PaperlayerView exact revision on arXiv

Source-linked AI summary

Contrastive Learning for Aspect Representation towards Explainable Recommendation

Emrul Hasan, Chen Ding

arXiv:2610.07761v1cs.IRcs.AI

TL;DR

Traditional recommendation approaches can lack interpretability, while prior methods that map user and item IDs into text increase complexity and can produce inconsistencies with natural-language explanations. CLARER combines ID-based features with aspect-aware review representations, and experiments report that it outperforms state-of-the-art baselines in accuracy and explainability.

  • Problem

    Traditional latent-feature recommendations can lack interpretability, and mapping IDs to text can increase complexity and create inconsistencies with natural-language explanations.

    Source: §I · p. 1

  • Method

    CLARER combines user- and item-ID features with aspect-aware review representations learned using domain-specific aspects and contrastive learning, then uses the combined representation for explanation generation.

    Source: §I · pp. 1–2, D. Integration of ID-based Features… · p. 4

  • Results

    CLARER outperforms state-of-the-art baselines in accuracy and explainability across three real-world datasets.

    Source: §I · p. 2, §VI · p. 7

  • Takeaways & Limitations

    Aspect-centric alignment between user preferences and item features is identified as important for recommendation quality in multi-criteria settings.

    Source: D. Discussion · p. 7

Abstract

from arXiv · show

In this work, we propose a novel recommendation model, CLARER (Contrastive Learning for Aspect Representation towards Explainable Recommendation) that integrates aspect features learned from textual reviews with rating information to improve the accuracy and explainability of recommendations. Our proposed framework learns user and item representations by combining rating-based features and aspect-based features from reviews. Specifically, rating-based features are learned through a multi-layer perceptron (MLP) model, while aspect-specific review representations are learned using a transformer encoder to capture the semantic information and contrastive learning to better distinguish user preferences. To provide explanations, we train a transformer decoder, using the final representations of users and items from both rating and aspect-based features as context. Experimental results in three benchmark data sets demonstrate that our model achieves superior performance compared to baseline methods in both recommendation (accuracy) and explanation generation.

I. INTRODUCTION

CLARER combines ID-based features with aspect-aware representations learned from reviews to improve recommendation and explanation generation. It addresses challenges in aligning user preferences with language and extracting useful aspects from noisy reviews.

  • ID-based features can misalign with natural language and, under sparse rating data, may represent users and items inaccurately, limiting personalization and performance.Source: §I · p. 1
  • External aspect-extraction tools can be less adaptable across domains, motivating a self-contained way to capture aspect-based preferences from noisy reviews.Source: §I · p. 1
  • CLARER jointly combines aspect-based representations learned from reviews with user-item latent factors to improve rating prediction and explanation generation.Source: §I · p. 1
  • The framework uses predefined domain-specific aspects with contrastive learning to distinguish aspect representations.Source: §I · p. 2
  • Across three real-world datasets, CLARER outperforms state-of-the-art baselines in both accuracy and explainability, with aspect-aware representations driving performance gains.Source: §I · p. 2

II. RELATED WORK

Prior work represents review aspects and uses user-item information or textual reviews to personalize explanations. CLARER targets gaps in semantic representation and in combining rating and aspect features for both tasks.

A. Aspect Representation

  • Topic-based review methods such as LDA, and aspect-aware extensions often rely on bag-of-words, limiting the semantic information they capture.Source: A. Aspect Representation · p. 2
  • Prior aspect-specific representations for explainable recommendation are underexplored; CLARER contrasts predefined aspects with whole-review representations rather than relying on external extraction tools.Source: A. Aspect Representation · p. 2

B. Explainable Recommendation Systems

III. METHODS

CLARER predicts ratings and generates explanations from user, item, review, and explanation representations. Its user-item representation uses an MLP to combine ID-based features with aspect-level review features.

III. METHODS

  • CLARER is presented as a framework with three main components.Source: §III · p. 2

A. Priliminaries

  • The task is to predict an overall user-item rating and, during training, an explanation that justifies the recommendation.Source: A. Priliminaries · p. 2

B. User and Item Representation

  • User and item embeddings are concatenated and passed through an MLP; its output is combined with review-derived aspect features for rating prediction.The MLP is chosen for simplicity, computational efficiency, and ability to model nonlinear interactions.Source: B. User and Item Representation · p. 3

C. Aspect Representation

CLARER learns review-based aspect representations using domain-specific aspects and contrastive learning, then combines them with ID-based features for recommendation and explanation.

