linearModel

LinearModel provide high-performance machine learning algorithoms written by golang. It contains logistic regression, factorization machine, field aware factorization machine.

Requirement

golang >= 1.18

benchmark

Criteo dataset benchmark

cd benchmark
bash run.sh

model	vcore	memory	time	auc
lr	30	16g	2m20s	0.794
ffm(8 field)	30	64g	8m30s	0.809
ffm(39 field)	30

TODO

1. feature count filter

Usage

build

git clone https://github.com/wanpengxie/linearmodel.git linearmodel
cd linearmodel && go build -o trainer main/trainer.go    

train

# model_name in (lr, fm, ffm), parallel means thread numbers being used
./trainer -v=3 -logtostderr -conf ${conf_path} -parallel 10 -model ${model_name}   

save model to file

# save_path='/tmp/model'
./trainer -v=3 -logtostderr -conf ${conf_path} -parallel 10 -model ${model_name}  -save ${save_path}

load

# load_path='/tmp/model'
./trainer -v=3 -logtostderr -conf ${conf_path} -parallel 10 -model ${model_name}  -load ${load_path}

Config file format

We using protobuf for config file, examples

optim_config {
  l1: 0.01
  l2: 0.02
  alpha: 2.0
  beta: 1.0
  emb_l1: 0.00
  emb_l2: 0.8
  emb_alpha: 0.8
  emb_beta: 1.0
  emb_size: 12
}

feature_list {
  name: "UserId"
  slot_id: 101
  vec_type: LEFT
  cross: 1
}

feature_list {
  name: "ItemId"
  slot_id: 102
  vec_type: RIGHT
  cross: 2
}

is_feature_signed: false
train_path_list:"path_to_train_data1"
train_path_list:"path_to_train_data2"
predict_path_list:"path_to_predict_data1"
predict_path_list:"path_to_predict_data2"

Data file format

Using libsvm dataformat , examples

1   101:user1 102:item1 103:8 104:wednesday
0   101:user2 102:item2 103:9 104:wednesday