mlpack  blog
Collaborative filtering package improvements - Week 3 + 4

Collaborative filtering package improvements - Week 3 + 4

Sumedh Ghaisas, 19 August 2014

Week 4 is over. Had fun with segmentation faults and undecodable generic programming errors. :) In past week the following things have been added to the AMF module -

1) The stopping criterion is shifted from minResidue to tolerance measurement between two iteration. This Helps terminate both NMF and SVD factorizations. 2) Each call to AMF apply (actual factorization function) will call Initialize on given update rule before starting the procedure of alternating updation. This will allow update rules to have local variables. 3) Added SVDBatchLearning update rule. Tested on MovieLens and matrices with neagative entries. 4) Added momentum to BatchLearning which decreases its time overhead while maintaining the efficiency.

Next week's task will be writing rigorous tests for this module and benchmarking it with known results. Thank you for visiting...