Nonlinear Machine Learning Models Research Articles

Symphony in the Latent Space: Provably Integrating High-Dimensional Techniques with Non-linear Machine Learning Models

This paper revisits building machine learning algorithms that involve interactions between entities, such as those between financial assets in an actively managed portfolio, or interactions between users in a social network. Our goal is to forecast the future evolution of ensembles of multivariate time series in such applications (e.g., the future return of a financial asset or the future popularity of a Twitter account). Designing ML algorithms for such systems requires addressing the challenges of high-dimensional interactions and non-linearity. Existing approaches usually adopt an ad-hoc approach to integrating high-dimensional techniques into non-linear models and recent studies have shown these approaches have questionable efficacy in time-evolving interacting systems. To this end, we propose a novel framework, which we dub as the additive influence model. Under our modeling assumption, we show that it is possible to decouple the learning of high-dimensional interactions from the learning of non-linear feature interactions. To learn the high-dimensional interactions, we leverage kernel-based techniques, with provable guarantees, to embed the entities in a low-dimensional latent space. To learn the non-linear feature-response interactions, we generalize prominent machine learning techniques, including designing a new statistically sound non-parametric method and an ensemble learning algorithm optimized for vector regressions. Extensive experiments on two common applications demonstrate that our new algorithms deliver significantly stronger forecasting power compared to standard and recently proposed methods.

Developing nonlinear k-nearest neighbors classification algorithms to identify patients at high risk of increased length of hospital stay following spine surgery.

Spondylolisthesis is a common operative disease in the United States, but robust predictive models for patient outcomes remain limited. The development of models that accurately predict postoperative outcomes would be useful to help identify patients at risk of complicated postoperative courses and determine appropriate healthcare and resource utilization for patients. As such, the purpose of this study was to develop k-nearest neighbors (KNN) classification algorithms to identify patients at increased risk for extended hospital length of stay (LOS) following neurosurgical intervention for spondylolisthesis. The Quality Outcomes Database (QOD) spondylolisthesis data set was queried for patients receiving either decompression alone or decompression plus fusion for degenerative spondylolisthesis. Preoperative and perioperative variables were queried, and Mann-Whitney U-tests were performed to identify which variables would be included in the machine learning models. Two KNN models were implemented (k = 25) with a standard training set of 60%, validation set of 20%, and testing set of 20%, one with arthrodesis status (model 1) and the other without (model 2). Feature scaling was implemented during the preprocessing stage to standardize the independent features. Of 608 enrolled patients, 544 met prespecified inclusion criteria. The mean age of all patients was 61.9 ± 12.1 years (± SD), and 309 (56.8%) patients were female. The model 1 KNN had an overall accuracy of 98.1%, sensitivity of 100%, specificity of 84.6%, positive predictive value (PPV) of 97.9%, and negative predictive value (NPV) of 100%. Additionally, a receiver operating characteristic (ROC) curve was plotted for model 1, showing an overall area under the curve (AUC) of 0.998. Model 2 had an overall accuracy of 99.1%, sensitivity of 100%, specificity of 92.3%, PPV of 99.0%, and NPV of 100%, with the same ROC AUC of 0.998. Overall, these findings demonstrate that nonlinear KNN machine learning models have incredibly high predictive value for LOS. Important predictor variables include diabetes, osteoporosis, socioeconomic quartile, duration of surgery, estimated blood loss during surgery, patient educational status, American Society of Anesthesiologists grade, BMI, insurance status, smoking status, sex, and age. These models may be considered for external validation by spine surgeons to aid in patient selection and management, resource utilization, and preoperative surgical planning.

Credit growth, the yield curve and financial crisis prediction: Evidence from a machine learning approach

We develop early warning models for financial crisis prediction applying machine learning techniques on macrofinancial data for 17 countries over 1870–2016. Most nonlinear machine learning models outperform logistic regression in out-of-sample predictions and forecasting. We identify economic drivers of our machine learning models by applying a novel framework based on Shapley values, uncovering nonlinear relationships between the predictors and crisis risk. Throughout, the most important predictors are credit growth and the slope of the yield curve, both domestically and globally. A flat or inverted yield curve is of most concern when nominal interest rates are low and credit growth is high.

