Towards Reliable Negative Sampling for Recommendation with Implicit Feedback via In-Community Popularity

Chen Chen; Haobo Lin; Yuanbo Xu

cs.IRFeb 21, 2026

Towards Reliable Negative Sampling for Recommendation with Implicit Feedback via In-Community Popularity

Authors: Chen Chen, Haobo Lin, Yuanbo Xu

Organizations: College of Computer Science and Technology, Jilin University, Changchun, Jilin, China

Abstract

Learning from implicit feedback is a fundamental problem in modern recommender systems, where only positive interactions are observed and explicit negative signals are unavailable. In such settings, negative sampling plays a critical role in model training by constructing negative items that enable effective preference learning and ranking optimization. However, designing reliable negative sampling strategies remains challenging, as they must simultaneously ensure realness, hardness, and interpretability. To this end, we propose ICPNS (In-Community Popularity Negative Sampling), a novel framework that leverages user community structure to identify reliable and informative negative samples. Our approach is grounded in the insight that item exposure is driven by latent user communities. By identifying these communities and utilizing in-community popularity, ICPNS effectively approximates the probability of item exposure. Consequently, items that are popular within a user's community but remain unclicked are identified as more reliable true negatives. Extensive experiments on four benchmark datasets demonstrate that ICPNS yields consistent improvements on graph-based recommenders and competitive performance on MF-based models, outperforming representative negative sampling strategies under a unified evaluation protocol.

Figures & tables

Figure 1. Overview of negative sampling ambiguity and the motivation of ICPNS. (A) Two types of unobserved items under implicit feedback. Triangles with darker colors indicate higher sample hardness; as hardness increases, the likelihood of false negatives also increases. (B) The inherent ambiguity of negative samples: true negatives arising from genuine disinterest and false negatives caused by lack of exposure. ICPNS approximates item exposure using in-community popularity, treating unclicked but community-popular items as more reliable negatives.

Strategy	Realness	Hardness	Interpretability
RNS	✗	✗	✗
PNS	✗	✗	✓
HNS	✗	✓	✗
ICPNS	✓	✓	✓

Table 1. Comparison of representative negative sampling strategies with respect to realness, hardness, and interpretability.

Figure 2. The detailed framework of ICPNS. After encoder is well-pretrained by RNS, the model switches to our ICPNS strategy. Here, users are clustered into latent communities based on pre-trained embeddings, and in-community popularity is leveraged to sample reliable negative items.

Dataset	Users	Items	Interactions	Sparsity
ML-100K	940	1017	80,393	91.59%
ML-1M	6,034	3,124	834,449	95.57%
Yelp	26,420	19,921	1,285,021	99.76%
Beauty	1,238	720	25,736	97.11%

Table 2. Statistics of four preprocessed datasets.

Model	Strategy	ML-100K				ML-1M				Yelp				Beauty
Model	Strategy	Rec	MRR	NDCG	Pre	Rec	MRR	NDCG	Pre	Rec	MRR	NDCG	Pre	Rec	MRR	NDCG	Pre
BPR	RNS	0.266	0.418	0.264	0.170	0.172	0.416	0.238	0.179	0.070	0.093	0.056	0.030	0.231	0.145	0.146	0.043
	PNS	0.253	0.408	0.254	0.162	0.174	0.416	0.237	0.177	0.070	0.094	0.056	0.029	0.231	0.145	0.144	0.043
	HNS	0.258	0.398	0.250	0.159	0.176	0.418	0.239	0.178	0.085	0.115	0.069	0.036	0.234	0.159	0.155	0.043
	ICPNS	0.269	0.413	0.257	0.162	0.176	0.415	0.238	0.178	0.082	0.113	0.068	0.035	0.224	0.147	0.147	0.041
NeuMF	RNS	0.253	0.415	0.259	0.162	0.161	0.380	0.217	0.166	0.063	0.079	0.049	0.025	0.232	0.150	0.149	0.043

Table 3. Performance comparison of various negative sampling strategies across different backbone models. Boldface indicate the best results for each model.

Panel of Figure 3 — Figure 3. Comparison of training time per epoch across datasets. The y-axis is plotted on a logarithmic scale. To improve readability, small and large datasets are reported separately. ICPNS consistently achieves lower computational overhead compared to HNS.

Panel of Figure 4 — Figure 4. Comparison of hardness across different negative sampling strategies, averaged over all training epochs across four datasets.

Dataset	Variant	Rec	MRR	NDCG	Pre
ML-100K	Default	0.311	0.514	0.332	0.204
	w/o pretraining	0.157	0.298	0.161	0.102
	w/o clustering	0.270	0.445	0.275	0.172
	w/o smoothing	0.277	0.455	0.279	0.174
ML-1M	Default	0.212	0.504	0.304	0.224
	w/o pretraining	0.166	0.422	0.237	0.176

Table 4. Ablation study evaluating the impact of key components, including pretraining, user clustering, and popularity smoothing, on recommendation performance across datasets. Default represents standard ICPNS.

