From Entity Reliability to Clean Feedback: An Entity-Aware Denoising Framework Beyond Interaction-Level Signals

TL;DR

The paper tackles noise in implicit-feedback for recommender systems by shifting focus from interaction-level to entity-level reliability. It introduces EARD, a lightweight, model-agnostic framework that assigns weights to training instances using both interaction-level loss-based signals and entity-level reliability estimated from average losses, via an ECDF-based base weighting and a rank-based mapping to entity reputations. The final weight for each (user, item) pair is the product of base, user, and item weights, enabling effective denoising with minimal hyperparameter tuning ($\alpha,\beta$). Extensive experiments across Yelp, Amazon-Book, and ML-1M on GMF, NeuMF, and CDAE show significant improvements over state-of-the-art baselines, with favorable efficiency compared to heavy denoising methods. The work demonstrates the practical value of entity-aware reliability modeling for robust recommendations in noisy, large-scale settings.

Abstract

Implicit feedback is central to modern recommender systems but is inherently noisy, often impairing model training and degrading user experience. At scale, such noise can mislead learning processes, reducing both recommendation accuracy and platform value. Existing denoising strategies typically overlook the entity-specific nature of noise while introducing high computational costs and complex hyperparameter tuning. To address these challenges, we propose \textbf{EARD} (\textbf{E}ntity-\textbf{A}ware \textbf{R}eliability-\textbf{D}riven Denoising), a lightweight framework that shifts the focus from interaction-level signals to entity-level reliability. Motivated by the empirical observation that training loss correlates with noise, EARD quantifies user and item reliability via their average training losses as a proxy for reputation, and integrates these entity-level factors with interaction-level confidence. The framework is \textbf{model-agnostic}, \textbf{computationally efficient}, and requires \textbf{only two intuitive hyperparameters}. Extensive experiments across multiple datasets and backbone models demonstrate that EARD yields substantial improvements over state-of-the-art baselines (e.g., up to 27.01\% gain in NDCG@50), while incurring negligible additional computational cost. Comprehensive ablation studies and mechanism analyses further confirm EARD's robustness to hyperparameter choices and its practical scalability. These results highlight the importance of entity-aware reliability modeling for denoising implicit feedback and pave the way for more robust recommendation research.

Figures (10)

• Figure 1: (a) User (blue) and item (yellow) noise rate distributions on ML-1M. As in ADT ADT, ratings $\leq$ 3 are labeled as noise. The broad spread shows that entity reliability varies widely. (b) Average loss of user (purple) and item (brown) groups with different noise levels on Yelp using GMF. Average loss tends to increase with noise rate.
• Figure 2: Overview of EARD. At the end of each epoch, the framework performs three steps: (1) Collect Loss: gather losses for all interactions and compute average loss per entity; (2) Entity Reliability and Base Weight Calculation: estimate entity reliability via a linear mapping of average loss from $\alpha$ to $\beta$, and assign base weights using the ECDF of negative losses; (3) Weight Fusion: combine interaction and entity weights through element-wise multiplication to generate training weights for the next epoch.
• Figure 3: Training loss distributions at the end of the first epoch across different models and datasets.
• Figure 4: Training dynamics of true positive (TP) and false positive (FP) samples across models and datasets. Base Weight (BW) is interaction-specific, while User Factor (UF) and Item Factor (IF) capture user and item reliability, respectively. (a) TP and FP loss over training epochs. EARD preserves a clear loss gap, with FP samples generally showing higher losses, supporting the small-loss assumption. (b) Assigned weights for TP and FP samples. EARD steadily gives higher weights to TP samples, highlighting its ability to identify clean interactions and suppress noise.
• Figure 5: (a) Recall@50 performance landscape over $[\alpha, \beta]$ using NeuMF on ML-1M. (b) Hessian-based concavity test at 15 representative points from (a).