Source-linked AI summary
Beyond a Single Story: Meta-Reviewing Sparse and Incomplete User-generated Contents for Recommendation
Hongren Wang, Tianjun Wei, Yingpeng Du, Jie Zhang, Yin-Leng Theng
TL;DR
UGC-based recommenders face missing reviews and incomplete attribute coverage, even though reviews provide fine-grained preferences. MOSAIC aggregates attribute-sentiment evidence from neighbor users, jointly models ratings and sentiment with MMoE, and personalizes the result with attention. Across four real-world datasets, it consistently outperforms state-of-the-art baselines and gains remain consistent for users with limited interaction history.
Problem
UGC reviews provide fine-grained preferences but are especially sparse, and individual reviews often omit relevant attributes.
Method
MOSAIC constructs attribute-level meta-reviews from neighbor reviews, jointly predicts ratings and meta-review sentiment with MMoE, and personalizes signals through attention.
Results
MOSAIC consistently outperforms state-of-the-art baselines on four real-world datasets, with consistent gains for users with limited interaction history.
Takeaways & Limitations
Neighbor-aggregated meta-reviews expand attribute coverage beyond single-user reviews while supporting refined predictions and attribute-level explanations.
Takeaways & Limitations
MOSAIC relies on predefined domain-specific attributes and neighbor links based on shared item interactions or optional social connections.
Abstract
from arXiv · showhide
Data sparsity remains a long-standing challenge in recommender systems, and it becomes more severe for methods relying on user-generated content (UGC) such as textual reviews, which capture fine-grained preferences but require more user efforts to produce. As a result, UGC exhibits (1) missing reviews, where interactions lack any review, and (2) incomplete reviews, where available reviews cover only a subset of relevant attributes. Existing approaches often overlook these UGC-specific issues, leading to degraded accuracy. Motivated by meta-review in academic peer review, we propose MOSAIC (Meta-review On Sparse And Incomplete user-generated Content), which constructs a meta-review for each target user by aggregating attribute-sentiment evidence from neighbor users' reviews. A multi-gate mixture-of-experts (MMoE) architecture jointly optimizes rating prediction and meta-review attribute-sentiment prediction, while an attention module personalizes the aggregated meta-review signals to each target user, yielding both refined rating predictions and attribute-level explanations. Experiments on four real-world datasets demonstrate that MOSAIC consistently outperforms state-of-the-art baselines in both recommendation accuracy and explanation quality, mitigating UGC sparsity and incompleteness while delivering consistent gains for users with limited interaction history.
1 Introduction
UGC offers fine-grained preference information but is especially sparse and incomplete because reviews require user effort and omit relevant attributes. MOSAIC addresses these gaps by aggregating attribute-level evidence from neighbor reviews and personalizing it for recommendation and explanation.
- UGC captures richer preferences than clicks and ratings but is even sparser because writing reviews requires considerable effort.
- Individual reviews often omit relevant attributes, leaving gaps that single-user review methods cannot fully capture.Users may mention one aspect while omitting another preference, such as battery life.
- MOSAIC aggregates attribute-sentiment evidence from neighbor users to complement absent or insufficient fine-grained preference signals.Its meta-review construction operates at the attribute–sentiment level rather than aggregating reviews only as whole instances.
- MOSAIC uses a multi-task framework and personalized attention to refine rating prediction while producing attribute-level explanations.The attention module selectively highlights attributes relevant to the target user.
- Experiments on four public benchmark datasets report superiority over state-of-the-art baselines, with consistent gains for users having limited interaction history.Ablation studies also assess the effectiveness of the proposed components.
2 Related Work
Related work uses review text for recommendation and explanation through neural, graph, contrastive, and LLM-based methods. These approaches generally depend on review content that is already available for the user or user–item pair.
- Explainable recommenders either extract existing review phrases or generate explanations with neural, pretrained-language, and LLM-based models.Hybrid methods ground generation on retrieved reviews or attributes to reduce hallucination.
- Existing review interpretation methods remain dependent on available review content, limiting supervision when user–item pairs lack reviews.
- Review-based recommenders encode review semantics with CNNs, attention, graph, contrastive-learning, and LLM-based aspect-extraction methods.
3 Motivation
The motivation is that individual reviews provide incomplete attribute coverage, while LLM extraction and neighbor aggregation can create broader, structured preference evidence. On TripAdvisor, aggregating reviews from as few as 10 neighbors raises attribute coverage above 80%.
- 3.1 LLM-Based Attribute-Sentiment Extraction: MOSAIC uses an LLM to extract concise attribute-sentiment pairs from unstructured reviews for interpretable explanations and recommendation features.
- 3.1 LLM-Based Attribute-Sentiment Extraction: The extraction uses a predefined domain-specific attribute set, while unmentioned or non-polarized attributes receive an empty sentiment set.Sentiment phrases are restricted to a vocabulary limited to three words.
- 3.1 LLM-Based Attribute-Sentiment Extraction: Extracted attribute-sentiment text is classified by fine-tuned BERT into four sentiment classes: positive, slightly positive, slightly negative, and negative.
