Small Dataset Size Research Articles

Imprecise global sensitivity analysis using bayesian multimodel inference and importance sampling

Global Sensitivity Analysis (GSA) aims to understand the relative importance of uncertain input variables to model response. Conventional GSA involves calculating sensitivity (Sobol’) indices for a model with known model parameter distributions. However, model parameters are affected by aleatory and epistemic uncertainty, with the latter often caused by lack of data. We propose a new framework to quantify uncertainty in probability model-form and model parameters resulting from small datasets and integrate these uncertainties into Sobol’ index estimates. First, the process establishes, through Bayesian multimodel inference, a set of candidate probability models and their associated probabilities. Imprecise Sobol’ indices are calculated from these probability models using an importance sampling reweighting approach. This results in probabilistic Sobol’ indices, whose distribution characterizes uncertainty in the sensitivity resulting from small dataset size. The imprecise Sobol’ indices thus provide a measure of confidence in the sensitivity estimate and, moreover, can be used to inform data collection efforts targeted to minimize the impact of uncertainties. Through an example studying the parameters of a Timoshenko beam, we show that these probabilistic Sobol’ indices converge to the true/deterministic Sobol’ indices as the dataset size increases and hence, distribution-form uncertainty reduces. The approach is then applied to assess the sensitivity of the out-of-plane properties of an E-glass fiber composite material to its constituent properties. This second example illustrates the approach for an important class of materials with wide-ranging applications when data may be lacking for some input parameters.

Relationship between temporal anomalies in PM2.5 concentrations and reported influenza/influenza-like illness activity

A small number of studies suggest atmospheric particulate matter with diameters 2.5 micron and smaller (PM2.5) may possibly play a role in the transmission of influenza and influenza-like illness (ILI) symptoms. Those studies were predominantly conducted under moderately to highly polluted outdoor atmospheres. The purpose of this study was to extend the data set to include a less polluted atmospheric environment. A relationship between PM2.5 and ILI activity extended to include lightly to moderately polluted atmospheres could imply a more complicated mechanism than that suggested by existing studies. We obtained concurrent PM2.5 mass concentration data, meteorological data and reported Influenza and influenza-like illness (ILI) activity for the light to moderately polluted atmospheres over the Tucson, AZ region. We found no relation between PM2.5 mass concentration and ILI activity. There was an expected relation between ILI, activity, temperature, and relative humidity. There was a possible relation between PM2.5 mass concentration anomalies and ILI activity. These results might be due to the small dataset size and to the technological limitations of the PM measurements. Further study is recommended since it would improve the understanding of ILI transmission and thereby improve ILI activity/outbreak forecasts and transmission model accuracies.

A generalizable and sensor-independent deep learning method for fault detection and location in low-voltage distribution grids

Power outages in electrical grids can have very negative economic and societal impacts rendering fault diagnosis paramount to their secure and reliable operation. In this paper, deep neural networks are proposed for fault detection and location in low-voltage smart distribution grids. Due to its key properties, the proposed method solves some of the drawbacks of the existing literature methods, namely a method that: 1) is not limited by the grid topology; 2) is branch-independent; 3) can localize faults even with limited data; 4) is the first to accurately detect and localize high-impedance faults in the low-voltage distribution grid. The generalizability of the method derives from the non-grid specific nature of the inputs that it requires, inputs that can be obtained from any grid. To evaluate the proposed method, a real low-voltage distribution grid in Portugal is considered and the robustness of the method is tested against several disturbances including large fault resistance values (up to 1000 Ω). Based on the case study, it is shown that the proposed methodology outperforms conventional fault diagnosis methods: it detects faults with 100% accuracy, identifies faulty branches with 83.5% accuracy, and estimates the exact fault location with an average error of less than 11.8%. Finally, it is also shown that: 1) even when reducing the available measurements to the bare minimum, the accuracy of the proposed method is only decreased by 4.5%; 2) while deep neural networks usually require large amounts of data, the proposed model is accurate even for small dataset sizes.

