Fast Data Reduction With Granulation-Based Instances Importance Labeling

THINGS-data reflects three large-scale neuroimaging and behavioral datasets of object processing in humans, comprising densely sampled functional MRI and magnetoencephalographic recordings, as well as 4.70 million similarity judgments in response to thousands of photographic images for up to 1854 objects.

In recent years, Deep Learning has gained popularity for its ability to solve complex classification tasks, increasingly delivering better results thanks to the development of more accurate models, the availability of huge volumes of data and the improved computational capabilities of modern computers. However, these improvements in performance also bring efficiency problems, related to the storage of datasets and models, and to the waste of energy and time involved in both the training and inference processes. In this context, data reduction can help reduce energy consumption when training a deep learning model. In this paper, we present up to eight different methods to reduce the size of a tabular training dataset, and we develop a Python package to apply them. We also introduce a representativeness metric based on topology to measure how similar are the reduced datasets and the full training dataset. Additionally, we develop a methodology to apply these data reduction methods to image datasets for object detection tasks. Finally, we experimentally compare how these data reduction methods affect the representativeness of the reduced dataset, the energy consumption and the predictive performance of the model.

High dimensionality data have become common in neuroimaging fields, especially group‐level functional magnetic resonance imaging (fMRI) datasets. fMRI connectivity analysis is a widely used, powerful technique for studying functional brain networks to probe underlying mechanisms of brain function and neuropsychological disorders. However, data‐driven technique like independent components analysis (ICA), can yield unstable and inconsistent results, confounding the true effects of interest and hindering the understanding of brain functionality and connectivity. A key contributing factor to this instability is the information loss that occurs during fMRI data reduction. Data reduction of high dimensionality fMRI data in the temporal domain to identify the important information within group datasets is necessary for such analyses and is crucial to ensure the accuracy and stability of the outputs. In this study, we describe an fMRI data reduction strategy based on an adapted neighborhood preserving embedding (NPE) algorithm. Both simulated and real data results indicate that, compared with the widely used data reduction method, principal component analysis, the NPE‐based data reduction method (a) shows superior performance on efficient data reduction, while enhancing group‐level information, (b) develops a unique stratagem for selecting components based on an adjacency graph of eigenvectors, (c) generates more reliable and reproducible brain networks under different model orders when the outputs of NPE are used for ICA, (d) is more sensitive to revealing task‐evoked activation for task fMRI, and (e) is extremely attractive and powerful for the increasingly popular fast fMRI and very large datasets.

An important factor affecting the rate of BCI is the number of EEG features. To reduce the number of features is an important way to improve the speed. In this paper, a method of data reduction be described, features marked be used to discrete the continuous EEG, and then choose the features from the discrete data with the help of this method. The results show that classification accuracy has not been reduced but the number of features is reduction.

The Internet of Things (IoT) is expected to grow faster than any other category of connected devices. IoT allows any device with an on-and-off switch to connect to the internet—a concept that has the ability to greatly change our lives and work. These modern systems collect inherently complex data streams due to the volume, velocity, value, variety, variability, and veracity of data, which leads to incremental growth in data traffic on networks and in the cloud. To fulfill the requirements of IoT, including geodistribution, low latency, location awareness, and mobility support, a new paradigm is proposed: edge computing. In edge computing, substantial computing and storage resources are placed at the edge of the network in mobile devices or sensors. The term “edge” is taken from network diagrams; normally, the edge of a network diagram represents the point at which data traffic enters or leaves the workable network. Using the concept of edge computing, an organization can shift huge amounts of data into processed data near the data origin, which helps to reduce data traffic in the network’s central repository (called the “cloud”). Edge computing uses a variety of data reduction techniques close to the data source at the network edge, including data pre-processing, local storage, and filtering. This approach can prevent some critical issues, such as I/O bottlenecks, storage and bandwidth limitations, data traffic increments, and high energy costs. A major advantage of edge computing is improvement of the request-response delay to milliseconds. Edge computing also supports security and network challenges. However, two major obstacles exist toward achieving the benefit of network-edge computing. First, the most efficient algorithms for data reduction in time series (one of the most common types of data in IoT) were developed to work posteriori upon big datasets, but they cannot make decisions for each incoming data item. Secondly, the state of the art lacks systems that can apply any of the possible data reduction methods without adding significant delays or major reconfigurations. Edge computing has also inherited some of the challenges of cloud computing, including data abstraction, naming, and programmability. This chapter presents a detailed taxonomic discussion of edge computing, along with its challenges, opportunities, and data reduction methods.

