Denoising Implicit Feedback for Robust Recommendation

19 concepts19 papers1 proposal

Recommender systems primarily rely on implicit feedback—such as clicks, views, and watch history—because it is incredibly abundant. However, implicit feedback suffers from a fundamental data-quality issue known as the interaction-preference gap. Traditional collaborative filtering models operate under a binary assumption: an observed interaction signifies a positive preference (label 1), while an unobserved interaction signifies a negative preference (label 0).

In reality, this assumption is deeply flawed. Observed interactions are riddled with false positives driven by external factors like clickbait, misleading headlines, or accidental taps. Conversely, unobserved interactions contain numerous false negatives—items the user would love but was never exposed to. As highlighted by Wenjie Wang et al. and Jujia Zhao et al., treating these noisy interactions as absolute ground truth misleads models into overfitting on non-representative behaviors, severely degrading recommendation quality. Furthermore, Karthik Raman et al. demonstrate that heuristics which aggressively prioritize clicked items without accounting for feedback noise cause the learning process to oscillate and collapse.

To bridge this gap, a prominent research direction focuses on implicit feedback denoising. Rather than taking user actions at face value, these techniques algorithmically separate genuine preferences from behavioral noise.


Exploiting the Memorization Effect

How can a model distinguish between a true preference and an accidental click without explicit user ratings? One powerful intuition relies on the memorization effect inherent in deep neural networks. During training, models tend to learn simple, clean, and consistent patterns first, only memorizing noisy, contradictory data points in later epochs.

Wenjie Wang et al. leverage this phenomenon through Adaptive Denoising Training (ADT). By monitoring the loss values of interactions during the early stages of training, ADT automatically identifies potential false positives. If an interaction consistently incurs a high loss early on, it means the model is struggling to reconcile it with the user's broader, cleaner preference patterns. The framework dynamically down-weights or filters these high-loss samples before the model has a chance to memorize them.

However, a high loss value can sometimes indicate a genuinely unusual but valid user preference rather than noise. To prevent the erasure of these unique preferences, Kaike Zhang et al. introduced Personalized Loss Denoising (PLD). Instead of evaluating loss against a global average, PLD analyzes an interaction's loss relative to a user's personal loss distribution. If a click aligns with a user's established historical patterns, it is retained, even if it looks like noise from a global perspective.


Quantifying Modeling Effort and Cross-Model Agreement

Another way to isolate noise is to measure how much "effort" a model must exert to fit an interaction, or to see if different model architectures agree on its validity.

True user preferences naturally form tight, coherent geometric structures in an embedding space. Noisy interactions, by contrast, act as structural outliers. Yanchao Tan et al. utilize Optimal Transport (ProRec) to calculate the mathematical "modeling effort" required to match users to items. Interactions that require an excessive transport cost to align are flagged as anomalous curiosity clicks and filtered out. In session-based settings, Jongwon Jeong et al. introduce a metric learning framework (FPAdaMetric) that actively pushes identified false positives away from the user's current session context, adapting the penalty based on the estimated "degree of dislike."

Alternatively, we can look for architectural consensus. Xin Xin et al. propose DeCA (Denoising with Cross-Model Agreement), which trains two distinct recommendation models (e.g., a graph neural network and a matrix factorization model) simultaneously. Because different architectures possess different inductive biases, they are unlikely to overfit to noise in the exact same way. If the models disagree on the predicted affinity for an observed interaction, the interaction is treated as a false positive driven by external biases like item popularity.


Generative and Structural Denoising

Rather than simply filtering or reweighting raw interaction data, several approaches clean the signals by transforming the underlying representations or graph structures.

Jujia Zhao et al. apply generative diffusion principles to recommendation with the Denoising Diffusion Recommender Model (DDRM). By treating interaction noise as a form of data corruption, DDRM uses a multi-step iterative reverse diffusion process directly on user and item embeddings to gradually strip away the noise, reconstructing a cleaner representation of core user interests.

[Noisy Interaction Graph] ──> [Bayesian Graph Learning / Attention] ──> [Latent Interest Center]
                                          │                                       │
                                          ▼                                       ▼
                             Smoothes out missing/false edges        Pushes away hard negatives

From a structural perspective, graph-based recommenders often suffer because they treat the observed interaction graph as an immutable ground truth. Soumyasundar Pal et al. address this by treating the graph structure itself as a random variable within a non-parametric Bayesian framework, allowing the model to learn and smooth out missing or noisy edges during training. Similarly, Jiangqiang Zhu et al. combine graph attention networks with self-supervised contrastive learning to build representations that are inherently robust against sparse and noisy edges. To bypass individual item noise entirely, Shipeng Song et al. replace individual positive items with a calculated "interest center" in a contrastive learning space, anchoring user profiles to aggregate intent rather than isolated clicks.


Dual-Sided Denoising and Error Imputation

While early paradigms focused almost exclusively on eliminating false positive clicks, modern frameworks recognize that true robustness requires a dual-sided approach that simultaneously corrects both false positives and false negatives.

In passive interaction environments—such as infinite scroll video feeds where users do not explicitly click or rate content—the absence of clear signals makes denoising exceptionally difficult. Guanyu Lin et al. demonstrate that simultaneously addressing both sides of the noise spectrum via an inverse meta-learning strategy yields massive accuracy gains over traditional negative sampling.

To formalize this mathematically, Haoxuan Li et al. introduce error-imputation-based estimators that handle selection bias and noisy feedback concurrently. These estimators predict the true underlying error for missing ratings while adjusting for the noise present in observed ratings. By treating denoising as a global optimization problem rather than a collection of local heuristics, frameworks like the bi-level optimization model (BOD) developed by Zongwei Wang 0002 et al. can learn optimal sample weights globally, ensuring the final preference matrix reflects true user affinity rather than accidental behavior.

Key Foundational Work

Denoising Implicit Feedback for Recommendation (2020) — contributed 1 concept elements to this direction.
Non Parametric Graph Learning for Bayesian Graph Neural Networks (2020) — contributed 1 concept elements to this direction.
Stable Coactive Learning via Perturbation (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_implicit_feedback_noise_3493e73f,
  title = {implicit feedback noise},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-feedback-noise/rdr_3493e73fecaa402b83cc8270f405dc4b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_implicit_feedback_noise_3493e73f_v1,
  title = {implicit feedback noise},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-feedback-noise/rdr_3493e73fecaa402b83cc8270f405dc4b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake