Dual-LightGCN+: Enhanced Dual Light Graph Convolutional Network for Discriminative Recommendation

The Dual-LightGCN has shown promise in recommendation systems but faces limitations in explicitly modeling similarity signals between users and items, which restricts its ability to uncover potential relationships and negatively impacts recommendation performance. To address this, we propose an enhanced approach, Dual-LightGCN+ (DLGN+), which introduces edges between two-hop neighbors in the original user-item interaction graph to construct user-user and item-item graphs. This modification allows for more effective capture of similarity signals. Furthermore, we employ a sampling strategy that retains the maximum spanning tree of these newly constructed graphs, balancing the trade-off between information preservation and computational cost. This strategy substantially reduces both space and time complexity by selectively maintaining the most relevant similarity information. We evaluated DLGN+ on two datasets of varying sizes and compared it with several state-of-the-art recommendation algorithms. Experimental results demonstrate that DLGN+ outperforms the widely used methods in recommendation accuracy.