- 3.1 LLM-Based Attribute-Sentiment Extraction: The LLM is used offline for extraction rather than during model training or inference, avoiding additional computational cost.
- 3.2 Incompleteness of User Review: Review incompleteness differs from missing reviews because an available review may still omit attributes the user considered.The paper identifies incompleteness as an overlooked form of sparsity in prior work.
- 3.2 Incompleteness of User Review: On TripAdvisor, half of six predefined attributes had coverage below 50% across all reviews.The same hotel could receive identical ratings in two reviews while different attributes were mentioned.
- 3.2 Incompleteness of User Review: Aggregating reviews from as few as 10 neighbors raises attribute coverage above 80% and supplies preference signals when the target user left no review.
4 Methodology
MOSAIC combines neighbor-derived attribute-sentiment meta-reviews with multitask prediction and target-user personalization. Its architecture jointly models ratings and sentiment signals, then uses attention and multiple losses to refine recommendations and explanations.
- Framework overview: MOSAIC jointly predicts ratings and attribute-sentiment labels from user-level and group-level meta-review signals.The framework uses MMoE with task-specific towers to balance the two prediction tasks and reduce negative transfer.
- Meta-review construction: Neighbor users who reviewed the same item are sampled to construct the meta-review, with majority voting selecting sentiment labels for each attribute.The method samples N = 15 neighbors and uses majority voting to reduce noise and sentiment heterogeneity.
- Input representation: Rating prediction uses target-user and item embeddings, whereas sentiment prediction additionally incorporates an aggregated neighbor embedding.The two task inputs therefore differ in whether they include neighborhood information.
- Multi-task integration: Task-specific MMoE gates dynamically weight shared experts, producing distinct representations for rating and meta-review sentiment prediction.The resulting representations feed separate towers: an MLP for ratings and an MLP with softmax for attribute-wise sentiment distributions.
- Personalized refinement: Personalized self-attention uses rating features as queries and predicted attribute-sentiment signals as keys and values to emphasize user-relevant attributes.The resulting sentiment representation is fused with rating features to produce a meta-review-guided rating adjustment.
- Loss functions: The final objective combines rating loss with meta-review sentiment loss based on classification and cosine-similarity supervision.Joint supervision uses user-level labels, majority-voted group-level labels, and the richer meta-review sentiment vector.
5 Experiments
MOSAIC is evaluated on four public datasets for rating prediction, attribute-sentiment prediction, efficiency, component contributions, sensitivity, and sparse-user performance. It consistently outperforms comparison methods, with especially large gains for sparse users and improvements linked to neighbor aggregation and personalized attention.
- Experimental setting: MOSAIC is evaluated on Yelp, TripAdvisor, Amazon Beauty, and Amazon Sports using rating and attribute-sentiment prediction metrics.Rating prediction uses RMSE and MAE; attribute-sentiment prediction uses Accuracy, F1 Score, Precision, and Recall.
- Overall performance: MOSAIC consistently outperforms all comparative methods on both rating-prediction metrics across all four datasets.Comparisons include collaborative filtering, review-based, joint rating–explanation, and LLM-based recommenders.
- Overall performance: MOSAIC achieves the best overall meta-review attribute-sentiment prediction performance, with more balanced F1 and recall than precision-oriented baselines.Some baselines obtain slightly higher precision, but their lower F1 and recall indicate more conservative positive-label predictions.
- Ablation and sensitivity analysis: Removing neighbor users causes the largest ablation drop, including RMSE worsening from 0.6759 to 0.7974 on TripAdvisor.Without neighbor-derived meta-reviews, the model loses broader attribute coverage and cross-user evidence for incomplete reviews.
- Ablation and sensitivity analysis: Increasing neighbor users generally improves RMSE, MAE, and label accuracy, while 10–15 neighbors provide over 80% attribute coverage.Replacing personalized attention with a uniform MLP also worsens performance, including Beauty RMSE increasing from 0.7840 to 0.9252.
- Performance on sparse users: MOSAIC outperforms all baselines across sparse, moderate, and active user groups, reducing sparse-user MAE by 30.38% on Yelp and 26.90% on Amazon Sports.The sparse group contains users with at most one training interaction.
6 Conclusion
MOSAIC addresses UGC sparsity and incompleteness by aggregating attribute-sentiment evidence from neighboring users. Its MMoE and personalized attention components jointly support refined rating prediction and attribute-level explanations, with consistent gains over state-of-the-art baselines.
- MOSAIC addresses UGC sparsity and incompleteness by aggregating attribute-sentiment evidence from neighbor users.
- The framework combines an MMoE-based multitask architecture with personalized attention conditioned on the target user.These components jointly produce refined rating predictions and attribute-level explanations.
- MOSAIC consistently outperforms state-of-the-art baselines across four real-world datasets.
- Meta-reviews expand attribute coverage beyond any single user’s reviews and maintain gains for users with limited interaction history.
