Solvers for Hierarchical Classification

Description

This repo currently includes the following folders:

• data: folder to hold experimental data, currently only have diatoms and reuters, both are taken from [1].
• util: contains code for preprocessing and evalution.
• code: proximal-newton, proximal-gradient and proximal-sgd solver for hierarchical classification
• experimental_result: hold some output and figures.

Quick Start

Run the following code to load data:

julia > include("code/readData.jl")

Load data matrix X, label matrix Y and labels' adjacency matrix A.

julia > X, Y = read("data/rcv1/train_remap.txt");
julia > K = size(Y, 1)
julia > A = read_cat_hier("data/rcv1/hr_remap.txt",K)
julia > Y = Y'

Add bias term to our data matrix X.

julia > N, D = size(X)
julia > X = [ones(N) X]
julia > K = size(Y,2)

Set hyper-parameters

julia > C = 1; lambda = 0.1

Run proximal-newton:

julia > include("code/prox_newton.jl")
julia > W = proximal_newton(X, Y, A, C, lambda,1e-4, 1e-4,200)

To run proximal-gradient:

julia > include("code/prox_newton.jl")
julia > W = proximal_gradient(X, Y, A, C, lambda, 1e-4, 300, 20)

To run proximal-SGD

julia > include("code/SGD.jl")
julia > W_res = mainStochastic(X, Y, K, A, C, lambda,stepsize=(i)->1e-8, maxIter=2e4)

References: [1] HR-SVM Project, https://sites.google.com/site/hrsvmproject/datasets-hier