Iterative Nearest Neighborhood Oversampling in Semisupervised Learning from Imbalanced Data

Algorithm 1

Iterative Nearest Neighborhood Oversampling (INNO).

Input: 𝑘NN graph, affinity matrix 𝑊, stop parameter 𝑠, imbalanced labeled dataset 𝑋_𝐿 and

unlabeled dataset 𝑋_𝑈;

Output: balanced or approximate balanced labeled dataset.

Procedure:

1 while var(𝑟)>𝑠

2 Initialization 𝑟_𝑗=min_{𝑗∈1…𝑐}𝑟, max=−∞, max𝑘=0;

3 for each labeled sample 𝑥_𝑗 in class 𝑗

4  for each neighbors 𝑥_𝑘 of 𝑥_𝑗

5   skip the 𝑥_𝑘 if it is in 𝑋_𝐿 or 𝑥_𝑘 has edges between labeled samples in other class;

6   if 𝑊_𝑖𝑗>max, then update max, max𝑘

7  end for

8 end for

9 if max𝑘=0 // all the neighbors of labeled samples in class 𝑗 have edges with labeled samples

in other classes, then 𝑟_𝑗=𝑟_max, continue;

10 label 𝑥_max𝑘 with class 𝑗, remove it from 𝑋_𝑈, add it to 𝑋_𝐿, 𝑟_𝑗=𝑟_𝑗+1;

11 end while