Modelling the Employment Rate in Türkiye Using Machine Learning Methods

Employment rate is a crucial economic indicator that measures the percentage of the working-age population that is employed. Considering this importance, the employment rate of Türkiye is modelled in this work. The employment rate, gross domestic product, gross fixed capital formation, government expenditures, export and import data of the 1991-2021 period are taken from the World bank database and then the seasonal-trend decomposition of these data are performed in EViews software. After inspecting the nonlinearity of the data, a nonlinear machine learning model namely a feedforward artificial neural network model is developed in Python programming language for the modelling of the employment rate dependent on the gross domestic product, gross fixed capital formation, government expenditures, export and import data. The 70% of the available data is used as the training data and the developed feedforward artificial neural network is trained with a successful convergence despite the low number of samples thanks to its optimal structure. Then the remaining 30% of the available data are utilized as the test data. The actual employment rate data and the result of the developed feedforward neural network model are plotted on the same axis pair which show continuous overlap in a wide range. Apart from the visual inspection of the results of the developed model, the performance metrics are also calculated in Python programming language such as the mean absolute percentage error, coefficient of determination, root mean square error and the mean absolute error. The performance metrics of the results of the developed model also verify the high accuracy of the model. It is argued that the developed feedforward neural network model can also be used for the situations where low number of samples need to be modelled thanks to the optimal structure. Keywords: Employment rate, modelling, machine learning, artificial neural networks, estimation.

Using machine learning to realize genetic site screening and genomic prediction of productive traits in pigs.

Genomic prediction, which is based on solving linear mixed-model (LMM) equations, is the most popular method for predicting breeding values or phenotypic performance for economic traits in livestock. With the need to further improve the performance of genomic prediction, nonlinear methods have been considered as an alternative and promising approach. The excellent ability to predict phenotypes in animal husbandry has been demonstrated by machine learning (ML) approaches, which have been rapidly developed. To investigate the feasibility and reliability of implementing genomic prediction using nonlinear models, the performances of genomic predictions for pig productive traits using the linear genomic selection model and nonlinear machine learning models were compared. Then, to reduce the high-dimensional features of genome sequence data, different machine learning algorithms, including the random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost) and convolutional neural network (CNN) algorithms, were used to perform genomic feature selection as well as genomic prediction on reduced feature genome data. All of the analyses were processed on two real pig datasets: the published PIC pig dataset and a dataset comprising data from a national pig nucleus herd in Chifeng, North China. Overall, the accuracies of predicted phenotypic performance for traits T1, T2, T3 and T5 in the PIC dataset and average daily gain (ADG) in the Chifeng dataset were higher using the ML methods than the LMM method, while those for trait T4 in the PIC dataset and total number of piglets born (TNB) in the Chifeng dataset were slightly lower using the ML methods than the LMM method. Among all the different ML algorithms, SVM was the most appropriate for genomic prediction. For the genomic feature selection experiment, the most stable and most accurate results across different algorithms were achieved using XGBoost in combination with the SVM algorithm. Through feature selection, the number of genomic markers can be reduced to 1 in 20, while the predictive performance on some traits can even be improved compared to using the full genome data. Finally, we developed a new tool that can be used to execute combined XGBoost and SVM algorithms to realize genomic feature selection and phenotypic prediction.

Comparison of linear and non-linear machine learning models for time-dependent readmission or mortality prediction among hospitalized heart failure patients

Although many models are available to predict prognosis of heart failure patients, most tools combining survival analysis are based on proportional hazard model. Non-linear machine learning algorithms would overcome the limitation of the time-independent hazard ratio assumption and provide more information in readmission or mortality prediction among heart failure patients.The present study collected the clinical information of 1796 hospitalized heart failure patients surviving during hospitalization in a Chinese clinical center from December 2016 to June 2019. A traditional multivariate Cox regression model and three machine learning survival models were developed in derivation cohort. Uno's concordance index and integrated Brier score in validation cohort were calculated to evaluate the discrimination and calibration of different models. Time-dependent AUC and Brier score curves were plotted to assess the performance of models at different time phases.

