Closed-Form Models of Accuracy Loss due to Subsampling in SVD Collaborative Filtering

We postulate and analyze a nonlinear subsampling accuracy loss (SSAL) model based on the root mean square error (RMSE) two SSAL models (MSE), suggested by extensive preliminary simulations. The predict in terms of parameters like fraction users dropped (FUD) items (FID). seek to investigate whether depend characteristics dataset constant way across datasets when using SVD collaborative filtering (CF) algorithm. considered include various densities rating matrix numbers items. Extensive simulations rigorous regression analysis led empirical symmetrical FID FUD whose coefficients only data characteristics. came out be multi-linear odds ratios dropping user (or an item) vs. not it. Moreover, one MSE deterioration turned linear where their interaction term has zero coefficient. Most importantly, are sense that they written closed-form (densities items). validated through 850 synthetically generated primary (pre-subsampling) matrices derived from 25M MovieLens dataset. Nearly 460 000 subsampled were then simulated subjected singular value decomposition (SVD) CF Further validation was conducted 1M Yahoo! Music Rating datasets. significant all 3