Generation Of Synthetic Samples Research Articles

PRM66 - Synthetic Sample Generation Representing the English Population Using Spearman Rank Correlation and Chomsky Decomposition

To generate a synthetic sample of 1 million individuals that reflect the characteristics of the population recorded in the Health Survey for England (HSE). We used data from the HSE to determine the age and gender-dependent distributions of continuous variable risk factors (height, weight, BMI, systolic blood pressure, total and HDL cholesterol and their ratio, number of cigarettes/day and units of alcohol/week) and prevalence of binary risk factors (smoking status, diabetes). Spearman rank correlations including age and gender were determined for these risk factors. A table of normally distributed random numbers was generated. Cholesky decomposition was used to replicate the observed Spearman rank correlations in the table of random numbers. Rank correlations that included binary variables were recalibrated to adjust for numerous tied values. The sample was then generated using a reverse look-up of the gamma distribution value using the random percentiles for continuous variables or setting a binary variable to 1 when the random percentile falls below the prevalence threshold. Differences between coefficients were no more than 0.5% for any continuous variable. The prevalence of binary factors in the SS was very well matched with the HSE sample. Smoker incidence rates were 18.8% and 16.7% in the SS versus 18.4% and 16.5% in the HSE sample, for males and females respectively. Prevalence of diabetes in the SS was 13.3% and 7.7% versus 13.2% and 7.8%, and for cardiovascular disease was 17.6% and 14.1% versus 18.2% and 14.6%. Comparing 25th, 50th and 75th quantiles, the maximum difference between the original and synthetic values for BMI and TC/HDL ratio were 0.6Kg and 0.3 respectively. Our new approach generates large synthetic samples with risk factor distributions very closely matching those of the real HSE population. This sample can be used to model the likely impact of new therapies or predict mortality.

REMEDIAL-HwR: Tackling multilabel imbalance through label decoupling and data resampling hybridization

The learning from imbalanced data is a deeply studied problem in standard classification and, in recent times, also in multilabel classification. A handful of multilabel resampling methods have been proposed in late years, aiming to balance the labels distribution. However these methods have to face a new obstacle, specific for multilabel data, as is the joint appearance of minority and majority labels in the same data patterns. We proposed recently a new algorithm designed to decouple imbalanced labels concurring in the same instance, called REMEDIAL (\textit{REsampling MultilabEl datasets by Decoupling highly ImbAlanced Labels}). The goal of this work is to propose a procedure to hybridize this method with some of the best resampling algorithms available in the literature, including random oversampling, heuristic undersampling and synthetic sample generation techniques. These hybrid methods are then empirically analyzed, determining how their behavior is influenced by the label decoupling process. As a result, a noteworthy set of guidelines on the combined use of these techniques can be drawn from the conducted experimentation.

A sigma-lognormal model-based approach to generating large synthetic online handwriting sample databases

This article describes a methodology to generate a large database of synthetic samples from a small set of original online handwriting specimens. The overall paradigm is based on the Kinematic Theory of rapid human movements and its sigma-lognormal model. The principal contributions of the present study include (i) development of a strategy for sigma-lognormal model-based generation of synthetic samples from real online handwriting samples of arbitrary scripts captured by arbitrary relevant devices and (ii) verification of the structural similarities, including the naturalness of such synthetic prototypes, through various human perception experiments, computer evaluations and statistical hypothesis testing. A database consisting of a large number of online synthetic handwritten word samples is used to train and evaluate the performance of three existing automatic online handwriting recognition systems. Training based on a combined set of original and synthetic samples improves the recognition accuracies on the test set. A combined training set is useful irrespective of the nature of the feature set used (online, offline or combined). Although the proposed method has primarily been developed and applied to the design of an online handwriting sample database of a popular Indian script, Bangla, it can be applied to the generation of large databases of any arbitrary script for example: English, Chinese and Arabic.

Using Genetic Approach for Learning from Imbalanced Text Corpora

Aiming at the ever-present problem of imbalanced data in text classification, the paper employs the Genetic Algorithm approach for tackling the imbalance problem in a binary classed text data. One of the inherent characteristics of imbalanced data is the highly uneven distribution of data among the classes. Consequentially, the traditional classifier algorithms such as the Nearest Neighbor have shown a decreased performance due to the under representation of the interested class. A hybrid approach has been used to incorporate the oversampling technique with the advantages of Genetic Algorithm for generation of the artificial patterns for the minority class. This approach employs avoidance of over fitting as the fitness function to decide the stopping criterion for generation of synthetic samples. Efficient evaluation measures analyze the increase in performance of the proposed hybrid-learning model.