Improving classification accuracy using data augmentation on small data sets

Data augmentation (DA) is a key element in the success of Deep Learning (DL) models, as its use can lead to better prediction accuracy values when large size data sets are used. DA was not very much used with earlier neural network models before 2012, and the reason might be related to the type of models and the size of the data sets used. We investigate in this work, applying several state-of-the-art models based on Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs), the effect of DA when using small size data sets, analyzing the results in terms of the prediction accuracy obtained according to the different characteristics of the training samples (number of instances and features, and class unbalance degree). We further introduce modifications to the standard methods used to generate the synthetic samples to alter the class balance representation, and the overall results indicate that with some computational effort a significant increase in prediction accuracy can be obtained when small data sets are considered.

Integrating Improved U-Net and Continuous Maximum Flow Algorithm for 3D Brain Tumor Image Segmentation

Abstract To address the issues of the relatively small size of brain tumor image datasets, severe class imbalance, and low precision in existing segmentation algorithms for brain tumor images, this study proposes a two-stage segmentation algorithm integrating convolutional neural networks (CNNs) and conventional methods. Four modalities of the original magnetic resonance images were first preprocessed separately. Next, preliminary segmentation was performed using an improved U-Net CNN containing deep monitoring, residual structures, dense connection structures, and dense skip connections. The authors adopted a multiclass Dice loss function to deal with class imbalance and successfully prevented overfitting using data augmentation. The preliminary segmentation results subsequently served as the a priori knowledge for a continuous maximum flow algorithm for fine segmentation of target edges. Experiments revealed that the mean Dice similarity coefficients of the proposed algorithm in whole tumor, tumor core, and enhancing tumor segmentation were 0.9072, 0.8578, and 0.7837, respectively. The proposed algorithm presents higher accuracy and better stability in comparison with some of the more advanced segmentation algorithms for brain tumor images.

Speeded-up Convolution Neural Network for classification tasks using multiscale 2-dimensional decomposition

In this paper, we propose a strategy for network simplification and acceleration. First, we propose to generate a suitable resized image using multiscale patching in the first convolutional layer, which can then be used for the rest of the network. We use p convolutional filters that operate on patches of size m × n, and we first select all the possible non-superposed m × n patches from the available images. If the number of such patches is not sufficient, the remaining ones are collected using scales a × b or c × d such that ab = cd = mn or a = 2 m or b = 2n. Patches generated from the former condition are directly extracted from the images, while the downsampled results are used in the latter case.We also introduce a 2-dimensional decomposition for patch compression, by stacking all the available image patches along the columns and applying a 2D PCA decomposition. Finally, a layer weight decomposition technique followed by module-based finetuning is adopted, for a new fast module-based CNN model.Extensive evaluations using public data sets like MNIST, Pascal VOC, and WebCASIA, and with different state-of-the art CNN architectures like Open-Face, VGG and DarkNet verify that our proposed model is able to accelerate the training process and even to provide higher classification accuracies for small-sized datasets. We obtain 8% increase in Top-1 and Top-5 recognition rates, and 5% increase in F1 score over general interpolation-based resizing.

Multi-Scale Self-Attention for Text Classification

In this paper, we introduce the prior knowledge, multi-scale structure, into self-attention modules. We propose a Multi-Scale Transformer which uses multi-scale multi-head self-attention to capture features from different scales. Based on the linguistic perspective and the analysis of pre-trained Transformer (BERT) on a huge corpus, we further design a strategy to control the scale distribution for each layer. Results of three different kinds of tasks (21 datasets) show our Multi-Scale Transformer outperforms the standard Transformer consistently and significantly on small and moderate size datasets.

