Groundwater resources are expensive to develop and use; they are difficult to monitor and data collected from monitoring wells are often sporadic, often only available at irregular, infrequent, or brief intervals Groundwater managers require an accurate understanding of historic groundwater storage trends to effectively manage groundwater resources, however, most if not all well records contain periods of missing data To understand long-term trends, these missing data need to be imputed before trend analysis We present a method to impute missing data at single wells, by exploiting data generated from Earth observations that are available globally We use two soil moisture models, the Global Land Data Assimilation System (GLDAS) model and National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center (CPC) soil moisture model to impute the missing data Our imputation method uses a machine learning technique called Extreme Learning Machine (ELM) Our implementation uses 11 input data-streams, all based on Earth observation data We train and apply the model one well at a time We selected ELM because it is a single hidden layer feedforward model that can be trained quickly on minimal data We tested the ELM method using data from monitoring wells in the Cedar Valley and Beryl-Enterprise areas in southwest Utah, USA We compute error estimates for the imputed data and show that ELM-computed estimates were more accurate than Kriging estimates This ELM-based data imputation method can be used to impute missing data at wells These complete time series can be used improve the accuracy of aquifer groundwater elevation maps in areas where in-situ well measurements are sparse, resulting in more accurate spatial estimates of the groundwater surface The data we use are available globally from 1950 to the present, so this method can be used anywhere in the world

/pdf/exploiting-earth-observation-data-to-impute-groundwater-3ncznrcu5a.pdf

Exploiting Earth Observation Data to Impute Groundwater Level Measurements with an Extreme Learning Machine

In the United States, stroke is the fifth prominent cause of fatality and it is a major reason of serious disability among the adult population [1]. Therefore, it is crucial that we can predict stroke accurately in order to be treated in early stages. Nowadays, use of Machine Learning (ML) algorithms have been in great demand to predict patient's condition in advance and inform the medical staff to avoid the risk of disease progression. Kaggle Healthcare dataset has been widely used by many researchers in this area for developing models for stroke prediction. The dataset has 43,400 instances and 10 features. This paper proposes a method for the analysis and prediction of stroke on the same dataset using Microsoft Azure Machine Learning (AzureML) which is a cloud-based platform. We have applied Chi-Squared test on the dataset for extracting the top features. The experiments are run on AzureML with the top 6 features as well as with all the features. In addition, we compare accuracy between the two models trained by the top 6 features and all the features. The performance of Two-class Decision Jungle with top 6 features has been set as the benchmark in our work. Two-Class Boosted Decision Tree, an ensemble learning method achieves 96.8% accuracy using the top 6 features. Our experimental results show that with the right features, we could improve the accuracy significantly for the stroke prediction, and it also takes less time to train the model.

Chi-Squared Based Feature Selection for Stroke Prediction using AzureML

—Brain Strokes are considered one of the deadliest brain diseases due to their sudden occurrence, so predicting their occurrence and treating the factors may reduce their risk. This paper aimed to propose a brain stroke prediction model using machine learning classifiers and a stacking ensemble classifier. The smote technique was employed for data balancing, and the standardization technique was for data scaling. The classifiers’ best parameters were chosen using the hyperparameter tuning technique. The proposed stacking prediction model was created by combining Random Forest (RF), K-Nearest Neighbors (KNN), Logistic Regression (LR), Support Vector Machine (SVM), and Naive Bayes (NB) as base classifiers, and meta learner was chosen to be Random Forest. The performance of the proposed stacking model has been evaluated using Accuracy, Precision, Recall, and F1 score. In addition, the Matthews Correlation Coefficient (MCC) has been also used for more reliable evaluation when having an unbalanced dataset, which is the case in most medical datasets. The results demonstrate that the proposed stacking model outperforms the standalone classifiers by achieving an accuracy of 97% and an MCC value of 94%.

A Machine Learning Ensemble Classifier for Prediction of Brain Strokes

Missing data is one of the main difficulties in working with fluviometric records. Database gaps may result from fluviometric stations components problems, monitoring interruptions and lack of observers. Incomplete series analysis generates uncertain results, negatively impacting water resources management. Thus, proper missing data consideration is very important to ensure better information quality. This work aims to analyze, comparatively, missing data imputation methodologies in monthly river-flow time series, considering, as a case study, the Doce River, located in Southeast Brazil. Missing data were simulated in 5%, 10%, 15%, 25% and 40% proportions following a random distribution pattern, ignoring the missing data generation mechanisms. Ten missing data imputation methodologies were used: arithmetic mean, median, simple and multiple linear regression, regional weighting, spline and Stineman interpolation, Kalman smoothing, multiple imputation and maximum likelihood. Their performances were compared through bias, root mean square error, absolute mean percentage error, determination coefficient and concordance index. Results indicate that for 5% missing data, any methodology for imputing can be considered, recommending caution for arithmetic mean method application. However, as the missing data proportion increases, it is recommended to use multiple imputation and maximum likelihood methodologies when there are support stations for imputation, and the Stineman interpolation and Kalman Smoothing methods when only the studied series is available. 
Keywords: Doce river, imputation, missing data.

