Imbalanced neural newsvendor

Over the past 20 years, the Internet saw an exponential grown of traffic, users, services and applications. Currently, it is estimated that the Internet is used everyday by more than 3.6 billions users, who generate 20 TB of traffic per second. Such a huge amount of data challenge network managers and analysts to understand how the network is performing, how users are accessing resources, how to properly control and manage the infrastructure, and how to detect possible threats. Along with mathematical, statistical, and set theory methodologies machine learning and big data approaches have emerged to build systems that aim at automatically extracting information from the raw data that the network monitoring infrastructures offer. In this thesis I will address different network monitoring solutions, evaluating several methodologies and scenarios. I will show how following a common workflow, it is possible to exploit mathematical, statistical, set theory, and machine learning methodologies to extract meaningful information from the raw data. Particular attention will be given to machine learning and big data methodologies such as DBSCAN, and the Apache Spark big data framework. The results show that despite being able to take advantage of mathematical, statistical, and set theory tools to characterize a problem, machine learning methodologies are very useful to discover hidden information about the raw data. Using DBSCAN clustering algorithm, I will show how to use YouLighter, an unsupervised methodology to group caches serving YouTube traffic into edge-nodes, and latter by using the notion of Pattern Dissimilarity, how to identify changes in their usage over time. By using YouLighter over 10-month long races, I will pinpoint sudden changes in the YouTube edge-nodes usage, changes that also impair the end users' Quality of Experience. I will also apply DBSCAN in the deployment of SeLINA, a self-tuning tool implemented in the Apache Spark big data framework to autonomously extract knowledge from network traffic measurements. By using SeLINA, I will show how to automatically detect the changes of the YouTube CDN previously highlighted by YouLighter. Along with these machine learning studies, I will show how to use mathematical and set theory methodologies to investigate the browsing habits of Internauts. By using a two weeks dataset, I will show how over this period, the Internauts continue discovering new websites. Moreover, I will show that by using only DNS information to build a profile, it is hard to build a reliable profiler. Instead, by exploiting mathematical and statistical tools, I will show how to characterize Anycast-enabled CDNs (A-CDNs). I will show that A-CDNs are widely used either for stateless and stateful services. That A-CDNs are quite popular, as, more than 50% of web users contact an A-CDN every day. And that, stateful services, can benefit of A-CDNs, since their paths are very stable over time, as demonstrated by the presence of only a few anomalies in their Round Trip Time. Finally, I will conclude by showing how I used BGPStream an open-source software framework for the analysis of both historical and real-time Border Gateway Protocol (BGP) measurement data. By using BGPStream in real-time mode I will show how I detected a Multiple Origin AS (MOAS) event, and how I studies the black-holing community propagation, showing the effect of this community in the network. Then, by using BGPStream in historical mode, and the Apache Spark big data framework over 16 years of data, I will show different results such as the continuous growth of IPv4 prefixes, and the growth of MOAS events over time. All these studies have the aim of showing how monitoring is a fundamental task in different scenarios. In particular, highlighting the importance of machine learning and of big data methodologies.

