Algorithm 1.

Training of node splitting.

1:  Randomly generates a set of feature subspace candidates {wj}

2:  For each wj, find the optimal ${\tau }_j^{*}=\underset{\tau }{\text{argmax}}G(w_j,\tau ,X,Y)$ that best splits the data.

3:  Among all {wj, ${\tau }_j^{*}$}, pick the one with largest information gain: $j^{*}=\underset{j}{\text{argmax}}G(w_j,{\tau }_j^{*},X,Y)$