China
Africa
Explore chapters and articles related to this topic
Unsupervised Learning for Large Scale Data
Published in Ioannis S. Triantafyllou, Mangey Ram, Statistical Modeling of Reliability Structures and Industrial Processes, 2023
Petros Barmpas, Sotiris Tasoulis, Aristidis G. Vrahatis, Panagiotis Anagnostou, Spiros Georgakopoulos, Matthew Prina, José Luis Ayuso-Mateos, Jerome Bickenbach, Ivet Bayes, Martin Bobaki, Francisco Félix Caballero, Somnath Chatterji, Laia Egea-Cortés, Esther García-Esquinas, Matilde Leonardi, Seppo Koskinen, Ilona Koupil, Andrzej Pająk, Martin Prince, Warren Sanderson, Sergei Scherbov, Abdonas Tamosiunas, Aleksander Galas, Josep MariaHaro, Albert Sanchez-Niubo, Vassilis P. Plagianakos, Demosthenes Panagiotakos
Biomedical and health technologies are constantly evolving generating ultra-high dimensional data since we have several features for each record. Sampling techniques aim to reduce the dataset’s size but still do not offer a solution for high-dimensional datasets. In such cases, Dimensionality Reduction precedes Clustering procedures as a preprocessing step (Kaski 1998; Yan et al. 2006). Dimensionality Reduction (DR) aims to solve the Curse of Dimensionality (Bellman 1957) depicting that when the dimensionality increases, the volume of the space increases at such a rate that the dataset becomes sparse, opposing statistical methods. The goal is to find low-dimensional representations of the data that retain their fundamental properties, typically in two or three dimensions (Ghodsi 2006; Sorzano, Vargas and Montano 2014). As such this process is also essential for data visualization in lower dimensions (Xia et al. 2017). Visualization tools can assist in identifying the data structure while plotting the data in two dimensions allows researchers to pinpoint any remaining technical variability source between samples, which should be removed by normalization (Rostom et al. 2017).
Developing a preliminary cost estimation model for tall buildings based on machine learning
Published in Jiuping Xu, Syed Ejaz Ahmed, Zongmin Li, Big Data and Information Theory, 2022
Muizz O. Sanni-Anibire, Rosli Mohamad Zin, Sunday Olusanya Olatunji
Real-world data suffers from the ‘curse of dimensionality’, and thus, feature selection is required to determine the subset of independent features contributing to the predictive performance of a model. The ‘CorrelationAttributeEval’ technique in Weka’s ‘select attributes’ module was used. It is based on the ranking of the correlation of various features in the dataset to the prediction output, and further selection based on the Recursive Feature Elimination (RFE) process (Akande, Owolabi, & Olatunji, 2015). In RFE, the entire feature set (V) ranked according to the correlation coefficient is split in half to derive the best V/2 features, and the worst V/2 features are eliminated. The splitting process continues recursively until only one best feature is left. Thereafter, the feature subset that achieved the best accuracy/or the best performance measure is finally chosen as the best subset to be used.
Multivariate Statistics Neural Network Models
Published in Basilio de Bragança Pereira, Calyampudi Radhakrishna Rao, Fábio Borges de Oliveira, Statistical Learning Using Neural Networks, 2020
Basilio de Bragança Pereira, Calyampudi Radhakrishna Rao, Fábio Borges de Oliveira
An important reason to reduce the dimension of the data is that some authors call: “the curse of dimensionality and the empty space phenomenon”. The curse of dimensionality phenomenon refers to the fact that in the absence of simplifying assumptions, the sample size needed to make inferences with a given degree of accuracy grows exponentially with the number of variables. The empty space phenomenon responsible for the curse of dimensionality is that high-dimensional spaces are inherently sparse. For example, for a one dimensional standard normal N(0, 1), about 70% of the mass is at points contained in the interval (sphere of radius of one standard deviation around the mean zero). For a 10-dimensional N(0, I), the same (hyper) sphere contains only 0, 02% of the mass, and a radius of more than 3 standard deviations is needed to contain 70%.
Multi-modal speech emotion detection using optimised deep neural network classifier
Published in Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2023
Sweta Nishant Padman, Dhiraj Magare
In the recent research area related to data mining and the machine learning approach, feature selection is a most significant step that manages the curse of dimensionality issues and promotes the recognition percentage. The required features for emotion recognition are selected from the feature vector established for the input audio-video input using the optimised feature selection technique, the Learner Memorizing algorithm, which minimises the origin of extension, expenses, the extent of work, and the processing time through the selection of informative features. From the total of features, only features are selected such that the selected features are informative and significant out of the features in . are denoted as such that the features are reduced by the utilisation of optimised feature selection technique.
Deep-Wavelets and convolutional neural networks to support breast cancer diagnosis on thermography images
Published in Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2023
Valter Augusto de Freitas Barbosa, Anderson Félix da Silva, Maíra Araújo de Santana, Rian Rabelo de Azevedo, Rita de Cássia Fernandes de Lima, Wellington Pinheiro dos Santos
We performed experiments with DWNN considering 2, 4 and 6 hidden layers. Each dataset was classified using SVM and ELM. For SVM, we used linear and RBF kernels. We varied gamma parameter for RBF kernel between the values 0.01, 0.25 and 0.50. ELM experiments were executed with 500 neurons in hidden layer and polynomial kernels of degree from 1 to 5. This experiments aimed to find a good settings for DWNN for the problem. For the datasets generated by feature extraction using CNNs, we performed the classification by using SVM with linear kernel. Each setup was run 30 times. So we can analyse statistically significant results. Since we have a reduced number of images, we employed 10-fold cross-validation with the aim to preventing overfitting. As can seen in Table 2, the six-layer DWNN and CNNs furnished a huge amount of attributes. Therefore, we performed a feature selection for their datasets, in order to avoid the curse of dimensionality. Attribute selection was executed by a random forest with 1000 trees. The number of attributes selected is given in Table 2.
Alcoholic EEG Signal Classification Using Multi-Heuristic Classifiers with Stochastic Gradient Descent Technique for Tuning the Hyperparameters
Published in IETE Journal of Research, 2023
Harikumar Rajaguru, A. Vigneshkumar, M. Gowri Shankar
Previously, neurophysiologists would interrupt EEG signals by visually analysing them. Later, it was possible to distinguish between normal and abnormal EEG patterns. This procedure also has a significant cost [8]. Alcoholic EEG features are non-linear, dynamic, and non-stationary features with a high degree of non-linearity [9]. EEG generates stochastic waves with variable amplitude and frequency. Alcoholic EEG features are examined in time-domain, frequency domain otherwise spectral domain, and matrix decomposition method for better pattern recognition purpose [10]. Electroencephalographers who used to train and recognize problems in EEG signals perform clinical EEG evaluation directly. One of the most well-known limitations in machine learning is the curse of dimensionality. Large dimensions may not necessarily give us relevant patterns about a underlying physiological disturbances in the brain [11]. EEG based on the Brain-Computer Interface (BCI) is an auspicious research strategy for machine learning approaches.