Source-linked AI summary
Generate Natural Language Explanations for Recommendation
Hanxiong Chen, Xu Chen, Shaoyun Shi, Yongfeng Zhang
TL;DR
Recommendation explanations matter, but templates restrict expression and reviews include sentences that do not explain products. The paper proposes HSS with feature-aware attention and auto-denoising for personalized explanation generation. It reports stronger rating prediction than BiasedMF and SVD++, and a comparable result to DeepCoNN.
Problem
Templates limit explanation expression and require manual effort, while review sentences may be noisy or unrelated to explanation.
Source: §1 · p. 2Method
HSS generates personalized explanations using hierarchical generation, feature-aware attention, and auto-denoising based on review sentences.
Source: §1 · p. 2Results
The model outperforms CF-based BiasedMF and SVD++ for rating prediction and achieves a comparable result to DeepCoNN.
Source: §6 · p. 9Takeaways & Limitations
The framework generates explanation sentences alongside rating predictions, with the stated aim of improving recommender effectiveness and trustworthiness.
Source: §6 · p. 9The authors identify wrong item descriptions as a recurring issue and suggest sparse data may leave item vectors insufficiently trained.
Source: §5.4 · p. 9
Abstract
from arXiv · showhide
Providing personalized explanations for recommendations can help users to understand the underlying insight of the recommendation results, which is helpful to the effectiveness, transparency, persuasiveness and trustworthiness of recommender systems. Current explainable recommendation models mostly generate textual explanations based on pre-defined sentence templates. However, the expressiveness power of template-based explanation sentences are limited to the pre-defined expressions, and manually defining the expressions require significant human efforts. Motivated by this problem, we propose to generate free-text natural language explanations for personalized recommendation. In particular, we propose a hierarchical sequence-to-sequence model (HSS) for personalized explanation generation. Different from conventional sentence generation in NLP research, a great challenge of explanation generation in e-commerce recommendation is that not all sentences in user reviews are of explanation purpose. To solve the problem, we further propose an auto-denoising mechanism based on topical item feature words for sentence generation. Experiments on various e-commerce product domains show that our approach can not only improve the recommendation accuracy, but also the explanation quality in terms of the offline measures and feature words coverage. This research is one of the initial steps to grant intelligent agents with the ability to explain itself based on natural language sentences.
1 INTRODUCTION
Recommendation explanations can support effectiveness, transparency, persuasiveness, and trust, but existing textual methods trade expressive freedom against human effort or reliance on existing sentences. The paper proposes HSS with auto-denoising and feature-aware attention for personalized explanation generation.
1 INTRODUCTION
- Explanations are presented as potentially supporting recommendation effectiveness, transparency, persuasiveness, and trustworthiness.Source: §1 · p. 1
- Template methods require extensive human effort and restrict expression, while retrieval methods cannot generate sentences beyond those already present in reviews.Templates fill predefined forms with words; retrieval methods select existing review sentences.Source: §1 · p. 1
- User reviews contain noise, users attend to different product features, and readability metrics alone do not establish explanation quality.The paper identifies data bias, personalization, and evaluation as three challenges.Source: §1 · pp. 1–2
- HSS collaboratively learns across multiple sentences from different users to generate explanation sentences.The proposed framework combines hierarchical sequence-to-sequence modeling with auto-denoising.Source: §1 · p. 2
- Item feature words guide sentence selection, while feature attention enhances the feature-level personalization of generated explanations.The model uses item feature words extracted from reviews to select explanation sentences for learning.Source: §1 · p. 2
- BLEU, ROUGE, and feature words coverage evaluate generated explanations, with feature coverage intended to reflect personalized preferences and possible explainability.BLEU is precision-based, while ROUGE is relatively recall-based.Source: §1 · p. 2
2 RELATED WORK
Related work spans explainable recommendation approaches using collaborative filtering, reviews, templates, and neural language generation. The paper positions its auto-denoising generation approach against limitations in templates and noisy review sentences.
2 RELATED WORK
- Early user- and item-based collaborative filtering used similar users or items, while later latent-factor approaches improved accuracy but made recommendations harder to explain [15, 26, 34, 36] [44] [49].Source: §2 · p. 2
- Review-based research incorporated user text to improve recommendation quality [49] and explainability [18] [39].Source: §2 · p. 2
- Template-based EFM fills templates with user-interested item features, but templates constrain expression and require substantial manual design.Neural review generation and fake-review approaches also face the problem that review sentences may not explain the recommendation [18] [9].Source: §2 · p. 2
- Neural language models use recurrent networks for sequential modeling, and the paper chooses word-level GRU generation over character-level modeling.Character-level models can address OOV issues but are generally outperformed by word-level models.Source: §2 · pp. 2–3
3 THE FRAMEWORK
HSS combines rating regression with hierarchical natural-language explanation generation using shared user and item latent factors. It weights review sentences by feature-word content, attends to item features, and trains both modules jointly.
3.1 Overview
- The framework jointly learns rating prediction and explanation generation from shared user and item latent factors.At test time, it uses those latent factors and extracted feature words.Source: §3.1 · p. 3
3.2 Neural Rating Regression
- The rating-regression module maps user and item latent factors through an MLP to predict ratings and optimize prediction loss.Additional tanh layers provide nonlinear transformations.Source: §3.1 · p. 3, §3.2 · p. 4
3.3.1 Auto-denoising.
- Sentence losses are weighted by feature-word proportion, weakening the training effect of sentences with few or no feature words.The model extracts product feature words from reviews to calculate the sentence-level supervised factor.Source: §3.3.1 · p. 4
3.3.2 Feature-aware Atention.
- Feature-aware attention scores product-feature words against hidden states, while user and item embeddings enrich word inputs for personalized generation.The normalized attention weights form a feature-aware context used to initialize sentence generation.Source: §3.3.2 · pp. 4–5
3.3 Personalized Natural Language Explanation Generation
- A hierarchical GRU generates multiple explanation sentences, with a context GRU supplying initial states to the word-generating GRU.The context representation is updated after each generated sentence.Source: §3.3.2 · p. 4
3.4 Multi-task Learning
- Multi-task learning combines rating-regression loss with weighted sentence-generation losses to train the framework’s neural parameters and embeddings.Each sentence-generation loss is scaled by its auto-denoising supervised factor.Source: §3.4 · p. 6
4.1 Datasets
The experiments use filtered Amazon review data from Electronics and Beauty, retaining users with at least ten shopping records and moving infrequent-item records into training to reduce cold-start cases.
