PaperlayerView exact revision on arXiv

Source-linked AI summary

Semantics-Aware Denoising: A PLM-Guided Sample Reweighting Strategy for Robust Recommendation

Xikai Yang, Yang Wang, Yilin Li, Sebastian Sun

arXiv:2602.15359v1cs.IR

TL;DR

Noisy clicks can misrepresent user preferences and degrade recommendation quality. SAID constructs textual user-interest profiles, compares them with item descriptions using PLM-based encoders, and converts similarity into loss weights; experiments show consistent gains and robustness under high noise. Its scope is limited when textual item descriptions are unavailable, and its simple profile construction and uniform noise treatment leave room for improvement.

  • Problem

    Implicit clicks include accidental, clickbait-induced, and exploratory interactions that may not represent genuine preferences, degrading recommendation quality.

  • Method

    SAID uses PLM-based semantic similarity between textual user-interest profiles and item descriptions to produce sample weights that downweight potentially noisy clicks without changing the backbone architecture.

  • Results

    SAID consistently outperforms all baselines on two datasets, achieving relative AUC improvements of 1.65% on MovieLens-1M and 2.18% on Amazon-Book over vanilla DeepFM, with stronger robustness as noise increases.

  • Takeaways & Limitations

    Semantic consistency provides a soft denoising signal that reduces noise impact while preserving valid exploratory clicks through a minimum-weight guarantee.

  • Takeaways & Limitations

    SAID assumes textual item descriptions are available, uses relatively simple user-interest profiles, and treats different noise types equally.

Abstract

from arXiv · show

Implicit feedback, such as user clicks, serves as the primary data source for modern recommender systems. However, click interactions inherently contain substantial noise, including accidental clicks, clickbait-induced interactions, and exploratory browsing behaviors that do not reflect genuine user preferences. Training recommendation models with such noisy positive samples leads to degraded prediction accuracy and unreliable recommendations. In this paper, we propose SAID (Semantics-Aware Implicit Denoising), a simple yet effective framework that leverages semantic consistency between user interests and item content to identify and downweight potentially noisy interactions. Our approach constructs textual user interest profiles from historical behaviors and computes semantic similarity with target item descriptions using pre-trained language model (PLM) based text encoders. The similarity scores are then transformed into sample weights that modulate the training loss, effectively reducing the impact of semantically inconsistent clicks. Unlike existing denoising methods that require complex auxiliary networks or multi-stage training procedures, SAID only modifies the loss function while keeping the backbone recommendation model unchanged. Extensive experiments on two real-world datasets demonstrate that SAID consistently improves recommendation performance, achieving up to 2.2% relative improvement in AUC over strong baselines, with particularly notable robustness under high noise conditions.

I. INTRODUCTION

Implicit clicks provide essential training signals for CTR prediction but include accidental, clickbait-induced, and exploratory interactions that may not reflect genuine interest. SAID uses semantic consistency between historical user interests and clicked-item content to identify and downweight potentially noisy samples.

  • I. INTRODUCTION: Noisy implicit-feedback samples can produce irrelevant recommendations and a feedback loop that further degrades data quality.Models may learn to predict accidental or exploratory clicks, leading to poor recommendations and more noisy interactions.
  • I. INTRODUCTION: Existing denoising methods use auxiliary confidence networks or contrastive learning but often add computational overhead and tuning requirements.
  • I. INTRODUCTION: SAID identifies potentially noisy clicks by comparing semantic consistency between users’ historical interests and clicked items.A semantically inconsistent click, such as a sudden fashion click from a user consistently engaging with technology content, is treated as more likely to be noise.
  • I. INTRODUCTION: SAID constructs textual user-interest representations, computes PLM-based similarity with item descriptions, and converts similarity scores into loss weights.The weighting mechanism downweights potentially noisy samples while leaving the recommendation backbone unchanged.

II. METHODOLOGICAL FOUNDATIONS

The methodological foundations connect CTR models and graph-based recommendation with semantic and structural approaches to handling noisy interaction data. SAID builds on encoder-based PLMs and sample-level weighting rather than using generative language models for direct recommendation.

