Ridge.py

#!/usr/bin/env python
# coding: utf-8

# In[1]:


import import_ipynb


# In[2]:


import numpy as np


class Kernelridge():

    def __init__(self, K, ID, eps=1e-5, lbda=0.1, solver=None):

        self.K = K
        self.ID = ID
        self.eps = eps
        self.lbda = lbda
        self.solver = solver

    def fit(self, X, y):

        self.Id_fit = np.array(X.loc[:, 'Id'])
        self.idx_fit = np.array([np.where(self.ID == self.Id_fit[i])[0] for i in range(len(self.Id_fit))]).squeeze()
        self.K_fit = self.K[self.idx_fit][:, self.idx_fit]
        self.y_fit, self.X_fit, = np.array(y.loc[:, 'Bound']), X
        self.n = self.K_fit.shape[0]
        self.a = np.dot(np.linalg.inv(self.K_fit + self.lbda * self.n * np.eye(self.n)), self.y_fit)
        # Align support vectors index with index from fit set
        self.idx_sv = np.where(np.abs(self.a) > self.eps)
        self.y_fit = self.y_fit[self.idx_sv]
        self.a = self.a[self.idx_sv]
        self.idx_sv = self.idx_fit[self.idx_sv]
        # Intercept
        self.y_hat = np.array([np.dot(self.a, self.K[self.idx_sv, i]).squeeze() for i in self.idx_sv])
        self.b = np.mean(self.y_fit - self.y_hat)

    def predict(self, X):

        # Align prediction IDs with index in kernel K
        self.Id_pred = np.array(X.loc[:, 'Id'])
        self.idx_pred = np.array([np.where(self.ID == self.Id_pred[i])[0] for i in range(len(self.Id_pred))]).squeeze()
        pred = []
        for i in self.idx_pred:
            pred.append(np.sign(np.dot(self.a, self.K[self.idx_sv, i].squeeze()) + self.b))
        return np.array(pred)

    def score(self, pred, y):
        """
        Compute accuracy of predictions according to y
        :param pred: np.array, predictions (-1/1)
        :param y: np.array or pd.DataFrame, true labels
        :return: float, percentage of correct predictions
        """
        label = np.array(y.loc[:, 'Bound']) if not isinstance(y, np.ndarray) else y
        assert 0 not in np.unique(label), 
        return np.mean(pred == label)


# In[ ]: