Source-linked AI summary
Explainable Recommendation via Multi-Task Learning in Opinionated Text Data
Nan Wang, Hongning Wang, Yiling Jia, Yue Yin
TL;DR
Recommendation systems often lack transparency, making it difficult for users to judge results before acting. The paper addresses this by jointly learning user preferences and opinionated content through tensor factorization, and reports effectiveness in recommendation and explanation tasks on Amazon and Yelp reviews, with positive user-study feedback.
Problem
Recommendation algorithms commonly optimize ratings or rankings without explaining the feature-level opinions behind users’ overall assessments, limiting transparency for decision-making.
Method
A joint tensor factorization integrates user preference modeling for recommendation with opinionated content modeling for feature- and phrase-level explanation.
Results
Experiments on Amazon and Yelp reviews show effectiveness in both recommendation and explanation generation, while user studies report positive feedback on practical value.
Takeaways & Limitations
The approach produces explainable recommendations that combine recommended items with detailed features and opinion-level descriptions of users’ preferences.
Takeaways & Limitations
The current model implicitly captures entity dependencies from rating and review histories and does not yet use external resources, advanced text synthesis, or real-world deployment evaluation.
Abstract
from arXiv · showhide
Explaining automatically generated recommendations allows users to make more informed and accurate decisions about which results to utilize, and therefore improves their satisfaction. In this work, we develop a multi-task learning solution for explainable recommendation. Two companion learning tasks of user preference modeling for recommendation} and \textit{opinionated content modeling for explanation are integrated via a joint tensor factorization. As a result, the algorithm predicts not only a user's preference over a list of items, i.e., recommendation, but also how the user would appreciate a particular item at the feature level, i.e., opinionated textual explanation. Extensive experiments on two large collections of Amazon and Yelp reviews confirmed the effectiveness of our solution in both recommendation and explanation tasks, compared with several existing recommendation algorithms. And our extensive user study clearly demonstrates the practical value of the explainable recommendations generated by our algorithm.
1 INTRODUCTION
The paper argues that explainable recommendation should address both users’ overall preferences and the feature-level opinions expressed in their reviews. It proposes jointly modeling recommendation and opinionated explanations, then evaluates the approach on Amazon and Yelp data.
- Motivation: Users need explanations to assess recommendations before acting and to make more informed decisions about which results to use.The paper frames explanations as a way to build trust while emphasizing decision quality rather than mere promotion.
- Motivation: Overall ratings summarize multiple factors, so recommendation methods based only on ratings or rankings overlook users’ feature-level decisions.The paper motivates companion learning tasks that model different aspects of users’ decisions over recommended items.
- Motivation: Opinionated review content can explain an overall rating by identifying positively and negatively assessed item features.For an example four-star rating, three positively commented features and one negatively commented feature contribute to the assessment.
- Prior approaches: Topic-modeling approaches may miss detailed feature opinions, causing biased recommendations or incorrect explanations when criticism appears in an otherwise positive review.Phrase-level sentiment analysis addresses finer-grained opinions, but the cited passage ends before describing its limitation.
- Approach: The proposed joint tensor factorization integrates user preference modeling with opinionated content modeling across users, items, features, and opinionated phrases.Three tensors represent user-item-feature preferences, user-feature-phrase descriptions, and item-feature-phrase comments; joint factorization exploits their relatedness.
- Evaluation: Experiments on Amazon and Yelp reviews report effectiveness for both item recommendation and explanation generation, while user studies provide positive practical feedback.The evaluation compares the solution with existing explainable recommendation algorithms and includes user studies of its utility.
2 RELATED WORK
Related work combines latent-factor recommendation with explanation techniques, but factorization-based models remain difficult to interpret. Existing approaches use interfaces, sentiment analysis, topic models, social relations, or explainability constraints to expose different aspects of recommendations.
- Explanation methods: Explanation research has evaluated interfaces and generation strategies to help users assess recommender systems.Examples include rating histograms from similar users and keyword-, neighbor-, and influence-style explanations.
- Latent factor models: Latent factor models map users and recommendation candidates into lower-dimensional spaces, but their latent nonlinear representations are difficult to explain.This interpretability problem persists despite the practical recommendation quality of matrix and tensor factorization.
- Factorization-based explanations: Prior factorization methods combine phrase-level sentiment analysis or explainability constraints, but some explain only overall ratings.These methods connect latent factors with sentiment or positively rated items without necessarily modeling detailed opinions.
- Probabilistic and topic models: Probabilistic latent-factor approaches combine recommendation with topic modeling to match item topics to users and support topic-based explanations.Subsequent work adds aspect-level sentiment and social relations to enable finer-grained or social explanations.
3 METHODOLOGY
MTER jointly models item recommendation and opinionated explanations by factorizing tensors that represent user–item–feature preferences and review content. The shared factorization uses feature-level sentiment, overall ratings, ranking constraints, and opinion phrases to produce recommendations with feature-level and textual explanations.
- 3 METHODOLOGY: MTER integrates user preference modeling and opinionated content modeling through a joint tensor factorization.The two tasks share entities and latent representations while corresponding core tensors capture task-specific interactions.