Computational time in optimization models scales with the number of time steps. To save time, solver time resolution can be reduced and input data can be down-sampled into representative periods such as one or a few representative days per month. However, such data reduction can come at the expense of solution accuracy. In this work, the impact of reduction of input data is systematically isolated considering an optimization which solves an energy system using representative days. A new data reduction method aggregates annual hourly demand data into representative days which preserve demand peaks in the original profiles. The proposed data reduction approach is tested on a real energy system and real annual hourly demand data where the system is optimized to minimize total annual costs. Compared to the full-resolution optimization of the energy system, the total annual energy cost error is found to be equal or less than 0.22% when peaks in customer demand are preserved. Errors are significantly larger for reduction methods that do not preserve peak demand. Solar photovoltaic data reduction effects are also analyzed. This paper demonstrates a need for data reduction methods which consider demand peaks explicitly.

As the number of pre-trained machine learning (ML) models is growing exponentially, data reduction tools are not catching up. Existing data reduction techniques are not specifically designed for pre-trained model (PTM) dataset files. This is largely due to a lack of understanding of the patterns and characteristics of these datasets, especially those relevant to data reduction and compressibility. This paper presents the first, exhaustive analysis to date of PTM datasets on storage compressibility. Our analysis spans different types of data reduction and compression techniques, from hash-based data deduplication, data similarity detection, to dictionary-coding compression. Our analysis explores these techniques at three data granularity levels, from model layers, model chunks, to model parameters. We draw new observations that indicate that modern data reduction tools are not effective when handling PTM datasets. There is a pressing need for new compression methods that take into account PTMs' data characteristics for effective storage reduction. Motivated by our findings, we design Elf, a simple yet effective, error-bounded, lossy floating-point compression method. Elf transforms floating-point parameters in such a way that the common exponent field of the transformed parameters can be completely eliminated to save storage space. We develop Elves, a compression framework that integrates Elf along with several other data reduction methods. Elves uses the most effective method to compress PTMs that exhibit different patterns. Evaluation shows that Elves achieves an overall compression ratio of 1.52×, which is 1.31×, 1.32× and 1.29× higher than a general-purpose compressor (zstd), an error-bounded lossy compressor (SZ3), and the uniform model quantization, respectively, with negligible model accuracy loss.

Instance selection and feature extraction is one of the most important task in data mining, due to the huge amount of data is constantly being produced in many fields. If the dataset is very large means most of the existing machine learning algorithms are inapplicable to handle such huge amount of data and computational cost is high. Two of the approaches have been used for solving this problem. One is scaling up algorithms and another one is data reduction. Scaling up data mining algorithm is not always feasible, but data reduction is possible. In this paper we take both, instance selection and feature extraction for data reduction. Instance selection is a technique that will reduce the size of the original training data. Feature extraction is input data having m dimension space that should be mapped into lower dimension space i.e., eliminate those components which are contributing less information. In this paper Cuttlefish optimization algorithm is used for instance selection, while principal component analysis is used for feature extraction. The combination of feature extraction and instance selection will reduce the large amount of computational time of training the classifiers. The optimal extracted subset of data points and reduced feature space are providing almost similar detection rate, accuracy rate, false positive rate and takes less amount of computational time for training the classifiers what we obtained from using original dataset.