Intelligent optimization for modelling superhydrophobic ceramic membrane oil flux and oil-water separation efficiency: Evidence from wastewater treatment and experimental laboratory

Due to the significant energy and economic losses brought on by the global oil spill, there has been an increased interest in oil-water separation. This study presents strong non-linear machine learning models (support vector regression (SVR) and Gaussian process regression (GPR)) with the Response surface method (RSM) to predict the oil flux and oil-water separation efficiency of wastewater using ceramic membrane technology. For the model development and prediction of oil flux (OF) and oil-water separation efficiency (OSE), oil concentration (mg/L), feed flow rate (mL/min), and pH were considered as input variables. The input variables are combined in three combinations to study the most contributing input features to the models’ performance. Mean square error (MSE) and Nash-Sutcliffe coefficient efficiency (NSE) were used to assess the prediction performances of the developed models with the different number of input combinations considered in the study. For the two target variables (OF and OSE), GPR and SVR models were used to separately predict them. For OF, the SVR-2 [Combo-2] model (MSE = 0.9255 and NSE = 2.7976) performed better with higher prediction accuracy compared to GPR-2 [Combo-2] model (MSE = 0.763 and NSE = 6.437). In addition, for OSE, the GPR-3 [Combo-3] model (MSE = 0.995 and NSE = 0.5544) performed slightly better than SVR-3 [Combo-3] model (MSE = 0.992 and NSE = 0.8066). The results showed that the SVR model with the combo-2 and GPR-3 models for OF and OSE variables are the proposed models with the best performance and accuracy. This machine learning study will aid in better evaluating the function of materials such as ceramic in membrane performance features such as oil flux and rejection prediction, separation efficiency, water recovery, membrane fouling, and so on. As for academics and manufacturers, this machine learning (ML) strategy will boost performance and allow a better understanding of system governance.

Development of a linear–nonlinear hybrid special model to predict monthly runoff in a catchment area and evaluate its performance with novel machine learning methods

Accurate forecasting of runoff as an important hydrological variable is a key task for water resources planning and management. Given the importance of this variable, in the current study, a multivariate linear stochastic model (MLSM) is combined with a multilayer nonlinear machine learning model (MNMLM) to generate a hybrid model for the spatial and temporal simulation of runoff in the Quebec basin, Canada. Monthly hydrological data from 2001 to 2013, including precipitation and runoff data from nine stations and Normalized Difference Vegetation Index (NDVI) extraction of MODIS data, are applied as input to the proposed hybrid model. At the first step of the hybrid modeling, data normality and stationary were examined by performing various tests. In the second step, MLSM was developed by defining four different scenarios and as a result 15 sub-scenarios. The first and second scenarios were developed based on one exogenous variable (precipitation or NDVI). In contrast, the second and third scenarios were developed based on two additional variables. In the first and third scenarios, the data are modeled without preprocessing. In the second and fourth scenarios, a preprocessing step is performed on the data. Then, in the third step, various combinations based on different time delays from runoff data were applied for developing nonlinear model. The comparisons are made between observed and simulated time series at various stations and based on the root mean squared error (RMSE), mean absolute error (MAE), correlation coefficient (R) and Akaike information criterion (AIC). The efficiency of the proposed hybrid model is compared with a novel machine learning model that was introduced in 2021 by Sultani et al., and it was also compared with the results obtained from the linear and nonlinear models. In most stations, delays (t-1) and (t-24) are identified as the most effective delays in hybrid and nonlinear modeling of runoff. Also, in most stations, the use of climatic parameters and physiographic factors as exogenous variables along with runoff data improves the results compared to the use of one variable. Results showed that at all stations, proposed hybrid model generally leads to more accurate estimates of runoff compared with various linear and nonlinear models. More accurate estimates of peak runoff values at all stations were another excellence of proposed hybrid model than other models.