Limitations
MOSAIC has limitations concerning its predefined attribute taxonomy, neighbor-selection strategy, and reliance on user-side evidence. Its LLM-based extraction pipeline also leaves residual hallucination-related errors that may propagate into meta-reviews.
- A predefined domain-specific attribute taxonomy cannot capture emerging or niche attributes outside the predefined set.Adapting MOSAIC to new domains requires reconstructing the attribute list.
- Neighbor selection may be sparse for highly isolated users because it relies on shared item interactions and optional social links.Graph-based or behavior-similarity neighbor finding is proposed as a possible extension.
- MOSAIC focuses on user-side neighbor-aggregated evidence, leaving complementary item-side reasoning for future hybrid designs.
- Residual extraction errors from LLM hallucinations may still propagate into meta-reviews despite constrained extraction, verification, and majority voting.The LLM is used offline, outputs are taxonomy- and vocabulary-constrained, sentiments are verified by BERT, and neighbor evidence is aggregated by attribute-level majority voting.
B.1 Datasets
The evaluation uses four public datasets spanning restaurant, hotel, and product-review domains, together with classical, graph-based, hypergraph, LLM-based, and explainable recommendation baselines. Three datasets retain approximately 100,000 processed interactions with attribute-sentiment pairs.
- Datasets: The datasets comprise Yelp restaurant reviews, TripAdvisor hotel reviews, and Amazon Beauty and Amazon Sports product reviews.They provide textual reviews, user-item interactions, and multi-aspect or subjective user feedback across domains.
- Datasets: Approximately 100,000 high-quality user-item interactions with attribute-sentiment pairs remain in Yelp, TripAdvisor, and Amazon Sports after preprocessing.
- Baselines: Baselines include PMF and SVD++, which estimate ratings using latent user and item factors.
- Baselines: The comparison also covers LLMRec, RGCL, APH, NETE, PETER, PEPLER, CER, and SERMON across graph, aspect, explanation, and multimodal recommendation settings.LLMRec uses gpt-4.1-nano here for fair comparison with MOSAIC’s extraction pipeline.
B.3 Experimental Settings
Experiments use per-user 8:1:1 train-validation-test splits, fixed optimization and embedding settings, and repeated runs on one GPU. Rating prediction and attribute-sentiment explainability are evaluated with separate regression and classification metrics.
- Training Setting: Each dataset is split per user into training, validation, and test sets at an 8:1:1 ratio, with at least one interaction in each set.
- Training Setting: All models use Adam with learning rate 1 × 10^-5, embedding size 512, and batch size 512; MOSAIC uses three MMoE experts.
- Training Setting: Experiments run five times on a single NVIDIA RTX A6000 48GB GPU with statistical significance analysis.
- Evaluation Setting: Rating prediction is evaluated with RMSE and MAE, while attribute-sentiment prediction uses Accuracy, F1 Score, Precision, and Recall.The explainability task predicts attribute-sentiment labels rather than generating textual explanations.
B.4 Computational Cost Analysis
MOSAIC’s computational analysis reports training-stage recommendation-model cost only, excluding offline preprocessing. The model uses relatively low memory and trains quickly compared with representative baselines.
- Training-cost reporting excludes offline preprocessing and covers only the recommendation models’ training stage.
- MOSAIC requires relatively low memory and achieves fast training speed compared with representative baselines.Its per-epoch and total training times are comparable to lightweight models and substantially below generation-based or complex architectures.
- The reported results indicate improved recommendation performance without additional computational overhead for practical deployment.
B.5 Additional Analysis of Gate Distributions
Additional gate analysis shows that increasing the number of experts does not eliminate specialization, but it fragments existing specialists and reduces performance at larger K.
- At K = 5, every attribute remains routed to a dominant expert and all five experts are used.
- RMSE degrades from 0.676 to 0.708 (+4.7%) when increasing the expert count to K = 5.
- The effective number of experts per attribute rises from 1.51 at K = 3 to 1.95 at K = 5 and 2.39 at K = 7.The measure is defined as the exponential of gate entropy, with 1.0 representing a dedicated expert.
- Additional experts fragment existing specialists, weakening training signals while increasing parameter count and reducing sample efficiency.
C.1 Explainable Recommender Systems
Explainable recommender systems derive justifications from reviews through extraction or generation, with newer hybrid methods retrieving attributes or reviews to reduce hallucination. However, these approaches remain limited by review dependence and incompleteness.
- Extraction-based explainers reuse existing phrases or reviews, whereas generation-based models produce natural-language explanations with neural language architectures.
- Hybrid explainers retrieve relevant attributes or historical reviews before generating explanations to reduce hallucination.
- Existing explanation methods depend on contextual review features and therefore remain subject to hallucination and incompleteness.
- Review-based recommendation leverages fine-grained semantic and opinion information from reviews to improve rating prediction beyond ID-based models.
- Prior recommendation methods encode review semantics using CNNs, attention, graph or contrastive learning, and LLM-based aspect extraction.