Source-linked AI summary
Denoising Implicit Feedback for Recommendation
Wenjie Wang, Fuli Feng, Xiangnan He, Liqiang Nie, Tat-Seng Chua
TL;DR
Implicit feedback is abundant but noisy, creating a gap between observed interactions and user satisfaction that can harm recommender training. The paper proposes Adaptive Denoising Training, which prunes or down-weights large-loss interactions without additional feedback, and reports improved recommendation quality across benchmarks. Its precision remains limited in identifying noisy interactions, and broader loss-function applications are left for future work.
Problem
Implicit feedback is abundant but often fails to reflect actual user satisfaction, while recommendation research has given limited attention to this noise.
Method
Adaptive Denoising Training dynamically discards or down-weights large-loss interactions during training, using Truncated Loss and Reweighted Loss formulations.
Results
ADT improves recommendation quality over normal training across three benchmarks and three representative recommenders.
Takeaways & Limitations
The two loss paradigms reduce the disturbance of noisy implicit feedback and can be applied across recommendation loss functions, neural recommenders, and optimizers.
Takeaways & Limitations
Truncated Loss has low precision, discarding many clean interactions; improving this precision and testing other loss functions remain future directions.
Abstract
from arXiv · showhide
The ubiquity of implicit feedback makes them the default choice to build online recommender systems. While the large volume of implicit feedback alleviates the data sparsity issue, the downside is that they are not as clean in reflecting the actual satisfaction of users. For example, in E-commerce, a large portion of clicks do not translate to purchases, and many purchases end up with negative reviews. As such, it is of critical importance to account for the inevitable noises in implicit feedback for recommender training. However, little work on recommendation has taken the noisy nature of implicit feedback into consideration. In this work, we explore the central theme of denoising implicit feedback for recommender training. We find serious negative impacts of noisy implicit feedback,i.e., fitting the noisy data prevents the recommender from learning the actual user preference. Our target is to identify and prune noisy interactions, so as to improve the quality of recommender training. By observing the process of normal recommender training, we find that noisy feedback typically has large loss values in the early stages. Inspired by this observation, we propose a new training strategy namedAdaptive Denoising Training(ADT), which adaptively prunes noisy interactions during training. Specifically, we devise two paradigms for adaptive loss formulation: Truncated Loss that discards the large-loss samples with a dynamic threshold in each iteration; and reweighted Loss that adaptively lowers the weight of large-loss samples. We instantiate the two paradigms on the widely used binary cross-entropy loss and test the proposed ADT strategies on three representative recommenders. Extensive experiments on three benchmarks demonstrate that ADT significantly improves the quality of recommendation over normal training.
1 INTRODUCTION
Implicit feedback is abundant but can contain false-positive interactions that obscure user satisfaction. The paper proposes Adaptive Denoising Training, which dynamically reduces the influence of large-loss interactions without additional data.
- Motivation: Implicit feedback is widely used because its volume alleviates sparsity, but clicks and purchases may not reflect actual user satisfaction.False-positive interactions include purchases followed by negative reviews or returns.
- Motivation: Prior approaches identify negative experiences or incorporate additional feedback, whereas this work denoises implicit interactions without extra data.The proposed approach distills signals from implicit interactions across users and items.
- Motivation: False-positive interactions typically have larger loss values during the early stages of recommender training.This observation motivates pruning or down-weighting hard interactions during optimization.
- Approach: Adaptive Denoising Training dynamically prunes large-loss interactions through Truncated Loss and Reweighted Loss paradigms.Truncated Loss discards large-loss interactions, while Reweighted Loss lowers their weights; both are instantiated on binary cross-entropy.
- Contributions: The paper formulates denoising for recommender training, applies the two loss paradigms to representative recommenders, and validates improved recommendation quality on three benchmarks.The contribution summary identifies three representative recommenders and extensive benchmark experiments.
2 STUDY ON FALSE-POSITIVE FEEDBACK
False-positive interactions substantially degrade recommender performance relative to clean training. Because reliable feedback is sparse in practice, the study motivates denoising implicit feedback rather than relying on clean labels alone.
- Empirical effects: Normal NeuMF training with false-positive interactions causes average performance drops of 15.69% in Recall@20 and 6.37% in NDCG@20 across two datasets.The comparison uses a clean testing set and contrasts normal training with training without false-positive interactions.
- Empirical effects: On Adressa and Amazon-book, normal training drops Recall@20 by 21.81% and 9.56%, respectively, compared with clean training.The reported result shows negative effects on recommending satisfying items.
- Practical implication: Clean training is impractical as a general solution because reliable feedback such as rating scores is sparse.The study therefore focuses on denoising abundant implicit signals such as clicks, views, or purchases.
3 METHOD
The method formulates recommender training around noisy implicit feedback and introduces Adaptive Denoising Training (ADT), which dynamically suppresses large-loss interactions. ADT uses loss behavior during training to distinguish likely false-positive interactions and applies truncation or reweighting within differentiable recommendation models.
