Disambiguation Enabled Linear Discriminant Analysis for Partial Label Dimensionality Reduction

As an emerging weakly supervised learning framework, partial label learning considers inaccurate supervision where each training example is associated with multiple candidate labels among which only one is valid. In this article, a first attempt toward employing dimensionality reduction to help improve the generalization performance of partial label learning system is investigated. Specifically, the popular linear discriminant analysis (LDA) techniques are endowed with the ability of dealing with partial label training examples. To tackle the challenge of unknown ground-truth labeling information, a novel learning approach named Delin is proposed which alternates between LDA dimensionality reduction and candidate label disambiguation based on estimated labeling confidences over candidate labels. On one hand, the (kernelized) projection matrix of LDA is optimized by utilizing disambiguation-guided labeling confidences. On the other hand, the labeling confidences are disambiguated by resorting to k NN aggregation in the LDA-induced feature space. Extensive experiments over a broad range of partial label datasets clearly validate the effectiveness of Delin in improving the generalization performance of well-established partial label learning algorithms.

Identifying temporal and spatial patterns of variation from multimodal data using MEFISTO

Factor analysis is a widely used method for dimensionality reduction in genome biology, with applications from personalized health to single-cell biology. Existing factor analysis models assume independence of the observed samples, an assumption that fails in spatio-temporal profiling studies. Here we present MEFISTO, a flexible and versatile toolbox for modeling high-dimensional data when spatial or temporal dependencies between the samples are known. MEFISTO maintains the established benefits of factor analysis for multimodal data, but enables the performance of spatio-temporally informed dimensionality reduction, interpolation, and separation of smooth from non-smooth patterns of variation. Moreover, MEFISTO can integrate multiple related datasets by simultaneously identifying and aligning the underlying patterns of variation in a data-driven manner. To illustrate MEFISTO, we apply the model to different datasets with spatial or temporal resolution, including an evolutionary atlas of organ development, a longitudinal microbiome study, a single-cell multi-omics atlas of mouse gastrulation and spatially resolved transcriptomics.

NeuMapper: A scalable computational framework for multiscale exploration of the brain’s dynamical organization

For better translational outcomes researchers and clinicians alike demand novel tools to distil complex neuroimaging data into simple yet behaviorally relevant representations at the single-participant level. Recently, the Mapper approach from topological data analysis (TDA) has been successfully applied on noninvasive human neuroimaging data to characterize the entire dynamical landscape of whole-brain configurations at the individual level without requiring any spatiotemporal averaging at the outset. Despite promising results, initial applications of Mapper to neuroimaging data were constrained by (1) the need for dimensionality reduction, and (2) lack of a biologically grounded heuristic for efficiently exploring the vast parameter space. Here, we present a novel computational framework for Mapper—designed specifically for neuroimaging data—that removes limitations and reduces computational costs associated with dimensionality reduction and parameter exploration. We also introduce new meta-analytic approaches to better anchor Mapper-generated representations to neuroanatomy and behavior. Our new NeuMapper framework was developed and validated using multiple fMRI datasets where participants engaged in continuous multitask experiments that mimic “ongoing” cognition. Looking forward, we hope our framework could help researchers push the boundaries of psychiatric neuroimaging towards generating insights at the single-participant level while scaling across consortium-size datasets.

Discovering cell types using manifold learning and enhanced visualization of single-cell RNA-Seq data

Identifying relevant disease modules such as target cell types is a significant step for studying diseases. High-throughput single-cell RNA-Seq (scRNA-seq) technologies have advanced in recent years, enabling researchers to investigate cells individually and understand their biological mechanisms. Computational techniques such as clustering, are the most suitable approach in scRNA-seq data analysis when the cell types have not been well-characterized. These techniques can be used to identify a group of genes that belong to a specific cell type based on their similar gene expression patterns. However, due to the sparsity and high-dimensionality of scRNA-seq data, classical clustering methods are not efficient. Therefore, the use of non-linear dimensionality reduction techniques to improve clustering results is crucial. We introduce a method that is used to identify representative clusters of different cell types by combining non-linear dimensionality reduction techniques and clustering algorithms. We assess the impact of different dimensionality reduction techniques combined with the clustering of thirteen publicly available scRNA-seq datasets of different tissues, sizes, and technologies. We further performed gene set enrichment analysis to evaluate the proposed method’s performance. As such, our results show that modified locally linear embedding combined with independent component analysis yields overall the best performance relative to the existing unsupervised methods across different datasets.

An outlook: machine learning in hyperspectral image classification and dimensionality reduction techniques

Hyperspectral imaging is used in a wide range of applications. When used in remote sensing, satellites and aircraft are employed to collect the images, which are used in agriculture, environmental monitoring, urban planning and defence. The exact classification of ground features in the images is a significant research issue and is currently receiving greater attention. Moreover, these images have a large spectral dimensionality, which adds computational complexity and affects classification precision. To handle these issues, dimensionality reduction is an essential step that improves the performance of classifiers. In the classification process, several strategies have produced good classification results. Of these, machine learning techniques are the most powerful approaches. As a result, this paper reviews three different types of hyperspectral image machine learning classification methods: cluster analysis, supervised and semi-supervised classification. Moreover, this paper shows the effectiveness of all these techniques for hyperspectral image classification and dimensionality reduction. Furthermore, this review will assist as a reference for future research to improve the classification and dimensionality reduction approaches.

Physical-oriented and machine learning-based emission modeling in a diesel compression ignition engine: Dimensionality reduction and regression

The development of internal combustion engines is affected by the exhaust gas emissions legislation and the striving to increase performance. This demands for engine-out emission models that can be used for engine optimization for real driving emission controls. The prediction capability of physically and data-driven engine-out emission models is influenced by the system inputs, which are specified by the user and can lead to an improved accuracy with increasing number of inputs. Thereby the occurrence of irrelevant inputs becomes more probable, which have a low functional relation to the emissions and can lead to overfitting. Alternatively, data-driven methods can be used to detect irrelevant and redundant inputs. In this work, thermodynamic states are modeled based on 772 stationary measured test bench data from a commercial vehicle diesel engine. Afterward, 37 measured and modeled variables are led into a data-driven dimensionality reduction. For this purpose, approaches of supervised learning, such as lasso regression and linear support vector machine, and unsupervised learning methods like principal component analysis and factor analysis are applied to select and extract the relevant features. The selected and extracted features are used for regression by the support vector machine and the feedforward neural network to model the NOx, CO, HC, and soot emissions. This enables an evaluation of the modeling accuracy as a result of the dimensionality reduction. Using the methods in this work, the 37 variables are reduced to 25, 22, 11, and 16 inputs for NOx, CO, HC, and soot emission modeling while maintaining the accuracy. The features selected using the lasso algorithm provide more accurate learning of the regression models than the extracted features through principal component analysis and factor analysis. This results in test errors RMSETe for modeling NOx, CO, HC, and soot emissions 19.22 ppm, 6.46 ppm, 1.29 ppm, and 0.06 FSN, respectively.