fri.py 文件源码

def community_detection(self,cutoff_threshold=0.55):
        '''
        Finding communities of features using pairwise differences of solutions aquired in the main LP step.
        '''
        svm_solution = self._svm_coef
        abs_svm_sol = np.abs(svm_solution)

        om = self._omegas
        mins = om[:,0,:]
        maxs = om[:,1,:]
        abs_mins = np.abs(mins)
        abs_maxs = np.abs(maxs)

        # Aggregate min and max solution values to obtain the absolute variation
        lower_variation = abs_svm_sol - abs_mins
        upper_variation = abs_maxs - abs_svm_sol
        variation = np.abs(lower_variation) + np.abs(upper_variation)

        # add up lower triangular matrix to upper one
        collapsed_variation = np.triu(variation)+np.tril(variation).T
        np.fill_diagonal(collapsed_variation, 0)
        #collapsed_variation = pd.DataFrame(collapsed_variation)

        # Create distance matrix
        dist_mat = np.triu(collapsed_variation).T + collapsed_variation
        # normalize
        dist_mat = 1- dist_mat/np.max(dist_mat)
        # get numpy array
        #dist_mat = dist_mat.values[:]
        # feature with itself has no distance
        np.fill_diagonal(dist_mat,0)

        # convert to squareform for scipy compat.
        dist_mat_square = squareform(dist_mat)

        # Execute clustering 
        link = linkage(dist_mat_square, method="ward")

        # Set cutoff at which threshold the linkage gets flattened (clustering)
        RATIO = cutoff_threshold
        threshold = RATIO * np.max(link[:,2]) # max of branch lengths (distances)

        feature_clustering = fcluster(link,threshold,criterion="distance")

        return feature_clustering, link