Over the recent decades, the amount of data generated has been growing exponentially, the existing machine learning algorithms are not feasible for processing of such huge amount of data. To solve such kind of issues, we have two commonly adopted schemes, one is scaling up the data mining algorithms and other one is data reduction. Scaling up the data mining algorithms is not a best way, but data reduction is fairly possible. In this paper, cuttlefish optimisation algorithm along with tabu search approach is used for data reduction. Dataset can be reduced mainly in two ways, one is the selecting optimal subset of features from the original dataset, in other words eliminating those features which are contributing lesser information another method is selecting optimal subset of instances from the original data set, in other words eliminating those instances which are contributing lesser information. Cuttlefish optimisation algorithm with tabu search finds both optimal subset of features and instances. Optimal subset of feature and instance obtained from the cuttlefish algorithm with tabu search provides a similar detection rate, accuracy rate, lesser false positive rate and the lesser computational time for training the classifier that we obtained from the original data set.

Despite the importance of data reduction as part of the processing of reflection-based classifications, this study represents one of the first in which the effects of both spatial and spectral data reductions on classification accuracies are quantified. Furthermore, the effects of approaches to data reduction were quantified for two separate classification methods, linear discriminant analysis (LDA) and support vector machine (SVM). As the model dataset, reflection data were acquired using a hyperspectral camera in 230 spectral channels from 401 to 879 nm (spectral resolution of 2.1 nm) from field pea (Pisum sativum) samples with and without internal pea weevil (Bruchus pisorum) infestation. We deployed five levels of spatial data reduction (binning) and eight levels of spectral data reduction (40 datasets). Forward stepwise LDA was used to select and include only spectral channels contributing the most to the separation of pixels from non-infested and infested field peas. Classification accuracies obtained with LDA and SVM were based on the classification of independent validation datasets. Overall, SVMs had significantly higher classification accuracies than LDAs (P < 0.01). There was a negative association between pixel resolution and classification accuracy, while spectral binning equivalent to up to 98% data reduction had negligible effect on classification accuracies. This study supports the potential use of reflection-based technologies in the quality control of food products with internal defects, and it highlights that spatial and spectral data reductions can (1) improve classification accuracies, (2) vastly decrease computer constraints, and (3) reduce analytical concerns associated with classifications of large and high-dimensional datasets.

Persistent homology is used for computing topological features of a space at different spatial resolutions. It is one of the main tools from computational topology that is applied to the problems of data analysis. Despite several attempts to reduce its complexity, persistent homology remains expensive in both time and space. These limits are such that the largest data sets to which the method can be applied have the number of points of the order of thousands in ℝ3. This paper explores a technique intended to reduce the number of data points while preserving the salient topological features of the data. The proposed technique enables the computation of persistent homology on a reduced version of the original input data without affecting significant components of the output. Since the run time of persistent homology is exponential in the number of data points, the proposed data reduction method facilitates the computation in a fraction of the time required for the original data. Moreover, the data reduction method can be combined with any existing technique that simplifies the computation of persistent homology. The data reduction is performed by creating small groups of similar data points, called nano-clusters, and then replacing the points within each nano-cluster with its cluster center. The persistence homology of the reduced data differs from that of the original data by an amount bounded by the radius of the nano-clusters. The theoretical analysis is backed by experimental results showing that persistent homology is preserved by the proposed data reduction technique.

In this paper, we propose a new SQL based incremental distributed algorithm for predicting the next location of a mobile user in a mobile web environments. Parallel and Distributed data mining algorithm is applied on moving logs stored in geographically distributed data grid to generate the mobility pattern, which provides various location based services to the mobile users. One of the existing works for deriving mobility pattern is re-executing the algorithm from scratch results in excessive computation. In our work, we have designed new incremental algorithm by maintaining infrequent mobility patterns, which avoids unnecessary scan of full database. We built data grid system on a cluster of workstation using open source Globus Toolkit (GT) and Message Passing Interface extended with Grid Services (MPICH-G2). The experiments were conducted on original data sets with incremental addition of data and the computation time was recorded for each data sets. We analyzed our results with various sizes of data sets and it shows the time taken to generate mobility pattern by incremental mining algorithm is less than re-computing approach.

The FORS1 instrument on the ESO Very Large Telescope was used to obtain low-resolution circular polarised spectra of nearly a thousand different stars, with the aim of measuring their mean longitudinal magnetic fields. A catalogue of FORS1 magnetic measurements would provide a valuable resource with which to better understand the strengths and limitations of this instrument and of similar low-dispersion, Cassegrain spectropolarimeters. However, FORS1 data reduction has been carried out by a number of different groups using a variety of reduction and analysis techniques. Our understanding of the instrument and our data reduction techniques have both improved over time. A full re-analysis of FORS1 archive data using a consistent and fully documented algorithm would optimise the accuracy and usefulness of a catalogue of field measurements. Based on the ESO FORS pipeline, we have developed a semi-automatic procedure for magnetic field determinations, which includes self-consistent checks for field detection reliability. We have applied our procedure to the full content of circular spectropolarimetric measurements of the FORS1 archive. We have produced a catalogue of spectro-polarimetric observations and magnetic field measurements for about 1400 observations of about 850 different objects. The spectral type of each object has been accurately classified. We have also been able to test different methods for data reduction is a systematic way. The resulting catalogue has been used to produce an estimator for an upper limit to the uncertainty in a field strength measurement of an early type star as a function of the signal-to-noise ratio of the observation. While FORS1 is not necessarily an optimal instrument for the discovery of weak magnetic fields, it is very useful for the systematic study of larger fields, such as those found in Ap/Bp stars and in white dwarfs.

This paper focuses on nowcasts of tail risk to GDP growth, with a potentially wide array of monthly and weekly information. We consider different models (Bayesian mixed frequency regressions with stochastic volatility, as well as classical and Bayesian quantile regressions) and also different methods for data reduction (either forecasts from models that incorporate data reduction or the combination of forecasts from smaller models). Our results show that, within some limits, more information helps the accuracy of nowcasts of tail risk to GDP growth. Accuracy typically improves as time moves forward within a quarter, making additional data available, with monthly data more important to accuracy than weekly data. Accuracy also typically improves with the use of financial indicators in addition to a base set of macroeconomic indicators. The better-performing models or methods include the Bayesian regression model with stochastic volatility, Bayesian quantile regression, some approaches to data reduction that make use of factors, and forecast averaging. In contrast, simple quantile regression performs relatively poorly.

Performing ground-based submillimetre observations is a difficult task as the measurements are subject to absorption and emission from water vapour in the Earth's atmosphere and time variation in weather and instrument stability. Removing these features and other artefacts from the data is a vital process which affects the characteristics of the recovered astronomical structure we seek to study. In this paper, we explore two data reduction methods for data taken with the Submillimetre Common-User Bolometer Array-2 (SCUBA-2) at the James Clerk Maxwell Telescope (JCMT). The JCMT Legacy Reduction 1 (JCMT LR1) and The Gould Belt Legacy Survey Legacy Release 1 (GBS LR1) reduction both use the same software (STARLINK) but differ in their choice of data reduction parameters. We find that the JCMT LR1 reduction is suitable for determining whether or not compact emission is present in a given region and the GBS LR1 reduction is tuned in a robust way to uncover more extended emission, which better serves more in-depth physical analyses of star-forming regions. Using the GBS LR1 method, we find that compact sources are recovered well, even at a peak brightness of only three times the noise, whereas the reconstruction of larger objects requires much care when drawing boundaries around the expected astronomical signal in the data reduction process. Incorrect boundaries can lead to false structure identification or it can cause structure to be missed. In the JCMT LR1 reduction, the extent of the true structure of objects larger than a point source is never fully recovered.