public static void Test_CollaborativeFilteringRSCostFunction_Evaluate(double lambda)
{
int num_users = 4; int num_movies = 5; int num_features = 3;
List <List <double> > X = DblDataTableUtil.LoadDataSet("X.txt");
List <List <double> > Y = DblDataTableUtil.LoadDataSet("Y.txt");
List <List <int> > R = IntDataTableUtil.LoadDataSet("R.txt");
List <List <double> > Theta = DblDataTableUtil.LoadDataSet("Theta.txt");
X = DblDataTableUtil.SubMatrix(X, num_movies, num_features);
Y = DblDataTableUtil.SubMatrix(Y, num_movies, num_users);
R = IntDataTableUtil.SubMatrix(R, num_movies, num_users);
Theta = DblDataTableUtil.SubMatrix(Theta, num_users, num_features);
Matrix <double> Y_matrix = Convert2Matrix(Y);
Matrix <double> X_matrix = Convert2Matrix(X);
Matrix <double> Theta_matrix = Convert2Matrix(Theta);
int[,] R_matrix = IntDataTableUtil.Convert2DArray(R);
int dimension = num_movies * num_features + num_users * num_features; //total number of entries in X and Theta
double[] theta_x = new double[dimension];
CollaborativeFilteringRSCostFunction.UnrollMatrixIntoVector(Theta_matrix, X_matrix, theta_x);
CollaborativeFilteringRSCostFunction f = new CollaborativeFilteringRSCostFunction(Y_matrix, R_matrix, num_movies, num_features, dimension);
f.RegularizationLambda = lambda;
double J = f.Evaluate(theta_x);
Console.WriteLine("Cost at loaded parameters: {0} (this value should be about 22.22)", J);
}
public static void Test_CollaborativeFilteringRSCostFunction_CalcGradient(double lambda = 0)
{
Matrix <double> X_t = CreateRandomMatrix(4, 3);
Matrix <double> Theta_t = CreateRandomMatrix(5, 3);
Matrix <double> Y = X_t.Multiply(Theta_t.Transpose());
int[,] R = new int[Y.RowCount, Y.ColumnCount];
for (int i = 0; i < Y.RowCount; ++i)
{
for (int j = 0; j < Y.ColumnCount; ++j)
{
if (mRand.NextDouble() > 0.5)
{
Y[i, j] = 0;
}
if (Y[i, j] == 0)
{
R[i, j] = 0;
}
else
{
R[i, j] = 1;
}
}
}
Matrix <double> X = CreateRandomMatrix(4, 3);
Matrix <double> Theta = CreateRandomMatrix(5, 3);
int num_users = Y.ColumnCount;
int num_movies = Y.RowCount;
int num_features = Theta_t.ColumnCount;
int dimension = num_movies * num_features + num_users * num_features; //total number of entries in X and Theta
double[] theta_x = new double[dimension];
CollaborativeFilteringRSCostFunction.UnrollMatrixIntoVector(Theta, X, theta_x);
CollaborativeFilteringRSCostFunction f = new CollaborativeFilteringRSCostFunction(Y, R, num_movies, num_features, dimension);
f.RegularizationLambda = lambda;
double[] numgrad = new double[dimension];
double[] grad = new double[dimension];
GradientEstimation.CalcGradient(theta_x, numgrad, (x_pi, constraints) =>
{
return(f.Evaluate(x_pi));
});
f.CalcGradient(theta_x, grad);
Console.WriteLine("The relative difference will be small:");
for (int i = 0; i < dimension; ++i)
{
Console.WriteLine("{0}\t{1}", numgrad[i], grad[i]);
}
}
public double ComputeCost(List <T> data_set, int num_features)
{
int n_m = data_set.Count;
if (n_m == 0)
{
throw new ArgumentException("Data set is empty!");
}
int x_dimension = num_features;
int n_u = data_set[0].UserRanks.Length;
Matrix <double> Y_matrix = new SparseMatrix(n_m, n_u, 0);
int[,] r_matrix = new int[n_m, n_u];
int theta_dimension = n_u * x_dimension;
int content_dimension = n_m * x_dimension;
int dimension = theta_dimension + content_dimension;
for (int i = 0; i < n_m; ++i)
{
T rating = data_set[i];
double[] Y = rating.UserRanks;
bool[] r = rating.IsRated;
for (int j = 0; j < n_u; ++j)
{
Y_matrix[i, j] = Y[j];
r_matrix[i, j] = r[j] ? 1 : 0;
}
}
CollaborativeFilteringRSCostFunction f = new CollaborativeFilteringRSCostFunction(Y_matrix, r_matrix, n_m, x_dimension, dimension);
f.RegularizationLambda = 0;
return(f.Evaluate(mThetaX));
}
public Matrix <double> Compute(List <T> data_set, int num_features)
{
int n_m = data_set.Count;
if (n_m == 0)
{
throw new ArgumentException("Data set is empty!");
}
int n_u = data_set[0].UserRanks.Length;
Matrix <double> Y_matrix = new SparseMatrix(n_m, n_u, 0);
int[,] r_matrix = new int[n_m, n_u];
int theta_dimension = n_u * num_features;
int content_dimension = n_m * num_features;
int dimension = theta_dimension + content_dimension;
for (int i = 0; i < n_m; ++i)
{
T rating = data_set[i];
rating.ItemIndex = i;
double[] Y = rating.UserRanks;
bool[] r = rating.IsRated;
for (int j = 0; j < n_u; ++j)
{
Y_matrix[i, j] = Y[j];
r_matrix[i, j] = r[j] ? 1 : 0;
}
}
CollaborativeFilteringRSCostFunction f = new CollaborativeFilteringRSCostFunction(Y_matrix, r_matrix, n_m, num_features, dimension);
f.RegularizationLambda = mRegularizationLambda;
double[] theta_x_0 = new double[dimension];
for (int d = 0; d < dimension; ++d)
{
theta_x_0[d] = NextRandomDouble();
}
ContinuousSolution solution = mLocalSearcher.Minimize(theta_x_0, f, mMaxSolverIteration);
mThetaX = solution.Values;
mTheta = new DenseMatrix(n_u, num_features);
for (int j = 0; j < n_u; ++j)
{
for (int k = 0; k < num_features; ++k)
{
int index = j * num_features + k;
mTheta[j, k] = mThetaX[index];
}
}
Matrix <double> X = new DenseMatrix(n_m, num_features);
for (int i = 0; i < n_m; ++i)
{
T rec = data_set[i];
if (rec.X == null)
{
rec.X = new double[num_features];
}
for (int k = 0; k < num_features; ++k)
{
int index = i * num_features + k + theta_dimension;
rec.X[k] = mThetaX[index];
X[i, k] = mThetaX[index];
}
}
Matrix <double> XThetaPrime = X.Multiply(mTheta.Transpose());
for (int i = 0; i < n_m; ++i)
{
T rec = data_set[i];
for (int j = 0; j < n_u; ++j)
{
if (r_matrix[i, j] == 0)
{
rec.UserRanks[j] = XThetaPrime[i, j];
}
}
}
return(XThetaPrime);
}