- 3 METHODOLOGY: The recommendation tensor represents how users appreciate individual features of items, while an overall-rating dummy feature connects feature-level and item-level assessments.Feature scores are derived from sentiment analysis and mapped to the target domain’s overall-rating range [1, N].
- 3 METHODOLOGY: A domain-specific sentiment lexicon converts reviews into feature, opinion, and sentiment tuples whose polarities are summed to obtain feature-level sentiment.Repeated mentions of a feature can contribute multiple sentiment polarities to its aggregate score.
- 3 METHODOLOGY: Tucker factorization completes missing user–item–feature affinities using nonnegative latent factors and a core tensor that models interactions among tensor modes.Different entity types may use different latent dimensions, and nonnegative constraints avoid negative predictions after scores are mapped to [1, N].
- 3 METHODOLOGY: The reconstructed feature-level assessments provide rating-based explanations, while the opinion-content task predicts detailed user comments for recommended item features.The joint model uses shared dependencies among recommendation and explanation tensors to generate opinionated textual explanations.
- 3 METHODOLOGY: Bayesian Personalized Ranking adds pairwise ranking optimization to reconstruction-error minimization, maintaining relative item orders while learning feature factors.The ranking loss is minimized when pairwise orders are maintained and score differences are increased.
4 EXPERIMENTATION
Experiments on sparse Amazon and Yelp review data evaluate MTER for personalized recommendation and opinionated content prediction. MTER improves recommendation quality over baselines and predicts user-specific features and opinion phrases for explanations, while its ranking performance depends on balancing reconstruction and BPR objectives.
- Experiment Setup: MTER is evaluated on Amazon cellphone/accessory reviews and Yelp restaurant reviews, both characterized by substantial sparsity.The Amazon data contain 401 extracted features, while Yelp contains 1,065; 73% of Amazon users and 47% of products have one review, and 54% of Yelp users have one review.
- Experiment Setup: The benchmark compares MTER with popularity ranking, NMF, BPRMF, JMARS, EFM, and a canonical-decomposition variant.The comparison includes both recommendation methods and explainable recommendation baselines, plus MTER-S(SA).
- Personalized Item Recommendation: MTER outperforms the baselines for personalized item recommendation, whereas NMF performs worse than the non-personalized popularity solution.BPR improves ranking in BPRMF, but BPRMF generally trails MTER and its variants because it does not exploit opinionated review content.
- Contribution of BPR: Increasing BPR weight improves ranking until excessive emphasis undermines content modeling and raises overfitting risk.At small relative BPR weight, reconstruction error dominates; as pairwise constraints dominate factor learning, content-modeling quality declines.
- Opinionated Textual Explanation: MTER ranks both user-attended features and detailed opinion phrases, and its latent factors capture complex user-item-feature associations beyond frequency.A case study places a battery-case product close to a user whose review history emphasizes cases and batteries, illustrating the learned relatedness.
5 USER STUDY
The user study evaluates whether MTER’s recommendations and explanations improve satisfaction, information, transparency, and understanding across Amazon and Yelp reviews. In an A/B test with 900 questionnaires, MTER generally outperformed the baselines, while participants’ feedback favored detailed opinionated explanations.
- Study design: Participants inferred an assigned user’s preferences from previous reviews before judging recommendations and explanations from that user’s perspective.The study used existing users from Amazon and Yelp review datasets and required participants to answer survey questions about generated results.
- Study design: The study evaluates recommendation satisfaction, informativeness, transparency, and the effectiveness of opinionated textual explanations.Questions Q1, Q2, and Q4 target satisfaction, effectiveness, and transparency; Q3 and Q5 assess opinionated explanations.
- Study design: 900 questionnaires were collected, with 150 responses for each of three models on each dataset.Participants were divided into six groups by model and dataset after invalid responses were filtered.
- Results & Analysis: Except for Q1 on Yelp, MTER outperformed both baselines across the user-study aspects, with support from paired t-tests.The comparison attributes gains on Q2 and Q4 to comprehensive explanations and gains on Q3 and Q5 to combining feature- and opinion-level explanations.
- Results & Analysis: Optional participant feedback described MTER’s bracketed descriptive words as clearer, thorough, and useful for understanding recommendations.The reported comments expressed interest in detailed explanations of automatically generated recommendations.
6 CONCLUSION
The paper concludes that joint tensor factorization provides a multi-task approach for improving recommendation and explanation together. Offline experiments and a user study support the model’s effectiveness, while the authors identify richer external resources, more natural text generation, and real-world deployment as future directions.
- Conclusion: The paper develops a multi-task explainable recommender using joint tensor factorization to improve recommendation and explanation jointly.The approach integrates user preference modeling with opinionated content modeling.
- Conclusion: Offline experiments and a user study show the model’s comprehensiveness and effectiveness.The conclusion summarizes evidence across both evaluation settings.
- Conclusion: The authors describe the work as a first step and propose external resources, neural language models, and real-world evaluation as future directions.Suggested resources include social networks and product taxonomies; deployment with real user populations remains to be evaluated.