github.com/bigtong/hector

hector

Golang machine learning lib. Currently, it can be used to solve binary classification problems.

Supported Algorithms

• Logistic Regression
• Factorized Machine
• CART, Random Forest, Random Decision Tree, Gradient Boosting Decision Tree
• Neural Network

Dataset Format

Hector support libsvm-like data format. Following is an sample dataset

1 	1:0.7 3:0.1 9:0.4
0	2:0.3 4:0.9 7:0.5
0	2:0.7 5:0.3
...

How to Run

Run as tools

In src folder, you will find two program with main function : hector-cv.go and hector-run.go

hector-cv.go will help you test one algorithm by cross validation in some dataset, you can run it by following steps:

cd src
go build hector-cv.go
./hector-cv --method [Method] --train [Data Path] --cv 10

Here, Method include

• lr : logistic regression with SGD and L2 regularization.
• ftrl : FTRL-proximal logistic regreesion with L1 regularization. Please review this paper for more details "Ad Click Prediction: a View from the Trenches".
• ep : bayesian logistic regression with expectation propagation. Please review this paper for more details "Web-Scale Bayesian Click-Through Rate Prediction for Sponsored Search Advertising in Microsoft’s Bing Search Engine"
• fm : factorization machine
• cart : classifiaction tree
• cart-regression : regression tree
• rf : random forest
• rdt : random decision trees
• gbdt : gradient boosting decisio tree
• linear-svm : linear svm with L1 regularization
• svm : svm optimizaed by SMO (current, its linear svm)
• l1vm : vector machine with L1 regularization by RBF kernel
• knn : k-nearest neighbor classification

hector-run.go will help you train one algorithm on train dataset and test it on test dataset, you can run it by following steps:

cd src
go build hector-run.go
./hector-run --method [Method] --train [Data Path] --test [Data Path]

Above methods will direct train algorithm on train dataset and then test on test dataset. If you want to train algorithm and get the model file, you can run it by following steps:

./hector-run --method [Method] --action train --train [Data Path] --model [Model Path]

Then, you can use model file to test any test dataset:

./hector-run --method [Method] --action test --test [Data Path] --model [Model Path]

Benchmark

Binary Classification

Following are datasets used in benchmarks:

• heart
• fourclass

I will do 5-fold cross validation on the dataset, and use AUC as evaluation metric. Following are the results:

DataSet	Method	AUC
heart	FTRL-LR	0.9109
heart	EP-LR	0.8982
heart	CART	0.8231
heart	RDT	0.9155
heart	RF	0.9019
heart	GBDT	0.9061
fourclass	FTRL-LR	0.8281
fourclass	EP-LR	0.7986
fourclass	CART	0.9832
fourclass	RDT	0.9925
fourclass	RF	0.9947
fourclass	GBDT	0.9958