Analyses of the effectiveness of infectious disease control interventions often rely on dynamic transmission models to simulate intervention effects. We aim to understand how the choice of network or compartmental model can influence estimates of intervention effectiveness in the short and long term for an endemic disease with susceptible and infected states in which infection, once contracted, is lifelong. We consider 4 disease models with different permutations of socially connected network versus unstructured contact (mass-action mixing) model and heterogeneous versus homogeneous disease risk. The models have susceptible and infected populations calibrated to the same long-term equilibrium disease prevalence. We consider a simple intervention with varying levels of coverage and efficacy that reduces transmission probabilities. We measure the rate of prevalence decline over the first 365 d after the intervention, long-term equilibrium prevalence, and long-term effective reproduction ratio at equilibrium. Prevalence declined up to 10% faster in homogeneous risk models than heterogeneous risk models. When the disease was not eradicated, the long-term equilibrium disease prevalence was higher in mass-action mixing models than in network models by 40% or more. This difference in long-term equilibrium prevalence between network versus mass-action mixing models was greater than that of heterogeneous versus homogeneous risk models (less than 30%); network models tended to have higher effective reproduction ratios than mass-action mixing models for given combinations of intervention coverage and efficacy. For interventions with high efficacy and coverage, mass-action mixing models could provide a sufficient estimate of effectiveness, whereas for interventions with low efficacy and coverage, or interventions in which outcomes are measured over short time horizons, predictions from network and mass-action models diverge, highlighting the importance of sensitivity analyses on model structure. • We calibrate 4 models-socially connected network versus unstructured contact (mass-action mixing) model and heterogeneous versus homogeneous disease risk-to 10% preintervention disease prevalence.• We measure the short- and long-term intervention effectiveness of all models using the rate of prevalence decline, long-term equilibrium disease prevalence, and effective reproduction ratio.• Generally, in the short term, prevalence declined faster in the homogeneous risk models than in the heterogeneous risk models.• Generally, in the long term, equilibrium disease prevalence was higher in the mass-action mixing models than in the network models, and the effective reproduction ratio was higher in network models than in the mass-action mixing models.

The utility of advanced mud gas (AMG) data has been limited to fluid typing and petrophysical correlations. There is the need to extend the utility to real-time reservoir characterization prior to wireline logging and geological core description. Our first attempt to predict reservoir rock porosity within a well yielded good result. This study improves on the previous effort by utilizing big data obtained from combining various wells in the study area. We used machine learning (ML) methodology in the absence of established physical relationship between AMG data, comprising light and heavy flare gas components, and reservoir rock porosity. More than 20,000 data points collected from representative wells were used to prove the concept of predicting the porosity in an interval or section of any well within the study area. Optimized models of artificial neural network (ANN), decision trees (DT) and random forest (RF) were applied to the combined dataset. The combined dataset was randomly split into training and validation subsets in 70:30 ratio. The 30% validation subset simulates a missing well interval or section. Comparing the results of the ANN, DT, and RF models using statistical model performance evaluation metrics, the RF model outperformed the others. The RF model gave a training and validation correlation coefficient (R-Squared) values of 0.94 and 0.83 respectively compared to 0.36 and 0.35 for the ANN and 0.84 and 0.73 for the DT models respectively. However, the p-value and mean errors show that the models are statistically acceptable. Having showed in a previous research that a multivariate linear regression model could not handle the complexity in the relationship between porosity and the flare gas components, these results have further confirmed the robustness of nonlinear solutions based on the ML methodology. We conclude that the ML approach to predicting reservoir rock porosity from advanced mud gas data is feasible and better results are achievable with more research. This study has confirmed the feasibility of predicting porosity based on a dataset of combined wells and the huge benefit in extending the utility of AMG data beyond the traditional workflows. This approach is capable of complementing existing reservoir characterization processes in assessing reservoir quality at the early stage of exploration. Future work will investigate the impact of integrating the AMG with surface drilling parameters to possibly increase the prediction accuracy.

