performance in portfolio optimization task, resulting in a gross capital increase of 41% in 84 days.