Explore unsupervised feature learning techniques using matrix decomposition in this comprehensive conference talk from ODSC East 2018. Delve into the world of unsupervised learning algorithms, focusing on dimension reduction and matrix factorization approaches for analyzing high-dimensional data without labeled classes. Gain insights into various matrix factorization techniques, including principal component analysis, correspondence analysis, non-negative matrix factorization, t-SNE, and autoencoders. Learn about extensions of these methods for simultaneous analysis of multiple datasets, such as canonical correlations analysis and multiple factor analysis. Discover how these approaches are applied to analyze tens of thousands of tumors, advancing precision medicine in oncology. The talk covers topics ranging from cancer microenvironment analysis to single-cell data analysis pipelines, classical dimension reduction techniques, and the integration of multiple datasets using advanced methods like multi-CIA and moGSA. Understand the application of these techniques in finding PanCancer Immune subtypes and exploring correlations between clusters, leucocyte fraction, and mutation load.
Unsupervised Feature Learning with Matrix Decomposition - Aedin Culhane, PhD | ODSC East 2018
Overview
Syllabus
Intro
Overview of Talk
Cancer Microenvironment, immune cells influence tumor progression, drug response
Many cell types
Exploratory data analysis (EDA)
Single Cell Data Analysis Pipeline
Classical Dimension Reduction Matrix Factorization approaches
Eigenvalues
Considerations when applying PCA
Correspondence Analysis
Multidimensional scaling (MDS)
Tensor Integration of 5 data sets (NC160) using multi-CIA
Reduce features to "groups of genes" to score get groups feature level single per case (moGSA)
Application of moGSA to finding PanCancer Immune subtypes
Correlation between 16 Clusters, leucocyte fraction and mutation load
Summary: multiple dataset integration
ENCODE
Taught by
Open Data Science