Polymer informatics: Expert-in-the-loop in QSPR modeling of refractive index

Simple SummaryIn vitro gas production systems are regularly utilized to screen feed ingredients for inclusion in ruminant diets. However, not all in vitro systems are set up to measure methane (CH4) production, nor do all papers report in vitro CH4. Therefore, the objective of this study was to develop models to predict in vitro production of CH4, a greenhouse gas produced by ruminants, from in vitro gas and volatile fatty acid (VFA) production data, and to identify the major drivers of CH4 production in these systems. Meta-analysis and machine learning (ML) methodologies were applied to predict CH4 production from in vitro gas parameters. Meta-analysis results indicate that equations containing apparent dry matter (DM) digestibility, total VFA production, propionate, valerate and feed type (forage vs. concentrate) resulted in best prediction of CH4. The ML models far exceeded the predictability achieved using meta-analysis, but further evaluation on an external database would be required to assess their generalization capacity. The models developed can be utilized to estimate CH4 emissions in vitro.In vitro gas production systems are utilized to screen feed ingredients for inclusion in ruminant diets. However, not all in vitro systems are set up to measure methane (CH4) production, nor do all publications report in vitro CH4. Therefore, the objective of this study was to develop models to predict in vitro CH4 production from total gas and volatile fatty acid (VFA) production data and to identify the major drivers of CH4 production in these systems. Meta-analysis and machine learning (ML) methodologies were applied to a database of 354 data points from 11 studies to predict CH4 production from total gas production, apparent DM digestibility (DMD), final pH, feed type (forage or concentrate), and acetate, propionate, butyrate and valerate production. Model evaluation was performed on an internal dataset of 107 data points. Meta-analysis results indicate that equations containing DMD, total VFA production, propionate, feed type and valerate resulted in best predictability of CH4 on the internal evaluation dataset. The ML models far exceeded the predictability achieved using meta-analysis, but further evaluation on an external database would be required to assess generalization ability on unrelated data. Between the ML methodologies assessed, artificial neural networks and support vector regression resulted in very similar predictability, but differed in fitting, as assessed by behaviour analysis. The models developed can be utilized to estimate CH4 emissions in vitro.

In the construction practice, at the stage of planning of a construction project, planners are trying to take into account the possibility of unfavorable situations and their consequences during the project execution. Therefore, planning decisions should use appropriate tools for uncertainty modeling as well as consider alternative options for the implementation of the entire undertaking or the most sensitive (critical) works. Taking into account the fact that the management and planning of construction projects is carried out in conditions of non-stochastic uncertainty, the authors propose to create an alternative model with a fuzzy decision node based on classical network models. This approach allows to model activities that are alternative to each other. An important methodological element is the fuzzy assessment of the considered alternatives with the use of fuzzy set theory. The proposed alternative network model with a fuzzy decision node gives the possibility to conduct a comprehensive analysis with the possibility of taking into account and modeling the uncertainty of input data in an easy way. The numerical example shows the possibility of quick selection of the best variant based on the adopted assessment criteria and allows to serialize work variants, which gives additional information in the alternative activities of the network model in the form of preferences. The combination of a decision model in the form of a decision node with a network model can be useful in both strategic and tactical planning of the implementation of construction projects.

Two characters, scale-free and hierarchical structure, are shown in lots of real systems all over the world. In this article, firstly, we build up a class of network models with hierarchical structure, denoted by N h (t), and prove it to be scale-free. By computing average degrees, we find these network models are all sparse. Secondly, based on their special topological structure, we research analytically average distance of each model belonging to N h (t) and present an iteration computational method of average distance corresponding to every one in N h (t). Finally, we discuss the affect by changing some tunable parameters to average distance in the set N h (t), and compare our hierarchical network models with previous ones, and show some meaning problems for gaining more universal methods of computing average distance in hierarchical networks.

The prediction of sediment levels in combined sewer system (CSS) would result in enormous savings in resources for their maintenance as a reduced number of inspections would be needed. In this paper, we benchmark different machine learning (ML) methodologies to improve the maintenance schedules of the sewerage and reduce the number of cleanings using historical sediment level and inspection data of the combined sewer system in the city of Barcelona. Two ML methodologies involve the use of spatial features for sediment prediction at critical sections of the sewer, where the cost of maintenance is high because of the dangerous access; one uses a regression model to predict the sediment level of a section, and the other one a binary classification model to identify whether or not a section needs cleaning. The last ML methodology is a short-term forecast of the possible sediment level in future days to improve the ability of operators to react and solve an imminent sediment level increase. Our study concludes with three different models. The spatial and short-term regression methodologies accomplished the best results with Artificial Neural Networks (ANN) with 0.76 and 0.61 R2 scores, respectively. The classification methodology resulted in a Gradient Boosting (GB) model with an accuracy score of 0.88 and an area under the curve (AUC) of 0.909.

