For a p dimension example x, FM models the data by
where w_i denotes by the i-th element of the p-length vector w, and v_i denotes by the i-th row of the p-by-k matrix V.
Given training data pairs (x,y), FM learns the model w and V by solving the following objective:
Here ℓ is the loss function such as logistic loss.
Go the root directory of wormhole, then build by make difacto. Next try
a small dataset using 2 worker and 1 server:
tracker/dmlc_local.py -n 2 -s 1 bin/difacto.dmlc learn/difacto/guide/demo.conf