Algorithm 2: FIS using threshold strategy.

Input: Number of feature fields: m; Threshold: $\epsilon$. Output: Hard vector of gated unit: $\mathbb{H}$. $\mathbb{W}\leftarrow$ Randomly initialize weights of features; $\mathbb{S}\leftarrow$ Compute Gumbel-softmax of $\mathbb{W}$; for $i=1$ to mdo