Source-linked AI summary
Aspect-Aware Latent Factor Model: Rating Prediction with Ratings and Reviews
Zhiyong Cheng, Ying Ding, Lei Zhu, Mohan Kankanhalli
TL;DR
Matrix factorization performs well but is limited by cold-start, opaque predictions, and globally optimized factors that may fit individual user-item pairs poorly. The paper combines review-derived aspect modeling with an aspect-aware latent factor model, and reports significant improvements over strong review-and-rating baselines, especially for users with few ratings. Its aspect ratings and pair-specific weights also support deeper recommendation interpretation.
Problem
Matrix factorization faces cold-start, limited transparency, and suboptimal local prediction because users and items can differ in aspect preferences and characteristics.
Method
ATM models review-derived user preferences and item features by aspect, while ALFM links rating factors to those aspects and combines weighted aspect ratings.
Results
The model significantly outperforms state-of-the-art methods using reviews and ratings on 19 Yelp and Amazon datasets, including for users with few ratings.
Takeaways & Limitations
Aspect-specific ratings and user-item-dependent aspect weights provide interpretable recommendations while addressing sparse-rating and local-prediction settings.
Takeaways & Limitations
The model assumes all items belong to the same category and therefore share the same set of aspects.
Abstract
from arXiv · showhide
Although latent factor models (e.g., matrix factorization) achieve good accuracy in rating prediction, they suffer from several problems including cold-start, non-transparency, and suboptimal recommendation for local users or items. In this paper, we employ textual review information with ratings to tackle these limitations. Firstly, we apply a proposed aspect-aware topic model (ATM) on the review text to model user preferences and item features from different aspects, and estimate the aspect importance of a user towards an item. The aspect importance is then integrated into a novel aspect-aware latent factor model (ALFM), which learns user's and item's latent factors based on ratings. In particular, ALFM introduces a weighted matrix to associate those latent factors with the same set of aspects discovered by ATM, such that the latent factors could be used to estimate aspect ratings. Finally, the overall rating is computed via a linear combination of the aspect ratings, which are weighted by the corresponding aspect importance. To this end, our model could alleviate the data sparsity problem and gain good interpretability for recommendation. Besides, an aspect rating is weighted by an aspect importance, which is dependent on the targeted user's preferences and targeted item's features. Therefore, it is expected that the proposed method can model a user's preferences on an item more accurately for each user-item pair locally. Comprehensive experimental studies have been conducted on 19 datasets from Amazon and Yelp 2017 Challenge dataset. Results show that our method achieves significant improvement compared with strong baseline methods, especially for users with only few ratings. Moreover, our model could interpret the recommendation results in depth.
1 INTRODUCTION
The paper motivates combining reviews with ratings to address matrix factorization’s cold-start, transparency, and local-prediction limitations. It proposes aspect-aware topic and latent-factor models that produce interpretable, user-item-specific rating predictions and reports improvements across sparse-rating settings.
- Motivation: Matrix factorization suffers from cold-start, lacks explanations for overall ratings, and can compromise prediction quality for individual users or items.Its global optimization learns user and item factors across all training ratings, while different users and items may differ in aspect preferences and characteristics.
- Motivation: Reviews provide complementary information about user preferences, item features, and the reasons underlying ratings.
- Approach: ATM extracts latent topics from reviews to model users and items across aspects and compute aspect importance for each user-item pair.
- Approach: ALFM associates latent factors with ATM’s aspects, estimates aspect ratings, and combines them using user-item-specific aspect weights.The model uses a weight matrix to connect factors to aspects and relaxes prior one-to-one mappings between latent topics and latent factors.
- Results: Experiments on 19 Yelp and Amazon datasets show significant improvements over methods using both reviews and ratings, including for users with few ratings.The paper also demonstrates recommendation interpretability through examples.
2 RELATED WORK
The related work covers sentiment-based, topic-based, and deep learning-based approaches for rating prediction from reviews and ratings. It positions the paper’s method in the topic-based category and distinguishes it from approaches that require reviews at prediction time.
- Problem setting: Real-world recommendation often requires predicting ratings for uncommented and unrated items, where the target review is unavailable.
- Sentiment-based: Sentiment-based methods analyze review sentiment or polarity before applying collaborative filtering or joint factorization.These approaches depend on external natural-language-processing tools for sentiment analysis.