Dataset	Clustering	Rec	MRR	NDCG	Pre
ML-100K	$k$ -means	0.311	0.514	0.332	0.204
	GMM	0.335	0.560	0.374	0.240
	DBSCAN	0.330	0.555	0.371	0.239
	AC	0.347	0.603	0.404	0.251
	SC	0.350	0.596	0.401	0.247
Beauty	$k$ -means	0.260	0.184	0.179	0.050

Table 5. Performance comparison of ICPNS under different clustering algorithms, reflecting the impact of clustering quality.

Panel of Figure 5 — Figure 5. Parameter sensitivity analysis of ICPNS on Beauty and ML-100K. The top row (a–b) illustrates the impact of $P$ , while the bottom row (c–d) displays the effect of $\alpha$ .

Panel of Figure 6 — Figure 6. Visualization of community structure before and after fine-tuning with ICPNS at $P=8$ .

Silhouette			Calinski–Harabasz
Before	After	$\boldsymbol{\Delta}$	Before	After	$\boldsymbol{\Delta}$
0.132	0.091	-0.041	83.231	65.624	-18.393

Table 6. Clustering quality metrics before and after fine-tuning on ML-100K. Lower Silhouette and Calinski–Harabasz scores indicate a more dispersed embedding structure.

\Delta

denotes After minus Before .

Community 3
Movie Name	Genres
The Net	Sci-Fi, Thriller
Hoodlum	Crime, Film-Noir
Hellraiser	Horror, Sci-Fi
Aladdin and the King of Thieves	Children’s
Bound	Crime, Thriller

Table 7. Community 3 and its top-5 popular movies.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7. on ML-100K

Panel of Figure 8 — Figure 8. Parameter sensitivity analysis of ICPNS on ML-1M and Yelp. The top row (a–b) illustrates the impact of $P$ , while the bottom row (c–d) displays the effect of $\alpha$ .

Figure 9. NDCG@10 trajectories during fine-tuning on ML-100K. Both methods start from the same pretrained checkpoint at Epoch 267. ICPNS consistently improves performance over RNS under the same warm-up protocol.

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 514 papers

Evidence Before Sampling: Interpretable Implicit Negative Candidate Discovery for Recommendation
Oct 6, 2026Shreya Rajpal, Sonia Sharma, Swapnil Parekh +4Recommender Systems Explainable Recommendation
Michigan State University, USA · Intuit, USA
TFPS: A Temporal Filtration-enhanced Positive Sample Set Construction Method for Implicit Collaborative Filtering
Feb 26, 2026Jiayi Wu, Zhengyu Wu, Xunkai Li +2Collaborative Filtering User Preference Modeling
Beijing Institute of Technology, Beijing, China
Denoising Implicit Feedback for Cold-start Recommendation
Jun 17, 2026Gaode Chen, Shicheng Wang, Shikun Li +8Cold-Start Recommendation Noisy-Label Learning
Kuaishou Technology Beijing, China · Hong Kong Baptist University Hong Kong, China · Institute of Information Engineering, Chinese Academy of Sciences Beijing, China

Oct 6, 2026J/K move · Enter open · S save

Evidence Before Sampling: Interpretable Implicit Negative Candidate Discovery for Recommendation

Shreya Rajpal, Sonia Sharma, Swapnil Parekh, Lisa Li+3

Michigan State University, USA · Intuit, USA

Recommender Systems Explainable Recommendation

Open paper AskPDF arXiv

Recommender systems learn from observed user-item interactions, but explicit negative feedback is often unavailable. Since deep learning models require negative signals for training, negative sampling methods typically treat selected unobserved interactions as negatives. However, a missing interaction does not explain why a user is uninterested in an item or whether there is sufficient evidence to label it negative. This is especially important in business recommendation, where negative signals should be interpretable and aligned with business objectives. We formulate implicit negative candidate discovery to identify unobserved interactions supported by observed customer behavior. We encode these patterns as symbolic rules, score them based on support, informativeness, and product relevance, and rank the retained rules by evidence. An LLM then interprets the retained rules using business objectives and domain knowledge; the interpretations are combined with the statistical evidence in the final report. We evaluate our method in an industrial B2B setting and across five public recommendation datasets. Candidate-quality evaluations in the industrial setting and three public datasets show higher precision than the evaluated baselines, while symbolic selection improves downstream test PR-AUC by 12.5% over random selection with four negatives per positive example in the industrial task. Our results show that negative candidate validity can be evaluated separately from downstream recommendation performance. This distinction enables evidence-based, business-aligned, and explainable negative selection, improving both interpretability and model training in sparse, skewed, real-world recommendation settings.

Abstract

Recommender Systems

Explore similar work

Evidence Before Sampling: Interpretable Implicit Negative Candidate Discovery for Recommendation

TFPS: A Temporal Filtration-enhanced Positive Sample Set Construction Method for Implicit Collaborative Filtering

Denoising Implicit Feedback for Cold-start Recommendation