- Task Formulation: Implicit-feedback training can learn poor models because noisy interactions mislead preference learning and reduce generalization on clean testing data.The formulation distinguishes reliable preference labels from observed interactions such as clicks and purchases.
- Observations: False-positive interactions have larger losses early in training, but deep recommenders eventually memorize both true- and false-positive interactions.This motivates identifying hard interactions before memorization reduces the usefulness of loss differences.
- Adaptive Denoising Training: ADT dynamically prunes large-loss interactions during training by either truncating their losses or assigning them smaller weights.The strategy estimates false-positive tendencies from training behavior without requiring additional data.
- Truncated Loss: Truncated Loss discards positive interactions whose cross-entropy loss exceeds a training-iteration-dependent threshold.The threshold changes as training proceeds, while the effective loss for interactions above it becomes zero.
- Truncated Loss: The truncation drop rate starts at zero, increases smoothly, and remains bounded so the model can distinguish true- and false-positive interactions gradually.A linear increase limits additional hyper-parameter tuning compared with more complex functions.
- Reweighted Loss: Reweighted Loss reduces the contribution of hard interactions through a dynamically adjustable weight function, instantiated here with cross-entropy loss.The weight is based on prediction scores, with smaller scores on positive interactions indicating larger cross-entropy loss.
4 RELATED WORK
Prior recommendation methods address noisy implicit feedback by identifying negative experiences, incorporating additional feedback, or applying denoising techniques, but often require extra information or target different robustness settings.
- Negative Experience Identification: Negative-experience identification predicts user satisfaction from additional behavioral feedback and item characteristics.These methods can require extensive manual labeling and feature design, limiting practical use as items change.
- Incorporating Various Feedback: Feedback-enrichment approaches incorporate signals such as dwell time, skips, or favorites into recommender training.Some methods distinguish click-complete, click-skip, and non-click items and assign different treatment or weights.
- Robustness of Recommender Systems: Robust recommender research evaluates stability under attacks, while denoising autoencoder methods corrupt inputs before reconstructing them.These approaches address robustness or denoising through settings distinct from the paper’s treatment of false-positive implicit interactions.
5 EXPERIMENT
Experiments evaluate ADT across three datasets, three recommenders, and clean-preference testing protocols. ADT improves recommendation performance and reduces the influence of false-positive interactions, though Truncated Loss can discard many clean interactions.
- Datasets: Experiments use Adressa, Amazon-book, and Yelp, with false-positive interactions identified from short dwell time or ratings below 3.Adressa uses clicks with dwell time under 10 seconds; Amazon-book and Yelp use ratings below 3.
- Experimental Protocol: Models train with false-positive interactions retained but are evaluated only on true-positive test interactions using Recall@K and NDCG@K.The metrics use dataset-specific K values: 50 and 100 for Amazon-book and Yelp, and 3 and 20 for Adressa.
- Overall Performance: ADT improves all tested cases, with NeuMF+T-CE outperforming vanilla NeuMF by 15.62% on average across the three datasets.The comparison includes standard CE, T-CE, and R-CE on GMF, NeuMF, and CDAE.
- Overall Performance: T-CE performs better than R-CE in most cases, while both outperform normal training by reducing the effect of hard interactions.T-CE discards large-loss interactions, whereas R-CE assigns them smaller weights.
- Comparison with Additional Feedback: Additional-feedback methods slightly outperform GMF but remain inferior to T-CE and R-CE, likely because the additional feedback is sparse.The comparison uses NMTR and NR on GMF.
- Memorization of False-positive Interactions: T-CE increases false-positive interaction losses while overall loss decreases, whereas R-CE keeps those losses decreasing but larger than the real training loss.These patterns indicate that T-CE avoids optimizing false-positive interactions and R-CE slows their fitting.
- Study of Truncated Loss: Truncated Loss discards nearly half of false-positive interactions but achieves only 10% precision, so many clean interactions are also removed.Improving precision to reduce clean-interaction loss is identified as future work.
- Hyper-parameter Sensitivity: T-CE performs best when 𝜖_max∈[0.1, 0.3], while performance drops significantly above 0.4; R-CE performs best when 𝛽 ranges from 0.15 to 0.3.The study also finds sensitivity to 𝜖_N, especially on Amazon-book, and a large T-CE hyper-parameter search space.
6 CONCLUSION AND FUTURE WORK
The paper concludes that Adaptive Denoising Training reduces the disturbance of noisy implicit feedback, while remaining broadly applicable and motivating further study across losses and domains.
- ADT’s Truncated Loss and Reweighted Loss paradigms effectively reduce the disturbance caused by noisy implicit feedback.The paper applies both paradigms with binary cross-entropy loss across three recommenders and three datasets.
- Both ADT paradigms are general and can be applied to different recommendation loss functions, neural recommenders, and optimizers.
- The work takes a first step toward denoising implicit feedback without using additional feedback for training.
- Future work includes testing ADT with Square, Hinge, and BPR losses and improving the paradigms’ precision.
- ADT may also be used to denoise implicit interactions in domains such as Web search and question answering.