AnyValue LogisticRegressionIRLS::transition(AbstractDBInterface &db,
AnyValue args) {
AnyValue::iterator arg(args);
// Initialize Arguments from SQL call
State state = *arg++;
double y = *arg++ ? 1. : -1.;
DoubleRow_const x = *arg++;
// See MADLIB-138. At least on certain platforms and with certain versions,
// LAPACK will run into an infinite loop if pinv() is called for non-finite
// matrices. We extend the check also to the dependent variables.
if (!boost::math::isfinite(y))
throw std::invalid_argument("Dependent variables are not finite.");
else if (!x.is_finite())
throw std::invalid_argument("Design matrix is not finite.");
if (state.numRows == 0) {
state.initialize(db.allocator(AbstractAllocator::kAggregate), x.n_elem);
if (!arg->isNull()) {
const State previousState = *arg;
state = previousState;
state.reset();
}
}
// Now do the transition step
state.numRows++;
// xc = x_i c
double xc = as_scalar( x * state.coef );
// a_i = sigma(x_i c) sigma(-x_i c)
double a = sigma(xc) * sigma(-xc);
// Note: sigma(-x) = 1 - sigma(x).
//
// sigma(-y_i x_i c) y_i
// z = x_i c + ---------------------
// a_i
double z = xc + sigma(-y * xc) * y / a;
state.X_transp_Az += trans(x) * a * z;
state.X_transp_AX += trans(x) * a * x;
// We use state.sumy to store the log likelihood.
// n
// --
// l(c) = -\ ln(1 + exp(-y_i * c^T x_i))
// /_
// i=1
state.logLikelihood -= std::log( 1. + std::exp(-y * xc) );
return state;
}
