“Deepfm: a factorization machine based neural network for CTR prediction” – paper abstract



The model is improved based on FM. In fact, NN is superimposed on FM.

  • Various sparse features are input
  • Next, go through the FM layer to get the implicit vector of features in the FM model
  • The input layer Z: z = (W0, Z1, Z2,…, Zn) of NN can be constructed through the hidden vector, and Zi = (WI, VI1, vi2,…, Vik), where wi is the hidden vector corresponding to the first-order weight VI in FM.
  • Finally, the NN layer is connected to get the final result.
  • Loss function: l (y, ˆ y) = −ylog ˆ y − (1 − y)log(1 − ˆ y)

The biggest advantage of FNN is that it does not need Feature Engineering, and its features are constructed by hidden vectors.

This work adoptsCC agreement, reprint must indicate the author and the link to this article

article!! Started on my blogStray_Camel(^U^)ノ~YO