Kernel dependence analysis and graph structure morphing for novelty detection with high-dimensional small size data set

In this study, we propose a new approach for novelty detection that uses kernel dependence techniques for characterizing the statistical dependencies of random variables (RV) and use this characterization as a basis for making inference. Considering the statistical dependencies of the RVs in multivariate problems is an important challenge in novelty detection. Ignoring these dependencies, when they are strong, may result in inaccurate inference, usually in the form of high false positive rates. Previously studied methods, such as graphical models or conditional classifiers, mainly use density estimation techniques as their main learning element to characterize the dependencies of the relevant RVs. Therefore, they suffer from the curse of dimensionality which makes them unable to handle high-dimensional problems. The proposed method, however, avoids using density estimation methods, and rather, employs a kernel method, which is robust with respect to dimensionality, to encode the dependencies and hence, it can handle problems with arbitrarily high-dimensional data. Furthermore, the proposed method does not need any prior information about the dependence structure of the RVs; thus, it is applicable to general novelty detection problems with no simplifying assumption. To test the performance of the proposed method, we apply it to realistic application problems for analyzing sensor networks and compare the results to those obtained by peer methods.

Outlier Detection in High Dimensional Data

High-dimensional data poses unique challenges in outlier detection process. Most of the existing algorithms fail to properly address the issues stemming from a large number of features. In particular, outlier detection algorithms perform poorly on dataset of small size with a large number of features. In this paper, we propose a novel outlier detection algorithm based on principal component analysis and kernel density estimation. The proposed method is designed to address the challenges of dealing with high-dimensional data by projecting the original data onto a smaller space and using the innate structure of the data to calculate anomaly scores for each data point. Numerical experiments on synthetic and real-life data show that our method performs well on high-dimensional data. In particular, the proposed method outperforms the benchmark methods as measured by [Formula: see text]-score. Our method also produces better-than-average execution times compared with the benchmark methods.

Sleep Stage Classification Using Time-Frequency Spectra From Consecutive Multi-Time Points.

Sleep stage classification is an open challenge in the field of sleep research. Considering the relatively small size of datasets used by previous studies, in this paper we used the Sleep Heart Health Study dataset from the National Sleep Research Resource database. A long short-term memory (LSTM) network using a time-frequency spectra of several consecutive 30 s time points as an input was used to perform the sleep stage classification. Four classical convolutional neural networks (CNNs) using a time-frequency spectra of a single 30 s time point as an input were used for comparison. Results showed that, when considering the temporal information within the time-frequency spectrum of a single 30 s time point, the LSTM network had a better classification performance than the CNNs. Moreover, when additional temporal information was taken into consideration, the classification performance of the LSTM network gradually increased. It reached its peak when temporal information from three consecutive 30 s time points was considered, with a classification accuracy of 87.4% and a Cohen’s Kappa coefficient of 0.8216. Compared with CNNs, our results indicate that for sleep stage classification, the temporal information within the data or the features extracted from the data should be considered. LSTM networks take this temporal information into account, and thus, may be more suitable for sleep stage classification.

Document-Level Text Classification Using Single-Layer Multisize Filters Convolutional Neural Network

The rapid growth of electronic documents are causing problems like unstructured data that need more time and effort to search a relevant document. Text Document Classification (TDC) has a great significance in information processing and retrieval where unstructured documents are organized into pre-defined classes. Urdu is the most favorite research language in South Asian languages because of its complex morphology, unique features, and lack of linguistic resources like standard datasets. As compared to short text, like sentiment analysis, long text classification needs more time and effort because of large vocabulary, more noise, and redundant information. Machine Learning (ML) and Deep Learning (DL) models have been widely used in text processing. Despite the major limitations of ML models, like learn directed features, these are the favorite methods for Urdu TDC. To the best of our knowledge, it is the first study of Urdu TDC using DL model. In this paper, we design a large multi-purpose and multi-format dataset that contain more than ten thousand documents organize into six classes. We use Single-layer Multisize Filters Convolutional Neural Network (SMFCNN) for classification and compare its performance with sixteen ML baseline models on three imbalanced datasets of various sizes. Further, we analyze the effects of preprocessing methods on SMFCNN performance. SMFCNN outperformed the baseline classifiers and achieved 95.4%, 91.8%, and 93.3% scores of accuracy on medium, large and small size dataset respectively. The designed dataset would be publically and freely available in different formats for future research in Urdu text processing.

