Home > Software design >  Too accurate to be true! (linear regression)
Too accurate to be true! (linear regression)

Time:08-28

I have been trying to code a gradient descent algorithm from scratch for multi-featured linear regression but when I'm predicting using my own training dataset I'm getting too accurate results.

class gradientdescent:

  
     def fit(self,X,Y):
        lr=0.005    *learning rate*
        b=0
        M=np.array(1)
        M=np.arange(X.shape[1])
        n=np.size(X,0)
        M.fill(1)     #initial value for gradient
        for i in range(10000):
             sum=0
             sum1=0
             for j in range(n):
                 sum=sum (np.dot(X[j],M) b-Y[j])*X[j]
                 sum1=sum1 (np.dot(X[j],M) b-Y[j])
             m_gradient=lr*sum/n
             b_gradient=lr*sum1/n
             M=M-m_gradient
             b=b-b_gradient
        self.b=b
        self.M=M
        self.n=n

This dataset I have taken below is too random, I had randomly entered values here in the X and Y array.

  X=np.array([[1,2,3,4,5],[2,1,4,3,5],[1,3,2,5,4],[3,0,1,2,4],[0,1,2,4,3]])
   Y=np.array([5,6,2,8,100])

my prediction function:

def predict(self,X):
           for i in range(self.n):
              print(np.dot(self.M,X[i]) self.b)

The predicted values:

5.000000000080892
5.999999999956618
1.9999999999655422
8.000000000004814
99.99999999998795

There is no way the plotted graph passes through the training dataset such closely as the data given was random so I had expected there to be little error. I even tried changing the data but still, it gives me these accurate results.

please tell me if there is any problem with my algorithm.

CodePudding user response:

Your X matrix is fully invertible thus the linear solution is just

w = X^-1 y

or in numpy

w = np.linalg.inv(X).dot(Y)  
# array([-13.98181818, -59.8       ,   2.74545455,  47.56363636,
         -11.98181818])

and then you get a perfect prediction

X.dot(w)
# array([  5.,   6.,   2.,   8., 100.])

This is happening because almost every "random" square matrix is invertible. The space of non-invertible matrices is actually pretty small, so you need to construct them somewhat carefully. Another situation where the error is to be expected is when you have less features than data points, and then linear regression will be much less likely to be perfect.

  • Related