Application of nested artificial neural network for the prediction of significant wave height

Significant wave height is the most important parameter in feasibility studies and the design of wave energy converters. In this study, the novel nested artificial neural networks were developed and applied to predict significant wave height at twenty selected locations of the North Sea. A nested artificial neural network applies nonlinear machine learning models as transfer functions in the neurons of networks. Two input parameters comprising wind speed and wind direction were implemented to train the derived models. The results show that the derived nonlinear machine learning models are about 18.39% more accurate than the linear regression technique. The statistical indices confirm that the nested artificial neural network may increases the accuracy of traditional models by up to 34%. Among all applied models, the nested artificial neural network developed based on the integration of particle swarm optimization algorithm and adaptive neuro-fuzzy inference system, with RMSE = 0.525m and R2 = 0.84, provides the most accurate prediction of wave heights. The high accuracy of the results indicates that if computational time is not a very critical factor for users, then the application of nested artificial neural networks may be recommended for the modeling of wave parameters and other complex problems.

Comparison of various chemometric methods on visible and near-infrared spectral analysis for wood density prediction among different tree species and geographical origins.

Visible and near-infrared (Vis-NIR) spectroscopy has been widely applied in many fields for the qualitative and quantitative analysis. Chemometric techniques including pre-processing, variable selection, and multivariate calibration models play an important role to better extract useful information from spectral data. In this study, a new de-noising method (lifting wavelet transform, LWT), four variable selection methods, as well as two non-linear machine learning models were simultaneously analyzed to compare the impact of chemometric approaches on wood density determination among various tree species and geographical locations. In addition, fruit fly optimization algorithm (FOA) and response surface methodology (RSM) were employed to optimize the parameters of generalized regression neural network (GRNN) and particle swarm optimization-support vector machine (PSO-SVM), respectively. As for various chemometric methods, the optimal chemometric method was different for the same tree species collected from different locations. FOA-GRNN model combined with LWT and CARS deliver the best performance for Chinese white poplar of Heilongjiang province. In contrast, PLS model showed a good performance for Chinese white poplar collected from Jilin province based on raw spectra. However, for other tree species, RSM-PSO-SVM models can improve the performance of wood density prediction compared to traditional linear and FOA-GRNN models. Especially for Acer mono Maxim, when compared to linear models, the coefficient of determination of prediction set ( ) and relative prediction deviation (RPD) were increased by 47.70% and 44.48%, respectively. And the dimensionality of Vis-NIR spectral data was decreased from 2048 to 20. Therefore, the appropriate chemometric technique should be selected before building calibration models.

On potential limitations of differential expression analysis with non-linear machine learning models

Recently, there has been a growing interest in bioinformatics toward the adoption of increasingly complex machine learning models for the analysis of next-generation sequencing data with the goal of disease subtyping (i.e., patient stratification based on molecular features) or risk-based classification for specific endpoints, such as survival. With gene-expression data, a common approach consists in characterising the emerging groups by exploiting a differential expression analysis, which selects relevant gene sets coupled with pathway enrichment analysis, providing an insight into the underlying biological processes. However, when non-linear machine learning models are involved, differential expression analysis could be limiting since patient groupings identified by the model could be based on a set of genes that are hidden to differential expression due to its linear nature, affecting subsequent biological characterisation and validation. The aim of this study is to provide a proof-of-concept example demonstrating such a limitation. Moreover, we suggest that this issue could be overcome by the adoption of the innovative paradigm of eXplainable Artificial Intelligence, which consists in building an additional explainer to get a trustworthy interpretation of the model outputs and building a reliable set of genes characterising each group, preserving also non-linear relations, to be used for downstream analysis and validation.

Option Return Predictability with Machine Learning and Big Data

Abstract Drawing upon more than 12 million observations over the period from 1996 to 2020, we find that allowing for nonlinearities significantly increases the out-of-sample performance of option and stock characteristics in predicting future option returns. The nonlinear machine learning models generate statistically and economically sizable profits in the long-short portfolios of equity options even after accounting for transaction costs. Although option-based characteristics are the most important standalone predictors, stock-based measures offer substantial incremental predictive power when considered alongside option-based characteristics. Finally, we provide compelling evidence that option return predictability is driven by informational frictions and option mispricing. Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online.

