Github user myui commented on the issue:

    https://github.com/apache/incubator-hivemall/pull/149
  
    @takuti 
    
    Linear term is not used in Libffm implementation. Better to do research 
about other FFM impl as well.
    https://github.com/chenhuang-learn/ffm/blob/master/ffm/src/ffm/FFMModel.java
    
https://github.com/superclocks/ffm/blob/master/libffm-ftrl-1.13/ffm-train.cpp
    https://github.com/chenhuang-learn/ffm
    https://github.com/gaterslebenchen/JLibFFM/
    https://github.com/yuantiku/ytk-learn
    https://github.com/RTBHOUSE/cuda-ffm/ (modified version of FFM)
    
    The default Initial learning rate effects largely to convergence as well.
    
    For instance-wise normalization, better to follow discussions in 
    https://markmail.org/message/jwtr5xygfutl55oz 
    It performs well only when all feature are categorical. For his dataset, 
instance-wise l2 normalization performed very badly...
    
    https://gist.github.com/myui/aaeef548a17eb90c4e88f824c3ca1bcd


---

Reply via email to