Fast Matrix-Vector Multiplications for Large-Scale Logistic Regression on Shared-Memory Systems

Shared-memory systems such as regular desktopsnow possess enough memory to store large data. However, thetraining process for data classification can still be slow if we donot fully utilize the power of multi-core CPUs. Many existingworks proposed parallel machine learning algorithms by modifyingserial ones, but convergence analysis may be complicated. Instead, we do not modify machine learning algorithms, butconsider those that can take the advantage of parallel matrixoperations. We particularly investigate the use of parallel sparsematrix-vector multiplications in a Newton method for largescalelogistic regression. Various implementations from easy tosophisticated ones are analyzed and compared. Results indicatethat under suitable settings excellent speedup can be achieved.