Digital Mapping of Soil Properties Using Ensemble Machine Learning Approaches in an Agricultural Lowland Area of Lombardy, Italy

Sustainable agricultural landscape management needs reliable and accurate soil maps and updated geospatial soil information. Recently, machine learning (ML) models have commonly been used in digital soil mapping, together with limited data, for various types of landscapes. In this study, we tested linear and nonlinear ML models in predicting and mapping soil properties in an agricultural lowland landscape of Lombardy region, Italy. We further evaluated the ability of an ensemble learning model, based on a stacking approach, to predict the spatial variation of soil properties, such as sand, silt, and clay contents, soil organic carbon content, pH, and topsoil depth. Therefore, we combined the predictions of the base learners (ML models) with two meta-learners. Prediction accuracies were assessed using a nested cross-validation procedure. Nonetheless, the nonlinear single models generally performed well, with RF having the best results; the stacking models did not outperform all the individual base learners. The most important topographic predictors of the soil properties were vertical distance to channel network and channel network base level. The results yield valuable information for sustainable land use in an area with a particular soil water cycle, as well as for future climate and socioeconomic changes influencing water content, soil pollution dynamics, and food security.

Pakistan CO2 Emission Modelling and Forecasting: A Linear and Nonlinear Time Series Approach.

Pakistan is considered among the top five countries with the highest CO2 emissions globally. This calls for pragmatic policy implementation by all stakeholders to bring finality to this alarming situation since it contributes greatly to global warming, thereby leading to climate change. This study is an attempt to make a comparative analysis of the linear time series models with nonlinear time series models to study CO2 emission data in Pakistan. These linear and nonlinear time series models were used to model and forecast future values of CO2 emissions for a short period. To assess and select the best model among these linear and nonlinear time series models, we used the root mean square error (RMSE) and the mean absolute error (MAE) as performance indicators. The outputs showed that the nonlinear machine learning models are the best among all other models, having the lowest RMSE and MAE values. Based on the forecasted value of the nonlinear machine learning neural network autoregressive model, Pakistan's CO2 emissions will be 1.048 metric tons per capita by 2028. The increasing trend in emissions is a frightening and clear warning, suggesting that innovative policies must be initiated to reduce the trend. We encourage the Pakistan government to price CO2 emissions by companies and entities per ton, adapt electricity production from hydro, wind, and different sources with no emissions of CO2, initiate rigorous planting of more trees in the populated areas of Pakistan as forest covers, provide incentives to companies, organisations, institutions, and households to come out with clean technologies or use technologies with no CO2 emissions or those with lower ones, and fund more studies to develop clean and innovative technologies with less or no CO2 emissions.

Application of Nonlinear Models Combined with Conventional Laboratory Indicators for the Diagnosis and Differential Diagnosis of Ovarian Cancer.

Existing biomarkers for ovarian cancer lack sensitivity and specificity. We compared the diagnostic efficacy of nonlinear machine learning and linear statistical models for diagnosing ovarian cancer using a combination of conventional laboratory indicators. We divided 901 retrospective samples into an ovarian cancer group and a control group, comprising non-ovarian malignant gynecological tumor (NOMGT), benign gynecological disease (BGD), and healthy control subgroups. Cases were randomly assigned to training and internal validation sets. Two linear (logistic regression (LR) and Fisher's linear discriminant (FLD)) and three nonlinear models (support vector machine (SVM), random forest (RF), and artificial neural network (ANN)) were constructed using 22 conventional laboratory indicators and three demographic characteristics. Model performance was compared. In an independent prospectively recruited validation set, the order of diagnostic efficiency was RF, SVM, ANN, FLD, LR, and carbohydrate antigen 125 (CA125)-only (AUC, accuracy: 0.989, 95.6%; 0.985, 94.4%; 0.974, 93.4%; 0.915, 82.1%; 0.859, 80.1%; and 0.732, 73.0%, respectively). RF maintained satisfactory classification performance for identifying different ovarian cancer stages and for discriminating it from NOMGT-, BGD-, or CA125-positive control. Nonlinear models outperformed linear models, indicating that nonlinear machine learning models can efficiently use conventional laboratory indicators for ovarian cancer diagnosis.