- Topic-based: Topic-based methods extract latent topics or aspects from reviews, shifting from manual domain labeling toward automatic topic modeling.Manual labeling requires domain knowledge and substantial labor.
- Deep learning-based: Deep learning-based recommendation applies neural collaborative filtering and review-based architectures such as DeepCoNN and TransNet.TransNet adds a layer that simulates the review for a target user-item pair before rating prediction.
3 PROPOSED MODEL
The proposed model combines review-derived aspects with rating-based latent factors to estimate aspect-aware, user-item-specific ratings. ATM models user preferences and item characteristics, while ALFM links latent factors to aspects and combines weighted aspect ratings into an overall prediction.
- Overall rating prediction: The overall rating is formed by summing aspect ratings weighted by aspect importance, with user, item, and global bias terms included in ALFM.Parameters are estimated by minimizing training-set rating prediction error with regularization; ℓ1 regularization encourages sparse aspect weights and ℓ2 regularization controls latent-factor magnitudes.
- Aspect rating estimation: The review-derived matching score su,i,a scales each aspect rating, so both review similarity and latent-factor interaction contribute to the prediction.The score is defined as 1 − JSD(θu,a,ψi,a), while the aspect rating is su,i,a · (wa ⊙pu)T(wa ⊙qi).
- Aspect rating estimation: ALFM introduces W to associate rating-based latent factors with aspects, making (wa ⊙pu)T(wa ⊙qi) an aspect rating.The weight vector wa is the a-th column of W, and element-wise products select the factors associated with aspect a.
- Aspect importance estimation: Aspect importance ρu,i,a combines a user-preference probability with user- and item-specific aspect probabilities.The model estimates ρu,i,a as πuλu,a + (1 −πu)λi,a, where ATM estimates the component parameters.
- Aspect-aware Topic Model: ATM learns latent aspects from reviews, representing each aspect through topic distributions for users’ preferences and items’ characteristics.User aspect distributions are learned from the user’s reviews, while item aspect distributions are learned from reviews written about the item.
4 EXPERIMENTAL STUDY
Experiments evaluate ALFM’s parameter sensitivity, rating-prediction performance, cold-start behavior, and interpretability across Amazon and Yelp datasets. ALFM generally outperforms comparison methods, including for users with few ratings, while producing aspect-based explanations.
- Research questions: The evaluation addresses parameter effects, model comparison, cold-start prediction, and interpretation of high or low ratings.These correspond to research questions RQ1–RQ4.
- Dataset and protocol: Experiments use Amazon product-review data across 18 categories and the Yelp Dataset Challenge 2017 dataset, with 5-core filtering for Amazon.The study reports evaluation on 19 datasets overall.
- Parameter analysis: Optimal numbers of latent topics and factors vary across datasets; more factors usually help, while the best topic count depends on the reviews.The results indicate that setting the numbers of factors and topics equal may not be optimal.
- Model comparison: ALFM achieves the best results on 18 of 19 datasets and improves average performance over BMF, CTR, RMR, RBLT, and TransNet.Reported relative improvements average 16.49% over BMF, 6.28% over CTR, 8.18% over RMR, 3.37% over RBLT, and 4.26% over TransNet.
- Cold-start setting: ALFM substantially improves prediction over BMF and other review-based baselines by modeling preferences and item features across aspects and estimating targeted aspect weights.RMSE is the evaluation metric, with lower values indicating better performance.
- Interpretability: The model discovers semantically interpretable aspects and explains ratings using aspect importance, preference matching, sentiment polarity, and aspect ratings.For user 1511, the illustrated aspects include value, comfort, accessories, shoes, and clothing.
5 CONCLUSIONS
The proposed model combines review-derived aspects with rating-based latent factors to estimate user-item-specific aspect ratings and weights. Across 19 datasets, it improves rating prediction, particularly for users with few ratings, while providing detailed recommendation interpretations.
- ALFM correlates review-derived latent topics and rating-derived latent factors through shared aspects without requiring one-to-one topic-factor mappings.This supports richer modeling of user preferences and item features than prior approaches constrained to one-to-one mappings.
- The model estimates aspect ratings and assigns pair-specific aspect weights based on a user’s preferences toward an item’s aspects.These weights combine with aspect ratings to model each user-item pair and support detailed recommendation explanations.
- The model greatly improves rating prediction accuracy over state-of-the-art methods, especially for users with few ratings.
- Extracted aspects, estimated aspect weights, and aspect ratings provide detailed interpretations of recommendation results.