In this paper, we use a unified loss function, called the soft insensitive loss function, for Bayesian support vector regression. We follow standard Gaussian processes for regression to set up the Bayesian framework, in which the unified loss function is used in the likelihood evaluation. Under this framework, the maximum a posteriori estimate of the function values corresponds to the solution of an extended support vector regression problem. The overall approach has the merits of support vector regression such as convex quadratic programming and sparsity in solution representation. It also has the advantages of Bayesian methods for model adaptation and error bars of its predictions. Experimental results on simulated and real-world data sets indicate that the approach works well even on large data sets.

Selecting a small set of informative features from a large number of possibly noisy candidates is a challenging problem with many applications in machine learning and approximate Bayesian computation. In practice, the cost of computing informative features also needs to be considered. This is particularly important for networks because the computational costs of individual features can span several orders of magnitude. We addressed this issue for the network model selection problem using two approaches. First, we adapted nine feature selection methods to account for the cost of features. We show for two classes of network models that the cost can be reduced by two orders of magnitude without considerably affecting classification accuracy (proportion of correctly identified models). Second, we selected features using pilot simulations with smaller networks. This approach reduced the computational cost by a factor of 50 without affecting classification accuracy. To demonstrate the utility of our approach, we applied it to three different yeast protein interaction networks and identified the best-fitting duplication divergence model. Supplementary materials, including computer code to reproduce our results, are available online.

Machine learning (ML), a subset of artificial intelligence (AI) centered on machines learning from extensive datasets, stands at the forefront of a technological revolution shaping various facets of society. Cardiovascular medicine has emerged as a key domain for ML applications, with considerable efforts to integrate these innovations into routine clinical practice. Within cardiac electrophysiology, ML applications, especially in the automated interpretation of electrocardiograms, have garnered substantial attention in existing literature. However, less recognized are the diverse applications of ML in cardiac electrophysiology and arrhythmias, spanning basic science research on arrhythmia mechanisms, both experimental and computational, as well as contributions to enhanced techniques for mapping cardiac electrical function and translational research related to arrhythmia management. This comprehensive review delves into various ML applications within the scope of this journal, organized into 3 parts. The first section provides a fundamental understanding of general ML principles and methodologies, serving as a foundational resource for readers interested in exploring ML applications in arrhythmia research. The second part offers an in-depth review of studies in arrhythmia and electrophysiology that leverage ML methodologies, showcasing the broad potential of ML approaches. Each subject is thoroughly outlined, accompanied by a review of notable ML research advancements. Finally, the review delves into the primary challenges and future perspectives surrounding ML-driven cardiac electrophysiology and arrhythmias research.

Anonymization techniques have tranquilized current social network users in terms of privacy leakage, however, it does not radically prevent adversaries from de-anonymizing users, as they may map the users to an un-anonymized network. Till now, researchers share a common thread in such de-anonymization attack: unveiling conditions leading to successful de-anonymization under the chosen network model. However, it has not yet been well understand how the structural property in different network models intrinsically determines de-anonymizability. We address the above issue in this paper by making the two contributions: (i) We discover that the automorphic degree and homomorphic degree of social networks determine their de-anonymizability universally. The automorphic degree characterizes the distinguishability of the users in a network, while the homomorphic degree models the similarities of users between two networks. We conclude that a smaller automorphic degree and a larger homomorphic degree conduce to a higher de-anonymizability. Such model-independent phenomenon refreshes us with a latitudinal study as it generalizes the essential commonness of de-anonymization in different network models. (ii) We derive explicit parametric bounds of the de-anonymizability for three classic network models, showing that such bounds correspond well to our conclusion about morphism property. We then algorithmically and experimentally show that such theoretical results literally make sense to adversaries. Such longitudinal study, including welding theory, algorithm and validation, promises applicability of our results on morphism property in real cases.