Benchmarking Small-Dataset Structure-Activity-Relationship Models for Prediction of Wnt Signaling Inhibition

Quantitative structure-activity relationship (QSAR) models based on machine learning algorithms are powerful tools to expedite drug discovery processes and therapeutics development. Given the cost in acquiring large-sized training datasets, it is useful to examine if QSAR analysis can reasonably predict drug activity with only a small-sized dataset (size <; 100) and benchmark these small-dataset QSAR models in application-specific studies. To this end, here we present a systematic benchmarking study on small-dataset QSAR models built for prediction of effective Wnt signaling inhibitors, which are essential to therapeutics development in prevalent human diseases (e.g., cancer). Specifically, we examined a total of 72 two-dimensional (2D) QSAR models based on 4 best-performing algorithms, 6 commonly used molecular fingerprints, and 3 typical fingerprint lengths. We trained these models using a training dataset (56 compounds), benchmarked their performance on 4 figures-of-merit (FOMs), and examined their prediction accuracy using an external validation dataset (14 compounds). Our data show that the model performance is maximized when: 1) molecular fingerprints are selected to provide sufficient, unique, and not overly detailed representations of the chemical structures of drug compounds; 2) algorithms are selected to reduce the number of false predictions due to class imbalance in the dataset; and 3) models are selected to reach balanced performance on all 4 FOMs. These results may provide general guidelines in developing high-performance small-dataset QSAR models for drug activity prediction.

Automatic Phase Picking From Microseismic Recordings Using Feature Extraction and Neural Network

High-accuracy microseismic phase picking is fundamental to microseismic signal processing. Phase picking methods based on deep learning show great potential dealing with low signal to noise ratio (SNR) data but need enormous training data. However, it's not easy to obtain a big size of field datasets and label them manually to train the neural network. In this paper, a novel method is proposed by applying feature extraction (Gammatone Feature) and neural networks to pick phases automatically. The feature extraction scheme aims to train the neural network using a relatively small size of training datasets. To test the performances of the proposed method, synthetic datasets were obtained by numerical simulation and used to train the neural network. Both synthetic datasets and field datasets were used to test the neural network for picking P- and S-phases of microseismic events. The results of phase picking illustrate that: (1) feature extraction scheme in the training stage can help reduce the size of training datasets; (2) The neural network can be trained well just by synthetic data and phase picking results are accurate and satisfying when the method was tested by both synthetic and field datasets.

A Multicluster Approach to Selecting Initial Sets for Clustering of Categorical Data

Aim/Purpose: This article proposes a methodology for selecting the initial sets for clustering categorical data. The main idea is to combine all the different values of every single criterion or attribute, to form the first proposal of the so-called multiclusters, obtaining in this way the maximum number of clusters for the whole dataset. The multiclusters thus obtained, are themselves clustered in a second step, according to the desired final number of clusters. Background: Popular cluster methods for categorical data, such as the well-known K-Modes, usually select the initial sets by means of some random process. This fact introduces some randomness in the final results of the algorithms. We explore a different application of the clustering methodology for categorical data that overcomes the instability problems and ultimately provides a greater clustering efficiency. Methodology: For assessing the performance of the proposed algorithm and its comparison with K-Modes, we apply both of them to categorical databases where the response variable is known but not used in the analysis. In our examples, that response variable can be identified to the real clusters or classes to which the observations belong. With every data set, we perform a two-step analysis. In the first step we perform the clustering analysis on data where the response variable (the real clusters) has been omitted, and in the second step we use that omitted information to check the efficiency of the clustering algorithm (by comparing the real clusters to those given by the algorithm). Contribution: Simplicity, efficiency and stability are the main advantages of the multicluster method. Findings: The experimental results attained with real databases show that the multicluster algorithm has greater precision and a better grouping effect than the classical K-modes algorithm. Recommendations for Practitioners: The method can be useful for those researchers working with small and medium size datasets, allowing them to detect the underlying structure of the data in an intuitive and reasonable way. Recommendation for Researchers: The proposed algorithm is slower than K-Modes, since it devotes a lot of time to the calculation of the initial combinations of attributes. The reduction of the computing time is therefore an important research topic. Future Research: We are concerned with the scalability of the algorithm to large and complex data sets, as well as the application to mixed data sets with both quantitative and qualitative attributes.

