public override IRegressionModel Build(IRegressionProblemData pd, IRandom random, CancellationToken cancellationToken, out int numberOfParameters)
{
if (pd.Dataset.Rows < MinLeafSize(pd))
{
throw new ArgumentException("The number of training instances is too small to create a linear model");
}
numberOfParameters = pd.AllowedInputVariables.Count() + 1;
double x1, x2;
var coeffs = ElasticNetLinearRegression.CalculateModelCoefficients(pd, 1, 0.2, out x1, out x2);
numberOfParameters = coeffs.Length;
return(ElasticNetLinearRegression.CreateSymbolicSolution(coeffs, pd).Model);
}
private void CalculateCoefficients(IEnumerable <BasisFunction> basisFunctions, out double[] lambda, out double[,] coeff, out double[] trainNMSE, out double[] testNMSE, out double[] intercept, out IRegressionProblemData elnetData)
{
elnetData = PrepareData(Problem.ProblemData, basisFunctions);
if (Verbose)
{
Results.Add(new Result(
"Basis Functions",
"Dataset which contains the Basis Functions after FFX Step 1.",
elnetData
));
}
// "real" iteration with all Basis Functions in X_b
ElasticNetLinearRegression.RunElasticNetLinearRegression(elnetData, Penalty, out lambda, out trainNMSE, out testNMSE, out coeff, out intercept, double.NegativeInfinity, double.PositiveInfinity);
}
private IEnumerable <BasisFunction> CreateBasisFunctions()
{
IEnumerable <BasisFunction> basisFunctions = CreateUnivariateBases(Problem.ProblemData);
if (Verbose)
{
Results.Add(new Result(
"Problem.ProblemData",
"The problem data the algorithm gets as input.",
Problem.ProblemData
));
}
// wraps the list of basis functions in a dataset, so that it can be passed on to the ElNet function
var elnetData = PrepareData(Problem.ProblemData, basisFunctions);
if (Verbose)
{
Results.Add(new Result(
"Univariate basis Functions",
"A Dataset consisting of the univariate basis functions.",
elnetData
));
}
if (ConsiderInteractions)
{
// for the purpose of efficiency, only the "most important" sqrt(n) basis functions are to be selected for the merge of multivariate bases step (see FFX paper)
ElasticNetLinearRegression.RunElasticNetLinearRegression(elnetData, Penalty, out var lambda, out var trainNMSE, out var testNMSE, out var coeff, out var intercept);
basisFunctions = FilterCoeffs(basisFunctions, coeff);
basisFunctions = CreateMultivariateBases(basisFunctions);
return(basisFunctions);
}
// add denominator bases to the already existing basis functions
if (ConsiderDenominations)
{
basisFunctions = CreateDenominatorBases(Problem.ProblemData, basisFunctions);
}
return(basisFunctions);
}
