Collaborative Filtering (CF) with implicit feedback is widely used in personalized recommender systems. In most real-world scenarios, only implicit feedback signals are available, making CF training heavily dependent on sampling-based paradigms—typically optimized via pairwise ranking losses such as Bayesian Personalized Ranking (BPR). This simple yet effective approach has achieved remarkable success and remains the foundation of many modern recommender models. However, despite its empirical success, little attention has been paid to the inherent training instability issue under this sampling-based paradigm. In this paper, we conduct an in-depth analysis of training stability and find that unstable training not only hinders convergence but also leads to fluctuating and suboptimal recommendation performance. We identify two fundamental sources of this instability in CF: (1) noisy or sparse positive samples, where a single observed interaction may not reliably reflect user preference; and (2) inconsistent negative samples, where randomly drawn negatives from unobserved space vary drastically in negative hardness, leading to uninformative or noisy gradient updates. To address these two challenges, we propose StabCF, a Stabilized Training Method for Collaborative Filtering, which improves training stability by synthesizing enriched positive samples from historical positives and constructing consistent hard negatives through user-aware negatives mixing. By replacing raw training triplets with synthesized positive-negative pairs, StabCF effectively smooths the training dynamics and improves convergence stability. Extensive experiments on three public datasets demonstrate that StabCF not only significantly stabilizes the training process but also achieves superior recommendation performance. Our PyTorch implementation is available at https://github.com/Wu-Xi/StabCF.

Negative sampling plays a pivotal role in training recommendation systems with implicit feedback, where the effectiveness of negatives directly impacts model convergence and recommendation quality. The key challenge is to efficiently mine high-quality hard negatives from the massive item space. Existing strategies typically rely on a single user perspective as the sampling anchor and use discrete arg max operations to select negatives. However, the single-anchor design introduces noisy negatives, and the discrete hard selection prevents end-to-end optimization. To address these limitations, we propose a differentiable dual-anchor negative sampling framework for graph-based recommendation. Our framework introduces a differentiable cross-hop sampling mechanism based on the Gumbel-Softmax trick, enabling hard negative selection while preserving gradient flow. Furthermore, we incorporate both the user and the corresponding positive item as complementary sampling anchors to improve the quality and stability of negative samples. Extensive experiments on three benchmark datasets demonstrate that our approach consistently improves recommendation performance.

Negative sampling plays a critical role in collaborative filtering (CF), as it accelerates convergence and improves recommendation accuracy. Among recent studies, mixup-based negative sampling has shown promising performance. However, existing methods primarily focus on increasing the similarity between the synthesized negative and the positive item, without considering the false positive issue commonly found in implicit feedback scenarios. Blindly training all positive samples with overly hard negatives can magnify the impact of false positives and hurt recommendation performance. To address this challenge, we first provide a theoretical analysis revealing that mixup-synthesized hard negatives implicitly reweight the similarity difference between the user's interactions and both the positive and negative boundaries, thereby shaping the training signal. Motivated by this, we propose a novel strategy named Dual Context-Aware Negative Sampling (DCANS), which enhances each positive item by assessing its alignment with the user's interest context, and simultaneously adjusts the hardness of synthesized negatives based on their relevance to the same interest context. This strategy optimizes the training direction toward the user's genuine preferences, mitigating the negative impact of false positives while preserving the benefits of hard negative sampling. Extensive experiments on three benchmark datasets demonstrate that our method achieves consistent improvements over state-of-the-art baselines. Our PyTorch implementation is available https://github.com/Wu-Xi/DCANS.

Collaborative filtering (CF) is a widely employed technique that predicts user preferences based on past interactions. Negative sampling plays a vital role in training CF-based models with implicit feedback. In this paper, we propose a novel perspective based on the sampling area to revisit existing sampling methods. We point out that current sampling methods mainly focus on Point-wise or Line-wise sampling, lacking flexibility and leaving a significant portion of the hard sampling area un-explored. To address this limitation, we propose Dimension Independent Mixup for Hard Negative Sampling (DINS), which is the first Area-wise sampling method for training CF-based models. DINS comprises three modules: Hard Boundary Definition, Dimension Independent Mixup, and Multi-hop Pooling. Experiments with real-world datasets on both matrix factorization and graph-based models demonstrate that DINS outperforms other negative sampling methods, establishing its effectiveness and superiority. Our work contributes a new perspective, introduces Area-wise sampling, and presents DINS as a novel approach that achieves state-of-the-art performance for negative sampling. Our implementations are available in PyTorch.