C-RNNCrispr: Prediction of CRISPR/Cas9 sgRNA activity using convolutional and recurrent neural networks

CRISPR/Cas9 is a hot genomic editing tool, but its success is limited by the widely varying target efficiencies among different single guide RNAs (sgRNAs). In this study, we proposed C-RNNCrispr, a hybrid convolutional neural networks (CNNs) and bidirectional gate recurrent unit network (BGRU) framework, to predict CRISPR/Cas9 sgRNA on-target activity. C-RNNCrispr consists of two branches: sgRNA branch and epigenetic branch. The network receives the encoded binary matrix of sgRNA sequence and four epigenetic features as inputs, and produces a regression score. We introduced a transfer learning approach by using small-size datasets to fine-tune C-RNNCrispr model that were pre-trained from benchmark dataset, leading to substantially improved predictive performance. Experiments on commonly used datasets showed C-RNNCrispr outperforms the state-of-the-art methods in terms of prediction accuracy and generalization. Source codes are available at https://github.com/Peppags/C_RNNCrispr.

Minimum variance-embedded deep kernel regularized least squares method for one-class classification and its applications to biomedical data

Deep kernel learning has been well explored for multi-class classification tasks; however, relatively less work is done for one-class classification (OCC). OCC needs samples from only one class to train the model. Most recently, kernel regularized least squares (KRL) method-based deep architecture is developed for the OCC task. This paper introduces a novel extension of this method by embedding minimum variance information within this architecture. This embedding improves the generalization capability of the classifier by reducing the intra-class variance. In contrast to traditional deep learning methods, this method can effectively work with small-size datasets. We conduct a comprehensive set of experiments on 18 benchmark datasets (13 biomedical and 5 other datasets) to demonstrate the performance of the proposed classifier. We compare the results with 16 state-of-the-art one-class classifiers. Further, we also test our method for 2 real-world biomedical datasets viz.; detection of Alzheimer’s disease from structural magnetic resonance imaging data and detection of breast cancer from histopathological images. Proposed method exhibits more than 5% F1 score compared to existing state-of-the-art methods for various biomedical benchmark datasets. This makes it viable for application in biomedical fields where relatively less amount of data is available.

Transform domain representation-driven convolutional neural networks for skin lesion segmentation

Automated diagnosis systems provide a huge improvement in early detection of skin cancer, and consequently, contribute to successful treatment. Recent research on convolutional neural network has achieved enormous success in segmentation and object detection tasks. However, these networks require large amount of data that is a big challenge in medical domain where often have insufficient data and even a pretrained model on medical images can be hardly found. Lesion segmentation as the initial step of skin cancer analysis remains a challenging issue since datasets are small and include a variety of images in terms of light, color, scale, and marks which have led researchers to use extensive augmentation and preprocessing techniques or fine tuning the network with a pretrained model on irrelevant images. A segmentation model based on convolutional neural networks is proposed in this study for the tasks of skin lesion segmentation and dermoscopic feature segmentation. The network is trained from scratch and despite the small size of datasets neither excessive data augmentation nor any preprocessing to remove artifacts or enhance the images are applied. Alternatively, we investigated incorporating image representations of the transform domain to the convolutional neural network and compared to a model with more convolutional layers that resulted in 6% higher Jaccard index and has shorter training time. The model improved by applying CIELAB color space and the performance of the final proposed architecture is evaluated on publicly available datasets from ISBI challenges in 2016 and 2017. The proposed model has resulted in an improvement of as much as 7% for the segmentation metrics and 17% for the feature segmentation, which demonstrates the robustness of this unique hybrid framework and its future applications as well as further improvement.

