Algorithm 2 Framework of Returnformer

Input: User–item bipartite graph $G=(U,I,E)$; user features $X_u$; item features $X_i$; Node2Vec embeddings $X_{n2v}$

Output: Edge-level prediction return probabilities $\widehat{Y}$

1:  Initialize model parameters $\theta$ randomly

2:  Feature fusion: $X_f=\mathrm{Attention}\mathrm{Fusion}(X_u,X_i,X_{n2v})$

3:  Initialize node embeddings: $H^{\left(0\right)}=[X_{fu}\parallel X_{fi}]$

4:  for each edge $\in H^{\left(0\right)}$ do

5:     $V_u=X_{fu}\left[u_{\mathrm{idx}}\right],V_i=X_{fi}\left[i_{\mathrm{idx}}\right]$

6:     $H_u^{l+1},H_i^{l+1}=\mathrm{Graph}\mathrm{Transformer}(V_u,V_i)$

7:     $H_u^{\mathrm{out}},H_i^{\mathrm{out}}=\mathrm{GEA}(H_u^{l+1},H_i^{l+1})$

8:     $X_{\mathrm{out}}=[H_u^{\mathrm{out}}\parallel H_i^{\mathrm{out}}]$

9:     $\widehat{Y}=\mathrm{KAN}\left(X_{\mathrm{out}}\right)$

10:  end for

11:  return $\widehat{Y}$