Comparing Matrix Decomposition Methods for Meta-Analysis and Reconstruction of Cognitive Neuroscience Results

摘要

The results of 2,256 neuroimaging experiments were analyzed using singular value decomposition (SVD) and non-negative matrix factorization (NMF) to extract patterns in the data. To evaluate the techniques' efficacy at capturing regularities in the data, one positive and one negative result from each of 100 random experiments were treated as missing, and the values were iteratively reconstructed using each technique for dimensionality reduction. Under the best conditions, precision and recall of roughly 78% was achieved for each method. Weighting the domain matrix and area matrix to have equal first eigenvalues before combining them, a technique known as blending, significantly improved results for both methods. While using unnormalized data appeared to produce a peak in results for 10-15 dimensions, normalizing to take into account variation in the popularity of experiment types removed the effect. The basis vectors produced by each method do not support the idea that current cognitive ontologies map well to individual brain areas.