Redundancy Removed Dual-Tree Discrete Wavelet Transform to Construct Compact Representations for Automated Seizure Detection

With the development of pervasive sensing and machine learning technologies, automated epileptic seizure detection based on electroencephalogram (EEG) signals has provided tremendous support for the lives of epileptic patients. Discrete wavelet transform (DWT) is an effective method for time-frequency analysis of EEG and has been used for seizure detection in daily healthcare monitoring systems. However, the shift variance, the lack of directionality and the substantial aliasing, limit the effects of DWT in some applications. Dual-tree discrete wavelet transform (DTDWT) can overcome those drawbacks but may increase information redundancy. For classification tasks with small dataset sizes, such redundancy can greatly reduce learning efficiency and model performance. In this work, we proposed a novel redundancy removed DTDWT (RR-DTDWT) framework for automated seizure detection. Energy and modified multi-scale entropy (MMSE) features in a dual tree wavelet domain were extracted to construct a complete picture of mental states. To the best of our knowledge, this is the first study to employ MMSE as an indicator of epileptic seizures. Moreover, a compact EEG representation can be obtained after removing useless information redundancy (redundancy between wavelet trees, adjacent channels and entropy scales) by a general auto-weighted feature selection framework via global redundancy minimization (AGRM). Through validation on Bonn and CHB-MIT databases, the proposed RR-DTDWT method can achieve better performance than previous studies.

Sustainable development, economic and financial factors, that influence the opportunity-driven entrepreneurship. An fsQCA approach

There are two main reasons when entrepreneurs decide to start a new venture: opportunity and necessity. Opportunity-driven entrepreneurship is expected to provide a stronger long-term positive impact than necessity-driven entrepreneurship. This study aimed to identify the combinations of the economic and sustainable development factors of countries that may be related to opportunity-driven entrepreneurship. In order to identify the combinations of the Sustainable Development and Economic aspects influencing opportunity-driven entrepreneurship, we analysed the data for 2017 from 57 countries. For this purpose, we conducted a cross-national analysis using the fsQCA methodology, which has proved suitable for small-sized datasets. Data were retrieved from four databases: Global Entrepreneurship Monitor, Country Risk Score, World Bank Database, and Sustainable Development Goals Index. Thus to analyse opportunity-driven entrepreneurial motivation, we considered economic and financial aspects jointly with social and gender equality, education, responsible production, innovation and infrastructure indicators.

Medical Image Classification Based On Normalized Coding Network with Multiscale Perception

Medical imaging classification is playing a vital role in identifying and diagnoses the diseases, which is very helpful to doctor. Conventional ways classify supported the form, color, and/or texture, most of tiny problematic areas haven’t shown in medical images, which meant less efficient classification and that has poor ability to identify disease. Advanced deep learning algorithms provide an efficient way to construct a finished model that can compute final classification labels with the raw pixels of medical images. These conventional algorithms are not sufficient for high resolution images due to small dataset size, advanced deep learning models suffer from very high computational costs and limitations in the channels and multilayers in the channels. To overcome these limitations, we proposed a new algorithm Normalized Coding Network with Multi-scale Perceptron (NCNMP), which combines high-level features and traditional features. The Architecture of the proposed model includes three stages. Training, retrieve, fuse. We examined the proposed algorithm on medical image dataset NIH2626. We got an overall image classification accuracy of 91.35, which are greater than the present methods.