Modelling & Simulation (M&S) and Machine Learning (ML) methodologies have undergone significant advancements, enabling transformative applications across various industries. The integration of M&S and ML into a Hybrid M&S-ML approach leverages the unique strengths of both fields, offering enhanced model precision, improved efficiency, and more effective decision support. This review explores the increasing convergence of ML algorithms with traditional M&S methods- namely Agent-Based Modelling & Simulation, Discrete Event Simulation, and System Dynamics- in healthcare applications. Through a systematic review of 90 relevant studies, this article provides a comprehensive synthesis of the current state-of-the-art Hybrid M&S-ML in healthcare. Specifically, it examines the M&S and ML methodologies employed, associated software tools and programming languages, analyses integration patterns and data exchange mechanisms, and explores application domains, as well as the types and motivations for hybridisation. Key findings highlight prominent methodological and technical trends, as well as opportunities for combining M&S with ML to address healthcare challenges. These insights provide direction for modellers and data scientists in developing hybrid M&S-ML approaches that more effectively combine simulation capabilities with data-driven learning. The review also demonstrates the potential of such approaches to advance methodological innovation and support evidence-based decision-making in diverse healthcare contexts.

The integration of machine learning (ML) in veterinary epidemiology offers transformative potential for data analysis and disease management, a significant shift from traditional statistical methods. This review explores the burgeoning role of ML, emphasizing its capacity to handle complex, high-dimensional data and uncover nonlinear relationships, which are pivotal in epidemiology. Key ML methodologies, including supervised, unsupervised, and reinforcement learning, provide robust frameworks for predictive modeling, pattern recognition, and decision-making processes. Applications in veterinary medicine are already evident in diagnostic imaging and animal behavior monitoring, showcasing ML's ability to enhance diagnostic accuracy and welfare monitoring.Despite these advancements, the field faces challenges such as imbalanced datasets, data quality issues, and the need for interdisciplinary collaboration. Strategies like Synthetic Minority Over-sampling Technique and ensemble methods help address class imbalance, while robust preprocessing techniques mitigate data noise. Future advancements in natural language processing and reinforced learning promise further integration, optimizing disease surveillance and intervention strategies.The review highlights the transformative potential of ML in veterinary epidemiology, advocating for continued research and collaboration to overcome existing hurdles. By leveraging ML's capabilities, veterinary professionals can improve disease prediction, develop targeted preventive programs, and enhance overall animal health and food security, marking a significant advancement in veterinary science.

This paper offers a comprehensive overview of machine learning (ML) methodologies and algorithms, highlighting their practical applications in the critical domain of water resource management. Environmental issues, such as climate change and ecosystem destruction, pose significant threats to humanity and the planet. Addressing these challenges necessitates sustainable resource management and increased efficiency. Artificial intelligence (AI) and ML technologies present promising solutions in this regard. By harnessing AI and ML, we can collect and analyze vast amounts of data from diverse sources, such as remote sensing, smart sensors, and social media. This enables real-time monitoring and decision making in water resource management. AI applications, including irrigation optimization, water quality monitoring, flood forecasting, and water demand forecasting, enhance agricultural practices, water distribution models, and decision making in desalination plants. Furthermore, AI facilitates data integration, supports decision-making processes, and enhances overall water management sustainability. However, the wider adoption of AI in water resource management faces challenges, such as data heterogeneity, stakeholder education, and high costs. To provide an overview of ML applications in water resource management, this research focuses on core fundamentals, major applications (prediction, clustering, and reinforcement learning), and ongoing issues to offer new insights. More specifically, after the in-depth illustration of the ML algorithmic taxonomy, we provide a comparative mapping of all ML methodologies to specific water management tasks. At the same time, we include a tabulation of such research works along with some concrete, yet compact, descriptions of their objectives at hand. By leveraging ML tools, we can develop sustainable water resource management plans and address the world’s water supply concerns effectively.