Source-linked AI summary
A Capsule Network for Recommendation and Explaining What You Like and Dislike
Chenliang Li, Cong Quan, Li Peng, Yunwei Qi, Yuming Deng, Libing Wu
TL;DR
Review-based recommendation uses review semantics, but existing approaches leave unclear whether a user likes or dislikes an item aspect given the user's viewpoint and to what extent. CARP pairs viewpoints with aspects and uses sentiment capsules to identify informative logic units and estimate ratings. Across seven real-world datasets, the authors report gains in rating accuracy, robustness, and fine-grained explanation.
Problem
Review-based models highlight important words or aspects, but do not explain whether a user likes or dislikes an item aspect given the user's viewpoint and to what extent.
Source: §1 · pp. 1–2Method
CARP pairs user viewpoints with item aspects as logic units and uses sentiment capsules with Routing by Bi-Agreement to identify informative units and estimate ratings.
Source: §1 · p. 2Results
Across seven diverse real-world datasets, CARP is reported to improve rating prediction accuracy, robustness, and fine-grained explanation.
Source: §1 · p. 2Takeaways & Limitations
The authors conclude that modeling viewpoints, aspects, and sentiments at a fine-grained level yields better recommendation performance and understanding.
Source: §5 · p. 9
Abstract
from arXiv · showhide
User reviews contain rich semantics towards the preference of users to features of items. Recently, many deep learning based solutions have been proposed by exploiting reviews for recommendation. The attention mechanism is mainly adopted in these works to identify words or aspects that are important for rating prediction. However, it is still hard to understand whether a user likes or dislikes an aspect of an item according to what viewpoint the user holds and to what extent, without examining the review details. Here, we consider a pair of a viewpoint held by a user and an aspect of an item as a logic unit. Reasoning a rating behavior by discovering the informative logic units from the reviews and resolving their corresponding sentiments could enable a better rating prediction with explanation. To this end, in this paper, we propose a capsule network based model for rating prediction with user reviews, named CARP. For each user-item pair, CARP is devised to extract the informative logic units from the reviews and infer their corresponding sentiments. The model firstly extracts the viewpoints and aspects from the user and item review documents respectively. Then we derive the representation of each logic unit based on its constituent viewpoint and aspect. A sentiment capsule architecture with a novel Routing by Bi-Agreement mechanism is proposed to identify the informative logic unit and the sentiment based representations in user-item level for rating prediction. Extensive experiments are conducted over seven real-world datasets with diverse characteristics. Our results demonstrate that the proposed CARP obtains substantial performance gain over recently proposed state-of-the-art models in terms of prediction accuracy. Further analysis shows that our model can successfully discover the interpretable reasons at a finer level of granularity.
1 INTRODUCTION
Review-based recommendation can improve rating prediction, yet existing models leave unclear how a user's viewpoint and an item's aspect combine with sentiment to shape a rating. CARP models those components as logic units and uses sentiment capsules to identify informative units and estimate ratings; the authors report gains across seven datasets.
1 INTRODUCTION
- Prior review models highlight important words or aspects, but do not explain how a user's viewpoint shapes liking or disliking an item aspect.The desired explanation includes whether the user likes or dislikes an aspect and to what extent.Source: §1 · pp. 1–2
- CARP pairs a user viewpoint with an item aspect as a logic unit, then uses positive and negative sentiment capsules to identify informative units and estimate their effects.Routing by Bi-Agreement jointly identifies logic units associated with positive and negative sentiments; capsule vector lengths indicate sentiment probabilities.Source: §1 · p. 2
- CARP extracts viewpoints and aspects from user and item documents, respectively, and derives representations for their possible pairings.The paper calls semantically plausible viewpoint-aspect pairs informative logic units.Source: §1 · p. 2
- A multi-task sentiment process uses rating scores to supply implicit sentiment information, without human labeling or external NLP tools.Reviews are classified for sentiment according to their rating scores during training.Source: §1 · p. 2
- Across seven diverse real-world datasets, CARP is reported to improve rating accuracy, robustness, and fine-grained explanation.The introduction reports superiority in rating prediction accuracy, performance robustness, and fine-grained explanation.Source: §1 · p. 2
2 RELATED WORK
Related work spans review-based recommendation, aspect-based explanation, and capsule architectures for language tasks. The paper identifies a remaining gap: prior methods do not resolve whether a user likes or dislikes an item aspect given the user's viewpoint and to what extent.
2.1 Review-based Recommender Systems
- Review-based recommenders progressed from topic and factor models to contextual neural models using CNNs, RNNs, and attention [3] [27] [39].The review notes that bag-of-words approaches miss contextual information, while some methods select useful reviews or words for prediction.Source: §2.1 · pp. 2–3
2.2 Aspect-based Recommender Systems
- Aspect recommenders use external NLP tools [31], graph structures, or internal aspect inference [6], but leave viewpoint-conditioned sentiment unresolved.External-tool approaches depend on toolkit performance; the paper says these methods cannot explain whether a user likes or dislikes an aspect given a viewpoint, or to what extent.Source: §2.2 · p. 3
2.3 Capsule Networks for NLP
- Capsule networks aggregate low-level features into higher-level features through dynamic routing [22]; CARP's RBiA uses inter- and intra-capsule agreements for joint objectives.The paper presents Routing by Bi-Agreement as a complement to existing capsule-network studies.Source: §2.3 · p. 3
3 THE PROPOSED MODEL
CARP extracts user viewpoints and item aspects from reviews, pairs them into logic units, and uses sentiment capsules with Routing by Bi-Agreement to predict ratings and represent likes and dislikes.