Features from the photoplethysmogram and the electrocardiogram for estimating changes in blood pressure

There is a growing emphasis being placed on the potential for cuffless blood pressure (BP) estimation through modelling of morphological features from the photoplethysmogram (PPG) and electrocardiogram (ECG). However, the appropriate features and models to use remain unclear. We investigated the best features available from the PPG and ECG for BP estimation using both linear and non-linear machine learning models. We conducted a clinical study in which changes in BP (DeltaBP) were induced by an infusion of phenylephrine in 30 healthy volunteers (53.8% female, 28.0 (9.0) years old). We extracted a large and diverse set of features from both the PPG and the ECG and assessed their individual importance for estimating DeltaBP through Shapley additive explanation values and a ranking coefficient. We trained, tuned, and evaluated linear (ordinary least squares, OLS) and non-linear (random forest, RF) machine learning models to estimate DeltaBP in a nested leave-one-subject-out cross-validation framework. We reported the results as correlation coefficient (rho _p), root mean squared error (RMSE), and mean absolute error (MAE). The non-linear RF model significantly (p<0.05) outperformed the linear OLS model using both the PPG and the ECG signals across all performance metrics. Estimating DeltaSBP using the PPG alone (rho _p = 0.86 (0.23), RMSE = 5.66 (4.76) mmHg, MAE = 4.86 (4.29) mmHg) performed significantly better than using the ECG alone (rho _p = 0.69 (0.45), RMSE = 6.79 (4.76) mmHg, MAE = 5.28 (4.57) mmHg), all p < 0.001. The highest ranking features from the PPG largely modelled increasing reflected wave interference driven by changes in arterial stiffness. This finding was supported by changes observed in the PPG waveform in response to the phenylephrine infusion. However, a large number of features were required for accurate BP estimation, highlighting the high complexity of the problem. We conclude that the PPG alone may be further explored as a potential single source, cuffless, blood pressure estimator. The use of the ECG alone is not justified. Non-linear models may perform better as they are able to incorporate interactions between feature values and demographics. However, demographics may not adequately account for the unique and individualised relationship between the extracted features and BP.

Peanut yield prediction with UAV multispectral imagery using a cooperative machine learning approach

&lt;abstract&gt; &lt;p&gt;The unmanned aerial vehicle (UAV), as a remote sensing platform, has attracted many researchers in precision agriculture because of its operational flexibility and capability of producing high spatial and temporal resolution images of agricultural fields. This study proposed machine learning (ML) models and their ensembles for peanut yield prediction using UAV multispectral data. We utilized five bands (red, green, blue, near-infra-red (NIR) and red-edge) multispectral images acquired at various growth stages of peanuts using UAV. The correlation between spectral bands and yield was analyzed for each growth stage, which showed that the maturity stages had a significant correlation between peanut yield and spectral bands: red, green, NIR and red edge (REDE). Using these four bands spectral data, we assessed the potential for peanut yield prediction using multiple linear regression and seven non-linear ML models whose hyperparameters were optimized using simulated annealing (SA). The best three ML models, random forest (RF), support vector machine (SVM) and XGBoost, were then selected to construct a cooperative yield prediction framework with both the best ML model and the ensemble scheme from the best three as comparable recommendations to the farmers.&lt;/p&gt; &lt;/abstract&gt;