/pdf/methodological-approaches-for-imputing-missing-data-into-1yubapjm.pdf

Methodological approaches for imputing missing data into monthly flows series

A review of the current publication trends on missing data imputation over three decades: direction and future research

In XML databases, the indexing process is based on a labelling or numbering scheme and generally used to label an XML document to perform an XML query using the path node information. Moreover, a labelling scheme helps to capture the structural relationships during query processing without the need to access the physical document. Two of the main problems for labelling XML schemes are duplicated labels, and cost efficiency regarding labelling time and size. This research presents a novel dynamic XML labelling scheme, called the Pentagonal Scheme, in which data are represented as ordered XML nodes with relationships between them. The update of these nodes from large-scale XML documents has been widely investigated and represents a challenging research problem, as it means relabelling a whole tree. Our algorithms provide an efficient dynamic XML labelling scheme to support data updates without duplicating labels or relabelling old nodes. Our work evaluates the labelling process in terms of size and time, and evaluates the labelling scheme’s ability to handle several insertions in XML documents. The findings indicate that the Pentagonal scheme shows a better initial labelling time performance than compared schemes. Particularly when using large size XML datasets. Moreover, it efficiently supports random skewed updates, has fast calculations and uncomplicated implementations to thus handle updates efficiently. In addition, the comparable evaluation of the query response time and relationships in Pentagonal scheme can be efficiently performed without presenting any extra cost. It was for this reason that our labelling scheme achieved the goal of this research.

Pentagonal scheme for dynamic XML prefix labelling

Machine learning algorithms became popular in many domains including applications in healthcare. However, in some cases, the datasets classifiers perform poorly due to several reasons. Studies have shown that combining classifiers may help improve performance and obtain better outcomes. Ensemble approach is a technique of combining two or more algorithms to make a robust system for predictions from all the base learners. Ensemble approach works based on supervised learning algorithms in which predictions of various learning algorithms are combined to make an estimation. The simplest type of ensemble learning is to train the base algorithms on random subsets of original data set and after that make a vote for by counting the most common classifications or by computing a form of averaging the predictions of the base algorithms. In this paper, various classifiers have been applied and compared for effective diagnosis of the Stroke data set. The Stroke dataset is used to demonstrate the effectiveness of the Ensemble approach for obtaining good predictions. Experimental results show that classifier Ensemble produces better prediction accuracy.

Ensemble of Multiple Classification Algorithms to Predict Stroke Dataset

Real world data often contain missing values. Data mining techniques have been actively used to overcome this problem by using methods of imputing the missing values. In particular, before applying any classification model, handling missing data in the dataset is an important task in pre-processing stages for ensuring the quality of classification results. Using the appropriate method of missing value imputation can help to generate complete datasets for improving the classifier’s performance. Many approaches have been proposed in the field of machine learning and data mining for handling missing values. Techniques used for imputing missing values can be divided into single and multiple methods. Some techniques, namely random forests, CART, k-NN imputation method and mean method, remove attributes and observations, predicting missed values by Multivariate Imputation method by Chained Equations (MICE) for example. In this study various approaches of treating missing values were applied on different decision trees algorithms to investigate how these techniques can be used effectively to improve the performance of selected classifiers. The Stroke data set was used in these experiments to check how well the methods of handling missing values work. Moreover, the paper reports how using data imputation methods affect classification results. The best results are obtained from the classifiers with removing variables that have missing data more than the rest of attributes. This work presents an attempt to analyse the chosen techniques with the purpose to investigate their strengths and weaknesses in handling missing values, and reports that both imputation methods (MIM and MICE) are efficient and yield similar accuracy.

Ebtesam Taktek

Papers

Pentagonal scheme for dynamic XML prefix labelling

Ensemble of Multiple Classification Algorithms to Predict Stroke Dataset

Performance Analysis of Missing Values Imputation Methods Using Machine Learning Techniques