jackstraw: jackstraw: Statistical Inference for Unsupervised Learning

In ncchung/jackstraw: Statistical Inference for Unsupervised Learning

jackstraw: Statistical Inference for Unsupervised Learning

Description

Test for association between the observed data and their estimated latent variables. The jackstraw package provides a resampling strategy and testing scheme to estimate statistical significance of association between the observed data and their latent variables. Depending on the data type and the analysis aim, the latent variables may be estimated by principal component analysis (PCA), factor analysis (FA), K-means clustering, and related algorithms. The jackstraw methods learn over-fitting characteristics inherent in this circular analysis, where the observed data are used to estimate the latent variables and used again to test against that estimated latent variables. When latent variables are estimated by PCA, the jackstraw enables statistical testing for association between observed variables and latent variables, as estimated by low-dimensional principal components (PCs). This essentially leads to identifying variables that are significantly associated with PCs. Similarly, unsupervised clustering, such as K-means clustering, partition around medoids (PAM), and others, finds coherent groups in high-dimensional data. The jackstraw estimates statistical significance of cluster membership, by testing association between data and cluster centers. Clustering membership can be improved by using the resulting jackstraw p-values and posterior inclusion probabilities (PIPs), with an application to unsupervised evaluation of cell identities in single cell RNA-seq.

Details

The jackstraw package provides a resampling strategy and testing scheme to estimate statistical significance of association between the observed data and their latent variables. Depending on the data type and the analysis aim, the latent variables may be estimated by principal component analysis (PCA), K-means clustering, and related algorithms. The jackstraw methods learn over-fitting characteristics inherent in this circular analysis, where the observed data are used to estimate the latent variables and used again to test against those estimated latent variables.

The jackstraw tests enable us to identify the data features (i.e., variables or observations) that are driving systematic variation, in a completely unsupervised manner. Using jackstraw_pca, we can find statistically significant features with regard to the top r principal components. Alternatively, jackstraw_kmeans can identify the data features that are statistically significant members of the data-dependent clusters. Furthermore, this package includes more general algorithms such as jackstraw_subspace for the dimension reduction techniques and jackstraw_cluster for the clustering algorithms.

Overall, it computes m p-values of association between the m data features and their corresponding latent variables. From m p-values, pip computes posterior inclusion probabilities, that are useful for feature selection and visualization.

Author(s)

Neo Christopher Chung [e-mail]

References

Chung and Storey (2015) Statistical significance of variables driving systematic variation in high-dimensional data. Bioinformatics, 31(4): 545-554 https://academic.oup.com/bioinformatics/article/31/4/545/2748186

Chung (2020) Statistical significance of cluster membership for unsupervised evaluation of cell identities. Bioinformatics, 36(10): 3107–3114 https://academic.oup.com/bioinformatics/article/36/10/3107/5788523

See Also

jackstraw_pca jackstraw_subspace jackstraw_kmeans jackstraw_cluster

ncchung/jackstraw documentation built on Aug. 22, 2023, 12:12 p.m.