MixDec Sampling: A Soft Link-based Sampling Method of Graph Neural Network for Recommendation

TL;DR

MixDec Sampling introduces soft link-based sampling for GNN-based recommender systems to overcome the limitations of binary hard sampling. It combines Mixup Sampling and BFS-driven Decay Sampling to synthesize new nodes and preserve graph structure, improving embedding learning for nodes with few neighbors. Across three datasets and three backbone GNNs, MixDec yields consistent gains in MRR and Hit@K, with notable benefits in dense graphs and sparse neighborhoods. The approach is plug-and-play, scalable, and offers a practical boost for real-world graph-based recommendations.

Abstract

Graph neural networks have been widely used in recent recommender systems, where negative sampling plays an important role. Existing negative sampling methods restrict the relationship between nodes as either hard positive pairs or hard negative pairs. This leads to the loss of structural information, and lacks the mechanism to generate positive pairs for nodes with few neighbors. To overcome limitations, we propose a novel soft link-based sampling method, namely MixDec Sampling, which consists of Mixup Sampling module and Decay Sampling module. The Mixup Sampling augments node features by synthesizing new nodes and soft links, which provides sufficient number of samples for nodes with few neighbors. The Decay Sampling strengthens the digestion of graph structure information by generating soft links for node embedding learning. To the best of our knowledge, we are the first to model sampling relationships between nodes by soft links in GNN-based recommender systems. Extensive experiments demonstrate that the proposed MixDec Sampling can significantly and consistently improve the recommendation performance of several representative GNN-based models on various recommendation benchmarks.

Figures (7)

• Figure 1: An overview of the MixDec sampling. MixDec consists of Mixup Sampling and Decay Sampling. 1) Based on sampled positive and negative items for the anchored user, Mixup Sampling linearly mixes $<$positive item, positive item$>$ pairs and $<$positive item, negative item$>$ pairs respectively to generate synthetic items. Then, the soft link is created between synthetic items and the user. 2) In Decay sampling, the decay item is the neighbor within the l-hop for the anchored user. A soft link between $<$user, decay item$>$ will be built. In MixDec, Mixup Samples and Decay Samples will be integrated.
• Figure 2: An illustration of the Mixup sampling. User $U$ is the anchored node. Due to directly connected to user $U$, $P_1$ and $P_2$ are positive samples. N is a negative sample. $M_1$ and $M_2$ are synthetic items, generated by mixing the feature of $<P_1, P_2>$ pair and $<P_1, N>$ pair according to the parameter $\lambda_1$ and $\lambda_2$, respectively. Then, we create a soft link with weight 1 between $<U, M_1>$ pair, and another soft link with weight $1-\lambda_2$ between $<U, M_2>$ pair.
• Figure 3: An illustration of the Decay sampling. $D_1$ and $D_2$ are decay items, which are the second-hop items found by anchored User $U$ through BFS. Then a soft link is built between $<D_1,U>$ and $<D_2,U>$. Using soft link, the structural information of the graph is effectively preserved.
• Figure 4: An example of BFS-based decay weights of soft links. For the anchored user, items with direct interactions are regarded as positive items, and decay items are 3-hop neighbors. For these four decay items, the number of pathways from the user to the decay items, as obtained by BFS, is 1, 1, 3 and 2. Correspondingly, the decay weights of soft links are 0.66, 0.66, 1.00 and 0.83, when the hyperparameter $\rho$ is set to 0.5.
• Figure 5: An illustration of the MixDec sampling. For example, in the figure, $P_1$, $P_2$, $N$, $M_1$ and $M_2$ are the same as the Mixup sampling phase \ref{['fig:mix_sam']}, and $D$ is the node searched by BFS in the Decay sampling phase. We synthesize $M_3$ by mixing $<P_1, D>$ pair, and also build a soft link between $<U, M_3>$. MixDec sampling integrates Mixup sampling and Decay sampling, and the structure and feature information of the graph will be reserved.