/**
* * Least-square solution y = X * b where: y_i = b_0 + b_1*x_1i + b_2*x_2i + ... + b_k*x_ki
* including intercep term y_i = b_1*x_1i + b_2*x_2i + ... + b_k*x_ki without intercep term
*
* @param datay
* @param dataX
*/
private void multipleLinearRegression(Matrix datay, Matrix dataX) {
Matrix X, y;
try {
X = dataX;
y = datay;
b = X.solve(y);
coeffs = new double[b.getRowDimension()];
for (int j = 0; j < b.getRowDimension(); j++) {
coeffs[j] = b.get(j, 0);
// System.out.println("coeff[" + j + "]=" + coeffs[j]);
}
// Residuals:
Matrix r = X.times(b).minus(y);
residuals = r.getColumnPackedCopy();
// root mean square error (RMSE)
rmse = Math.sqrt(MathUtils.sumSquared(residuals) / residuals.length);
// Predicted values
Matrix p = X.times(b);
predictedValues = p.getColumnPackedCopy();
// Correlation between original values and predicted ones
correlation = MathUtils.correlation(predictedValues, y.getColumnPackedCopy());
} catch (RuntimeException re) {
throw new Error("Error solving Least-square solution: y = X * b");
}
}
// Given a set of coefficients and data predict values applying linear equation
// This function can be used to test with data that was not used in training
// c[] is the number of the columns in the file not the indexFeatures
public void predictValues(
String fileName, int indVariable, int[] c, boolean interceptTerm, int rowIni, int rowEnd) {
try {
BufferedReader reader = new BufferedReader(new FileReader(fileName));
Matrix data = Matrix.read(reader);
reader.close();
int rows = data.getRowDimension() - 1;
int cols = data.getColumnDimension() - 1;
if (rowIni < 0 || rowIni > rows)
throw new RuntimeException(
"Problem reading file, rowIni=" + rowIni + " and number of rows in file=" + rows);
if (rowEnd < 0 || rowEnd > rows)
throw new RuntimeException(
"Problem reading file, rowIni=" + rowIni + " and number of rows in file=" + rows);
if (rowIni > rowEnd)
throw new RuntimeException(
"Problem reading file, rowIni < rowend" + rowIni + " < " + rowEnd);
Matrix indVar =
data.getMatrix(
rowIni,
rowEnd,
indVariable,
indVariable); // dataVowels(:,0) -> last col is the independent variable
data =
data.getMatrix(
rowIni, rowEnd, c); // the dependent variables correspond to the column indices in c
int numCoeff;
if (interceptTerm) numCoeff = c.length + 1;
else numCoeff = c.length;
if (b != null) {
if (b.getRowDimension() == numCoeff) {
if (interceptTerm) { // first column of X is filled with 1s if b_0 != 0
int row = data.getRowDimension();
int col = data.getColumnDimension();
Matrix B = new Matrix(row, col + 1);
Matrix ones = new Matrix(row, 1);
for (int i = 0; i < row; i++) ones.set(i, 0, 1.0);
B.setMatrix(0, row - 1, 0, 0, ones);
B.setMatrix(0, row - 1, 1, col, data);
data = B;
}
// Residuals:
Matrix r = data.times(b).minus(indVar);
residuals = r.getColumnPackedCopy();
// root mean square error (RMSE)
rmse = Math.sqrt(MathUtils.sumSquared(residuals) / residuals.length);
// Predicted values
Matrix p = data.times(b);
predictedValues = p.getColumnPackedCopy();
for (int i = 0; i < predictedValues.length; i++)
if (predictedValues[i] < 0.0)
System.out.println(
"*** WARNING predictedValue < 0.0 : predictedValues["
+ i
+ "]="
+ predictedValues[i]);
// Correlation between original values and predicted ones
correlation = MathUtils.correlation(predictedValues, indVar.getColumnPackedCopy());
System.out.println("Correlation predicted values and real: " + correlation);
System.out.println("RMSE (root mean square error): " + rmse);
} else {
throw new RuntimeException(
"Number of columns of data is not the same as number of coeficients");
}
} else {
throw new RuntimeException("Regression coefficients are not loaded");
}
} catch (Exception e) {
throw new RuntimeException("Problem reading file " + fileName, e);
}
}