Estimation of seismic building structural types using multi-sensor remote sensing and machine learning techniques

Detailed information about seismic building structural types (SBSTs) is crucial for accurate earthquake vulnerability and risk modeling as it reflects the main load-bearing structures of buildings and, thus, the behavior under seismic load. However, for numerous urban areas in earthquake prone regions this information is mostly outdated, unavailable, or simply not existent. To this purpose, we present an effective approach to estimate SBSTs by combining scarce in situ observations, multi-sensor remote sensing data and machine learning techniques. In particular, an approach is introduced, which deploys a sequential procedure comprising five main steps, namely calculation of features from remote sensing data, feature selection, outlier detection, generation of synthetic samples, and supervised classification under consideration of both Support Vector Machines and Random Forests. Experimental results obtained for a representative study area, including large parts of the city of Padang (Indonesia), assess the capabilities of the presented approach and confirm its great potential for a reliable area-wide estimation of SBSTs and an effective earthquake loss modeling based on remote sensing, which should be further explored in future research activities.

Evaluation of technique to overcome small dataset problems during neural-network based contamination classification of packaged beef using integrated olfactory sensor system

Similar to most biological studies, beef contamination classification studies using artificial neural networks are restricted to small datasets. This study evaluates multivariate normal (MVN) technique of synthetic sample generation on small datasets associated with Salmonella contamination in beef. Six experiments were conducted to evaluate the performance of integrated sensor system towards identification of Salmonella contaminated beef packages. Pattern recognition involved using wavelet packet transform for feature extraction from sensor array responses and radial basis function network (RBFN) based classification of contaminated beef packages from uncontaminated packages. The MVN generated synthetic olfactory sensor signatures were used to train and test the RBFN classifiers. For the datasets analyzed in this study, genetic algorithm optimized RBF networks conferred average contamination test classification accuracies of 90.33% ± 7.68% (mean ± std. dev.) which were higher compared to the bootstrapped quadratic discriminant analysis based average accuracies. RBFN classifier based average overall classification accuracies of six synthetically generated datasets were in the range of 86.66%–98.89% with highest average overall classification accuracies of 98.89% ± 1.92%.

Neural-Network-Based Classification of Meat: Evaluation of Techniques to Overcome Small Dataset Problems

One of the objectives of our multidisciplinary research group is to develop sensors for the detection of Salmonella contamination in beef. Similar to most biological studies, beef contamination classification studies using artificial neural networks (ANNs) are restricted to small datasets. This study evaluates selected techniques of data domain expansion and synthetic sample generation on small datasets associated with meat contamination. Mega-trend-diffusion (MTD) and functional virtual population (FVP) techniques for data domain expansion and synthetic sample generation were assessed on the small datasets. The datasets used were obtained from a thin-film (TF) module electronic nose system in response to the headspace of control and Salmonella-inoculated packaged meat samples. Back-propagation neural networks (BPNNs) were used to determine classification accuracies of the synthetically expanded datasets. For aged beef datasets, the maximum mean of average overall classification accuracies provided by FVP technique was 90%. The maximum mean of average overall classification accuracies obtained by FVP technique was about 81% for fresh beef datasets. MTD technique also provided similar accuracies (in the lower 80s). Both techniques were found useful for expanding the domain range of the small dataset in order to test and evaluate BPNN-based classification models.

Initialization for generating single‐site and multisite low‐order periodic autoregressive and moving average processes

Single‐site and multisite periodic autoregressive and moving average (PARMA) processes have been suggested for modeling and simulation of hydrologic processes. Generation of synthetic samples based on such models requires knowledge of the initial conditions, i.e., the values of the process at times prior to t = 1. In this paper an initialization procedure for generating univariate and multivariate PAR (1), PAR (2), and PARMA (1, 1) processes is described. The procedure is based on generating the initial values of the process in such a way that the variance‐covariance structure of the process is preserved. The initialization procedure introduced herein is exact and easy to use. Examples are included to explain the applicability of the proposed approach.