Statistical Time Series Forecasting Models for Pandemic Prediction

COVID-19 has significantly impacted the growth prediction of a pandemic, and it is critical in determining how to battle and track the disease progression. In this case, COVID-19 data is a time-series dataset that can be projected using different methodologies. Thus, this work aims to gauge the spread of the outbreak severity over time. Furthermore, data analytics and Machine Learning (ML) techniques are employed to gain a broader understanding of virus infections. We have simulated, adjusted, and fitted several statistical time-series forecasting models, linear ML models, and nonlinear ML models. Examples of these models are Logistic Regression, Lasso, Ridge, ElasticNet, Huber Regressor, Lasso Lars, Passive Aggressive Regressor, K-Neighbors Regressor, Decision Tree Regressor, Extra Trees Regressor, Support Vector Regressions (SVR), AdaBoost Regressor, Random Forest Regressor, Bagging Regressor , AuoRegression, MovingAverage, Gradient Boosting Regressor, Autoregressive Moving Average (ARMA), Auto-Regressive Integrated Moving Averages (ARIMA), SimpleExpSmoothing, Exponential Smoothing, Holt-Winters, Simple Moving Average, Weighted Moving Average, Croston, and naive Bayes. Furthermore, our suggested methodology includes the development and evaluation of ensemble models built on top of the best-performing statistical and ML-based prediction methods. A third stage in the proposed system is to examine three different implementations to determine which model delivers the best performance. Then, this best method is used for future forecasts, and consequently, we can collect the most accurate and dependable predictions.

Machine learning for predicting elections in Latin America based on social media engagement and polls

The popularity of contemporary social media (SM) has impacted democratic practices, and the success of presidential campaigns is frequently attributed to SM performance. Within this new scenario, many methodological proposals that use SM data have been put forward for predicting election results. However, the most common approach, based on the volume and sentiment analysis of mentions on Twitter, has been frequently criticized and challenged. Thus, recent surveys have indicated new directions, such as the use of data from more than one SM platform, the adoption of nonlinear machine learning (ML) models, and the validation of methodologies and experiments in different elections. In this context, the present paper proposes SoMEN, the Social Media framework for Election Nowcasting, a framework composed of a process and an ML model for nowcasting election results based on SM performance as features and with offline polls as labeled data. It also defines SoMEN-DC, an execution strategy for SoMEN that enables continuous prediction during the campaign (DC). The proposed metrics and framework were applied to some of the main recent presidential elections in Latin America: Argentina (2019), Brazil (2018), Colombia (2018), and Mexico (2018). More than 65,000 posts were collected from the SM profiles of candidates on Facebook, Twitter, and Instagram with data from 195 presidential polls. Results have demonstrated that it was possible to achieve a high level of accuracy in predicting the final vote share of the candidates and to make daily predictions, providing competitive or better results than the traditional polls. The strategies put forward in this study have attempted to address several of the current challenges in this research area and have indicated a new manner of how to face the problems. They may also be directly used for predicting future elections in similar scenarios.

Solubility Characteristics of Acetaminophen and Phenacetin in Binary Mixtures of Aqueous Organic Solvents: Experimental and Deep Machine Learning Screening of Green Dissolution Media

The solubility of active pharmaceutical ingredients is a mandatory physicochemical characteristic in pharmaceutical practice. However, the number of potential solvents and their mixtures prevents direct measurements of all possible combinations for finding environmentally friendly, operational and cost-effective solubilizers. That is why support from theoretical screening seems to be valuable. Here, a collection of acetaminophen and phenacetin solubility data in neat and binary solvent mixtures was used for the development of a nonlinear deep machine learning model using new intuitive molecular descriptors derived from COSMO-RS computations. The literature dataset was augmented with results of new measurements in aqueous binary mixtures of 4-formylmorpholine, DMSO and DMF. The solubility values back-computed with the developed ensemble of neural networks are in perfect agreement with the experimental data, which enables the extensive screening of many combinations of solvents not studied experimentally within the applicability domain of the trained model. The final predictions were presented not only in the form of the set of optimal hyperparameters but also in a more intuitive way by the set of parameters of the Jouyban-Acree equation often used in the co-solvency domain. This new and effective approach is easily extendible to other systems, enabling the fast and reliable selection of candidates for new solvents and directing the experimental solubility screening of active pharmaceutical ingredients.