  • II. METHODOLOGICAL FOUNDATIONS: CTR architectures such as Wide & Deep, DeepFM, and DIN capture feature interactions or user interests but remain vulnerable to noisy training data.
  • II. METHODOLOGICAL FOUNDATIONS: Graph recommenders model structural user-item signals, yet NGCF and LightGCN are likewise susceptible to noise in interactions.
  • II. METHODOLOGICAL FOUNDATIONS: Graph-transformer and temporal-graph approaches motivate using structural and temporal consistency to distinguish informative signals from noise.
  • II. METHODOLOGICAL FOUNDATIONS: Encoder-based PLMs such as BERT and Sentence-BERT provide semantic representations and similarity computation for recommendation-related tasks.
  • II. METHODOLOGICAL FOUNDATIONS: SAID uses semantic understanding for denoising and combines contrastive-knowledge-transfer and robust-optimization ideas with sample-level weighting.

III. METHODOLOGY

The problem formulation distinguishes observed clicks from latent true preferences in implicit-feedback CTR prediction. SAID targets positive samples that are observed clicks but do not correspond to genuine preference by downweighting them during training.

  • III. METHODOLOGY: CTR prediction uses observed clicks as positive samples and sampled non-clicked items as negative samples for user-item probability estimation.
  • III. METHODOLOGY: Not every positive click reflects genuine user interest, so observed labels can differ from unobserved true preference labels.
  • III. METHODOLOGY: For noisy interactions, the observed click label is 1 while the true preference label is 0.
  • III. METHODOLOGY: The training objective is to identify and downweight noisy positive samples rather than treating every click as equally reliable.

B. Semantic Similarity Computation

SAID constructs user-interest profiles from textual historical-item information, encodes profiles and candidate descriptions with a pretrained text encoder, and measures their semantic similarity. Low similarity indicates potential noise and supplies the signal for subsequent sample reweighting.

  • B. Semantic Similarity Computation: User-interest profiles aggregate textual information from historical interactions, including titles, categories, and content descriptions.The stated goal is to represent stable and transient user-intent patterns more comprehensively than a single textual facet.
  • B. Semantic Similarity Computation: The implementation can concatenate the most recent k item titles for computational efficiency.
  • B. Semantic Similarity Computation: A pretrained text encoder maps user profiles and item descriptions into semantic embeddings.Sentence-BERT is used as the encoder for semantic similarity computation.
  • B. Semantic Similarity Computation: Semantic similarity between a user and item is computed with cosine similarity between their embeddings.
  • B. Semantic Similarity Computation: A low similarity score signals semantic inconsistency between a clicked item and historical interests, suggesting potential noise.

C. Sample Reweighting

SAID uses soft semantic reweighting to reduce the influence of suspicious clicks without discarding potentially valid exploratory behavior.

  • C. Sample Reweighting: Soft denoising reduces suspicious samples’ influence while preserving their contribution, because semantic inconsistency can reflect exploration or interest drift.The method avoids hard filtering so potentially valid exploratory clicks remain available during training.
  • C. Sample Reweighting: The weighting function gives higher weights to more similar items, assigns non-zero weights to low-similarity samples, and changes smoothly between them.These constraints are intended to avoid abrupt decisions and complete exclusion of exploratory clicks.
  • C. Sample Reweighting: The sigmoid weighting function uses α as the minimum weight, β to control transition sharpness, and μ as the similarity threshold.The minimum weight ensures every sample contributes to training.
  • C. Sample Reweighting: The global mean similarity score supplies μ, producing stable and reproducible assignments across batches.A global threshold keeps treatment consistent across batch compositions and deployment settings.
  • C. Sample Reweighting: Consistent clicks receive weights near 1, whereas inconsistent clicks receive reduced weights bounded below by α.The supplied passage links this bounded soft weighting to avoiding the performance harm of hard filtering.

D. Weighted Loss Function

SAID incorporates semantic sample weights into the training loss, while leaving negative samples unweighted and targeting false-positive click noise.

