Contents

function [z, history] = distr_l1_logreg(A, b, mu, N, rho, alpha)
% distr_l1_logreg   Solve distributed L1 regularized logistic regression
%
% [x, history] = distr_l1_logreg(A, b, mu, N, rho, alpha)
%
% solves the following problem via ADMM:
%
%   minimize   sum( log(1 + exp(-b_i*(a_i'w + v)) ) + m*mu*norm(w,1)
%
% where A is a feature matrix and b is a response vector. The scalar m is
% the number of examples in the matrix A.
%
% This solves the L1 regularized logistic regression problem. It uses
% a custom Newton solver for the x-step. This version solves a distributed
% version of L1 regularized logistic regression.
%
% The solution is returned in the vector x = (v,w).
%
% history is a structure that contains the objective value, the primal and
% dual residual norms, and the tolerances for the primal and dual residual
% norms at each iteration.
%
% N is the number of subsystems to use to split the examples. This code
% will (serially) solve N x-updates with m / N examples per subsystem.
% Therefore, the number of examples, m, should be divisible by N. No error
% checking is done.
%
% rho is the augmented Lagrangian parameter.
%
% alpha is the over-relaxation parameter (typical values for alpha are
% between 1.0 and 1.8).
%
% This example requires the "MATLAB Interface for L-BFGS-B" and L-BFGS-B
% installed. These can be acquiured at
% http://www.cs.ubc.ca/~pcarbo/lbfgsb-for-matlab.html.
%
%
% More information can be found in the paper linked at:
% http://www.stanford.edu/~boyd/papers/distr_opt_stat_learning_admm.html
%


t_start = tic;

Global constants and defaults

QUIET    = 0;
MAX_ITER = 1000;
ABSTOL   = 1e-4;
RELTOL   = 1e-2;

Preprocessing

[m, n] = size(A);
m = m / N;  % should be divisible

ADMM solver

x = zeros(n+1,N);
z = zeros(n+1,N);
u = zeros(n+1,N);


if ~QUIET
    fprintf('%3s\t%10s\t%10s\t%10s\t%10s\t%10s\t%10s\n', 'iter', '# bfgs', ...
      'r norm', 'eps pri', 's norm', 'eps dual', 'objective');
end
p = size(z,1);
C = [-b -A]';

global BFGS_ITERS;  % global variable to keep track of bfgs iterations
bfgs_iters = zeros(N,1);

for k = 1:MAX_ITER

    % serial x-update
    for i = 1:N,
        K = C(:,1+(i-1)*m:i*m)';
        x(:,i) = bfgs_update(K, u(:,i), z(:,i), rho, N, x(:,i));
        bfgs_iters(i) = BFGS_ITERS;
    end

    % z-update with relaxation
    zold = z;
    x_hat = alpha*x + (1-alpha)*zold;
    ztilde = mean(x_hat + u,2);
    ztilde(2:end) = shrinkage( ztilde(2:end), (m*N)*mu/(rho*N) );

    z = ztilde*ones(1,N);

    % u-update
    u = u + (x_hat - z);

    % diagnostics, reporting, termination checks
    history.objval(k)  = objective(A, b, mu, x, z(:,1));

    history.r_norm(k)  = norm(x - z, 'fro');
    history.s_norm(k)  = norm(rho*(z - zold),'fro');

    history.LBFGS_iters(:,k) = bfgs_iters;

    history.eps_pri(k) = sqrt(p*N)*ABSTOL + RELTOL*max(norm(x,'fro'), norm(z,'fro'));
    history.eps_dual(k)= sqrt(p*N)*ABSTOL + RELTOL*norm(rho*u,'fro');

    if ~QUIET
        fprintf('%3d\t%10d\t%10.4f\t%10.4f\t%10.4f\t%10.4f\t%10.2f\n', k, sum(bfgs_iters), ...
            history.r_norm(k), history.eps_pri(k), ...
            history.s_norm(k), history.eps_dual(k), history.objval(k));
    end


    if history.r_norm(k) < history.eps_pri(k) && ...
       history.s_norm(k) < history.eps_dual(k)
        break;
    end
end

if ~QUIET
    toc(t_start);
end
z = z(:,1);
end

function obj = objective(A, b, mu, x, z)
    m = size(A,1);
    obj = sum(log(1 + exp(-A*z(2:end) -b*z(1)))) + m*mu*norm(z(2:end),1);
end

function [x t] = bfgs_update(C, u, z, rho, N, x0)
    % solve the x update
    %   minimize [ -logistic(x_i) + (rho/2)||x_i - z^k + u^k||^2 ]
    % via L-BFGS

    [m n] = size(C);

    auxdata{1} = C;
    auxdata{2} = z;
    auxdata{3} = u;
    auxdata{4} = rho;

    x = lbfgsb(x0, -Inf*ones(n,1), +Inf*ones(n,1), 'l2_log', 'l2_log_grad', auxdata, 'record_bfgs_iters');
end


function z = shrinkage(a, kappa)
    z = max(0, a-kappa) - max(0, -a-kappa);
end