C. Aspect Representation

  • Review-derived aspect representations are combined with ID-based features for rating prediction and explanation generation.The method treats aspects as domain-dependent.Source: C. Aspect Representation · p. 3
  • TripAdvisor aspects come from its platform, while Amazon and Yelp aspects are selected and defined through data analysis.The paper gives topic analysis as an example.Source: C. Aspect Representation · p. 3, Table I · p. 3
  • Pre-trained sentence-transformer embeddings represent predefined aspects, and a transformer encoder generates review representations for contrast with those embeddings.The encoder processes token inputs through self-attention and feed-forward layers, with positional encodings added to token embeddings.Source: C. Aspect Representation · p. 3
  • Contrastive learning treats the review as a positive counterpart to each aspect and samples negative aspect embeddings to train aspect-level representations.The review is assumed to reflect the aspect; loss is averaged across aspects.Source: C. Aspect Representation · pp. 3–4

D. Integration of ID-based Features with Aspect Representation

CLARER uses a sigmoid gate to integrate ID-based user and item representations with review-derived aspect features into a combined representation.

D. Integration of ID-based Features with Aspect Representation

E. Rating Prediction

The model predicts ratings from combined rating-and-review features and generates explanations autoregressively from user-item context and encoder output.

E. Rating Prediction

F. Explanation Generation

  • The transformer decoder generates explanations autoregressively, attending to previous user-item representations and encoder output while masking future tokens.Source: F. Explanation Generation · p. 4
  • Explanation generation uses cross-entropy training, with the total explanation loss defined over training examples and ground-truth explanations.Source: F. Explanation Generation · p. 4

G. Multi-task Learning

CLARER trains rating prediction, aspect learning, and explanation generation together to account for personalization in explanations.

G. Multi-task Learning

  • A multi-task objective trains rating prediction, aspect learning, and explanation generation simultaneously.Source: G. Multi-task Learning · p. 4
  • Joint training is proposed because individual task training ignores the influence of personalization in explanation generation.Source: G. Multi-task Learning · p. 4
  • The objective weights user-item representation, aspect learning, and explanation generation losses, with a separate hyperparameter controlling parameter regularization strength.Source: G. Multi-task Learning · p. 4

IV. EXPERIMENTS

The experiments evaluate rating prediction and explanation generation on three review datasets, using rating-error and text-similarity metrics. The paper compares performance with state-of-the-art models.

A. Dataset

  • The study uses Hong Kong TripAdvisor, Yelp Challenge 2019, and Amazon 5-core Movies & TV data, filtering out users and items with fewer than 20 interactions.For explanation generation, one review sentence containing a feature is selected as ground truth; dataset statistics are reported in Table II.Source: A. Dataset · p. 4, Table II · p. 5

B. Evaluation Metrics

  • Rating prediction is evaluated with RMSE and MAE; explanation generation uses BLEU-1, BLEU4, ROUGE-2, and ROUGE-L F1 scores.Higher text-generation scores indicate better alignment with the ground-truth text.Source: B. Evaluation Metrics · p. 5

C. Baseline Methods

  • The experiments compare task performance with state-of-the-art models discussed in the following baseline sections.Table III is identified as reporting rating-prediction performance using MAE and RMSE.Source: C. Baseline Methods · p. 5, Table III · p. 5

1) Rating Prediction:

CLARER outperforms rating-focused and joint recommendation baselines on rating prediction across three datasets. The section also describes comparison models and the implementation choices used in the experiments.