  • D. Weighted Loss Function: The weighted binary cross-entropy loss incorporates sample weights into the backbone model’s CTR prediction objective.The supplied equation passages identify the loss as weighted binary cross-entropy and define the prediction as backbone-model CTR.
  • D. Weighted Loss Function: Negative samples receive weight 1, so weighting applies only to positive samples to reduce the impact of noisy clicks.This design focuses denoising on false-positive interactions while keeping non-clicked items unchanged.

E. Framework Overview

SAID separates semantic weight computation from CTR prediction and connects both pipelines through a weighted loss without changing the recommendation backbone.

  • E. Framework Overview: Two parallel pipelines compute similarity-based weights and generate CTR predictions, with the weighted loss integrating them for end-to-end training.The framework preserves the backbone recommendation model architecture.
  • E. Framework Overview: Semantic embeddings can be cached offline, leaving training with only weight computation based on simple arithmetic operations.The text encoder requires one forward pass per item for embedding computation.

A. Experimental Setup

The evaluation uses two benchmark datasets, several recommendation baselines, specified implementation settings, repeated runs, and AUC and Logloss metrics.

  • A. Experimental Setup: Experiments use MovieLens-1M and 5-core Amazon-Book, with MovieLens ratings binarized into 575,281 positive interactions.MovieLens-1M contains approximately 1 million ratings from 6,040 users and 3,706 movies.
  • A. Experimental Setup: SAID is compared with DeepFM, DeepFM+T-CE, DeepFM+DNS, and LightGCN under the CTR prediction framework.The baselines include denoising, dynamic negative-sampling, and graph-based recommendation approaches.
  • A. Experimental Setup: Implementation uses all-MiniLM-L6-v2, profiles built from the most recent 10 items, and a DeepFM backbone with 64-dimensional embeddings and three hidden layers.The reported settings include α=0.4, β=5, and μ equal to the global mean similarity score.
  • A. Experimental Setup: Each experiment runs five times with different random seeds, and the reported values are means.Repeated runs provide the stated aggregation procedure for the evaluation results.
  • A. Experimental Setup: Evaluation uses AUC and Logloss, where higher AUC and lower Logloss indicate better performance.These metrics assess ranking discrimination and prediction loss under the stated evaluation setup.

B. Overall Performance

SAID consistently improves recommendation performance over baselines on MovieLens-1M and Amazon-Book, with larger gains on the sparser Amazon-Book dataset and stronger robustness as noise increases.

  • B. Overall Performance: SAID outperforms all baselines on both datasets, improving AUC over vanilla DeepFM by 1.65% on MovieLens-1M and 2.18% on Amazon-Book.Logloss also improves, indicating better calibrated probability estimates.
  • B. Overall Performance: SAID improves both evaluation metrics on both datasets, with more pronounced gains on the sparser Amazon-Book dataset.The results suggest semantic information is particularly valuable when behavioral signals are limited.
  • B. Overall Performance: As injected noise increases from 0% to 50%, all methods degrade, but SAID maintains significantly better performance on MovieLens-1M.The experiment flips randomly selected negative labels to positive to evaluate robustness under controlled noise.

D. Parameter Sensitivity

SAID’s soft denoising design balances noise reduction with preservation of valid exploratory clicks, achieving its best performance at α=0.4 without strong parameter sensitivity.

  • D. Parameter Sensitivity: Setting α=0 causes hard filtering that removes low-similarity samples entirely, hurting performance by discarding valid exploratory clicks.Exploratory clicks can represent genuine interest drift, so suspicious interactions should retain some training influence.
  • D. Parameter Sensitivity: The optimal minimum weight α=0.4 balances noise reduction and information preservation, while the relatively flat curve indicates limited sensitivity around the optimum.At α=1, all samples receive equal weight, reducing SAID to the baseline without denoising.
  • D. Parameter Sensitivity: SAID’s current limitations include requiring textual item descriptions, using relatively simple user-profile construction, and treating different noise types equally.The paper identifies attention-based aggregation, generative LLM summarization, and noise-specific handling as possible future directions.
Loading 2602.15359v1…