3.1 Viewpoint and Aspect Extraction
- Viewpoint extraction applies convolutional context encoding and viewpoint-specific gating, then uses intra-attention to form user viewpoints; the same procedure extracts item aspects.The item extraction procedure uses a separate set of parameters.Source: §3.1 · pp. 3–4
3.2 Sentiment Capsule
- Each logic unit pairs a user viewpoint with an item aspect, and informative units are those with a semantically plausible association between them.CARP forms M² candidate pairings for each user-item pair.Source: §3.1 · p. 4, §3.2 · p. 4
- Positive and negative sentiment capsules jointly select logic units and produce outputs whose direction represents sentiment and whose length represents its probability.The capsule outputs encode the extent to which a user likes or dislikes an item.Source: §3.2 · p. 4
- Routing by Bi-Agreement combines inter-capsule and intra-capsule agreement when assigning logic-unit coupling coefficients.It updates the coefficients using agreement between each feature vector and the sentiment-capsule output at the current iteration.Source: §3.2 · p. 5
3.3 Rating Prediction and Optimization
- The predicted rating combines sentiment-specific scores weighted by capsule-output lengths with user and item biases, then maps the result into the rating range.A one-layer highway network adds flexibility to model scenarios where positive and negative causes coexist.Source: §3.3 · pp. 5–6
- Training combines rating-score MSE with a sentiment task whose labels assign positive sentiment above a rating threshold and negative sentiment otherwise.The sentiment objective is intended to improve sentiment analysis beyond matching observed ratings alone.Source: §3.3 · p. 6
- Because low-rating pairs are less common, CARP adds mutual-exclusion supervision for the opposite sentiment, while the authors note its constraint is too strict to hold in reality.The paper sets ε = 0.8 and describes the constraint as providing additional weak supervision for both sentiments.Source: §3.3 · p. 6
4 EXPERIMENTS
Experiments on seven datasets compare CARP with rating-only and review-based methods, analyze its parameter choices, and examine whether its logic units support fine-grained explanations. CARP achieves the best rating-prediction performance across the seven datasets, and the examples illustrate explanations combining viewpoints, item aspects, and sentiment.
4.1 Experimental Setup
- Experiments use seven datasets from Amazon, Yelp, and RateBeer, with 80:20 train/test splits and 10% of training records reserved for validation.The datasets include five Amazon domains, Yelp16-17, and Beer; empty-review records are filtered out.Source: §4.1 · p. 6
4.2 Performance Evaluation
- CARP achieves the best MSE across seven datasets, improving over TARMF, MPCN, and ANR by up to 39.0%, 9.7%, and 8.0%, respectively.Table 2 reports overall performance in MSE; the percentages are relative improvements over the named review-based alternatives.Source: Table 2 · p. 7, §4.2 · p. 8
- PMF is worst on all seven datasets, while no review-based baseline dominates; ANR leads three datasets and other baselines lead the remaining four.DeepCoNN performs worse than other solutions across datasets, while RBLT is relatively competitive, especially on sparser datasets.Source: §4.2 · pp. 7–8
4.3 Analysis of CARP
- Three or four routing iterations are optimal; RBiA yields sharper positive-capsule coupling ratios than RA, while M=5/7 is preferred in more datasets.The authors select three routing iterations and M=5 for experiments because M performance varies little across datasets.Source: §4.3 · p. 8, Figure 3 · p. 8, Table 4 · p. 8
- Across three datasets, performance is consistently optimal around λ=0.5; removing mutual exclusion worsens results to a level comparable with λ=1.The authors fix λ at 0.5, while noting it may not be optimal for some datasets.Source: §4.3 · p. 8, Figure 3 · p. 8
4.4 Explainability Analysis
- Examples illustrate mixed causes for a low-rated file box and a high-rated Post-it item, including weak positive and negative effects.For the Post-it pair, the negative capsule output length of 0.293 indicates a weak negative effect.Source: §4.4 · p. 9
- In a Musical Instruments example, CARP captures agreement on space saving and cost performance while downweighting an appearance mismatch; the authors report finer-grained explanations overall.The appearance mismatch has a lower relative negative value of 0.753 and a vector length of 0.295.Source: §4.4 · p. 9
5 CONCLUSION
Experiments indicate that fine-grained modeling of causes and effects yields better recommendation performance and understanding. The authors describe CARP as a step toward reasoning about human activity and identify several possible future applications.
5 CONCLUSION
- Experiments demonstrate better recommendation performance and understanding through fine-grained modeling of causes and effects.Source: §5 · p. 9
- The authors frame the capsule network model as a step toward reasoning about a particular kind of human activity.Source: §5 · p. 9
- The authors identify cross-domain recommendation, user profiling, and user-oriented review summarization as future possibilities.Source: §5 · p. 9