1) Rating Prediction:

  • PMF learns user/item latent factors; NARRE and DAML model documents for representations, while RGCL uses graph contrastive learning and outperforms several state-of-the-art models.NRT, CAML, PETER [41], and SERMON focus on rating prediction and explanation generation; ReXPlug generates explanation text while predicting ratings. PEPLER uses prompts to produce explainable recommendations.Source: 1) Rating Prediction · p. 5, 2) Explanation Generation · p. 5

D. Experimental details and Reproducibility

  • The implementation uses 384-dimensional Sentence Transformer embeddings, six Transformer layers, eight attention heads, and a 1024-dimensional hidden size.The feed-forward dimensions are [256 →128 →64]; Adam uses a 0.001 learning rate and batch size 64, with early stopping after three epochs without improvement.Source: D. Experimental details and… · p. 5

A. Rating Prediction Accuracy

  • Against SERMON, CLARER reduces Yelp RMSE by 29% and MAE by 35%, TripAdvisor RMSE by 14% and MAE by 16%, and Amazon RMSE by 19% and MAE by 18%.The authors report that CLARER achieves the best RMSE and MAE across all three datasets, compared with rating-only and joint rating-and-explanation models.Source: A. Rating Prediction Accuracy · p. 5
  • Review-based NARRE, DAML, and RGCL have lower prediction errors than rating-based PMF, while recent explainable models outperform both baseline types.The comparison associates the review-based models with textual user and item representations, and PMF with latent factors.Source: A. Rating Prediction Accuracy · pp. 5–6, 1) Rating Prediction · p. 5

B. Explanation Generation

CLARER performs strongly on explanation generation across three datasets. Ablations compare overall-rating and review features, criteria-level ratings, and aspect-based representations.

B. Explanation Generation

  • CLARER achieves the best ROUGE-2 scores on all three datasets, the best ROUGE-L on Yelp, and the second-best ROUGE-L on Amazon.It outperforms all baselines on every reported metric in Yelp, and has the highest BLEU-1 on Amazon and BLEU4 on TripAdvisor.Source: B. Explanation Generation · p. 6

C. Ablation Studies

  • The CLARER ablation progresses from overall-rating and review features, to criteria-level ratings, then to aspect-based features with contrastive learning.Criteria-level ratings are unavailable for Amazon and Yelp, so those datasets are evaluated with overall-rating and aspect-based representations.Source: C. Ablation Studies · p. 6
  • Across the three datasets, aspect-based CLARER has the lowest MAE; criteria-level ratings also lower MAE on TripAdvisor but remain less effective than aspects.The reported CLARER MAE values are 0.49 for TripAdvisor, 0.70 for Amazon, and 0.51 for Yelp.Source: C. Ablation Studies · p. 7, Table V · p. 7
  • CLARER outperforms the two variants on most explanation-generation measures, with slightly worse BLEU1 on Yelp.The reported ROUGE-L comparison also favors CLARER over the other variants; the highest score is on Yelp, while the lowest is from TripAdvisor CLARER-O.Source: C. Ablation Studies · p. 7

D. Discussion

CLARER combines user- and item-ID features with aspect-based review representations, and its contrastive learning framework is reported to improve representation robustness, reduce noise, and enhance generalization. The discussion notes that explanation-generation performance varies across datasets, while the conclusion highlights meaningful justifications and trust and transparency as goals.

D. Discussion

  • Aspect-centric alignment is described as critical to improving recommendation quality in multi-criteria settings.The discussion attributes more accurate and personalized predictions to contrastive learning for aspect-based representation.Source: D. Discussion · p. 7
  • Across three datasets, CLARER does not consistently outperform all baselines in explanation generation, although it achieves the best results across all evaluation metrics on Yelp.Source: D. Discussion · p. 7
  • TripAdvisor’s longer, more varied reviews are proposed as a reason for lower BLEU-1, while Amazon’s product diversity and category-specific review differences are linked tentatively to lower BLEU-4 and ROUGE-L.The authors suggest exploring domain- or dataset-tailored aspects and a potentially larger transformer-based encoder for aspect representation.Source: D. Discussion · p. 7

VI. CONCLUSION

  • CLARER outperforms existing state-of-the-art recommendation models across multiple benchmark datasets, with contrastive learning reported to reduce noise and enhance generalization.The conclusion also describes its representations as more robust.Source: §VI · p. 7
  • The conclusion presents meaningful, interpretable recommendation justifications as a way to enhance system trust and transparency, and proposes multimodal data as a possible future direction.Source: §VI · p. 7
Loading 2610.07761v1…