4.1 Datasets
- The data originate from Amazon user reviews and metadata spanning May 1996 through July 2014, with duplicated records removed.The source collection covers 24 product categories; Electronics and Beauty are selected for the experiments.Source: §4.1 · p. 6
- Electronics contains 45,224 users, 61,687 items, 744,453 reviews, and 434 extracted feature words; Beauty contains 5,122 users, 11,616 items, 90,247 reviews, and 518 feature words.The two datasets were chosen to represent different domains and scales.Source: §4.1 · p. 6
- The authors filter Amazon 5-core data to users with at least ten shopping records and move less-frequent-item records into training to avoid cold-start cases in testing.They retain punctuation and numbers and impose no sentence-length threshold, leaving noisy text.Source: §4.1 · p. 6
4.2 Rating Regression Evaluation
Rating prediction is evaluated against matrix-factorization and review-based baselines using RMSE. The supplied passages specify the comparison methods and metric, but provide no outcome values.
4.2.1 Baselines.
- The rating-prediction baselines are BiasedMF, SVD++, and DeepCoNN, with the first two using ratings and DeepCoNN incorporating user reviews.The authors identify DeepCoNN as a state-of-the-art deep-learning method and use an existing implementation.Source: §4.2.1 · p. 6
- BiasedMF adds user and item biases to matrix factorization, while SVD++ integrates implicit feedback into latent-factor modeling.The passage cites BiasedMF and SVD++ [14].Source: §4.2.1 · p. 6
4.2.2 Evaluation Metric.
- Rating-prediction performance is measured with Root Mean Square Error (RMSE).Source: §4.2.2 · p. 6
4.3 Explanation Sentence Generation Evaluation
The evaluation compares HSS with Att2SeqA and assesses generated explanations using BLEU, ROUGE, and feature words coverage. Att2SeqA is selected for similarities in inputs, attention, and review data, despite differences in their generation goals.
4.3.1 Baseline.
- Att2SeqA is chosen as the comparison model because both systems use user and item factors, attention, and user reviews, despite different text-generation goals.Att2SeqA generates reviews using user, item, and rating inputs; HSS predicts ratings and generates explanation sentences.Source: §4.3.1 · p. 7
4.3.2 Evaluation Metrics.
- BLEU measures the precision of candidate-text n-grams overlapping with human references, with each word count clipped to the largest count in any single reference.The evaluation describes BLEU as precision-based and uses clipped counts for its n-gram precision calculation.Source: §4.3.2 · p. 7
- ROUGE measures how much human-reference text appears in generated text; the experiments use recall, precision, and F-measure for ROUGE-1, ROUGE-2, ROUGE-L, and ROUGE-SU4.The variants represent unigram, bigram, longest-common-subsequence, and skip-gram measures, respectively.Source: §4.3.2 · p. 7
- Feature words coverage measures how well generated explanation sentences capture users’ personalized preferences and is used to evaluate their explainability.It is based on reference feature words and the number covered in machine-generated sentences.Source: §4.3.2 · p. 7
4.4 Experimental Settings
The experimental settings specify model and decoding configurations for HSS and Att2SeqA. The generated sequences are limited to their first two sentences, which the authors consider a reasonable length for explanations.
4.4 Experimental Settings
- HSS uses 300-dimensional user and item factors, 300-dimensional hidden and word vectors, and a four-layer rating regressor with a three-layer explanation generator.Training uses batch size 100, gradient clipping at 1.0, L2 weight 0.001, dropout 0.1, beam size 4, learning rate 0.002, and SGD with momentum 0.9.Source: §4.4 · p. 7
- Both HSS and Att2SeqA generate sequences up to 100 tokens, but only the first two sentences are retained for evaluation.The remainder is discarded after searching for the end-of-sentence tag.Source: §4.4 · p. 7
- Table 3 reports RMSE values for rating regression.The supplied caption identifies the table’s scope but does not provide cell values.Source: Table 3 · p. 8
- Two sentences are considered a reasonable explanation length: shorter explanations are quicker to grasp, while one sentence may omit information needed to improve recommendation quality.The authors use two sentences as a compromise between quick comprehension and information coverage.Source: §4.4 · p. 8
5 RESULTS AND DISCUSSIONS
HSS performs well on rating prediction and generates explanations that generally score better than the baseline on ROUGE measures. Its feature-aware attention supports feature coverage, while multi-sentence generation trades slightly lower precision for better recall; sparse data can also yield incorrect item descriptions.
5.1 Ratings
- HSS outperforms all baselines on Beauty ratings and beats BiasedMF and SVD++ on Electronics RMSE, while remaining comparable to DeepCoNN on Electronics.Both HSS and DeepCoNN use user reviews, unlike traditional collaborative-filtering methods that use only ratings.Source: §5.1 · p. 8
5.2 Personalized Explanation Sentence Generation Qality
- HSS almost always beats the baseline across ROUGE measures, with only slightly lower ROUGE-SU4 recall; it also achieves higher precision and BLEU than At2SeqA.Feature-word attention is designed to improve feature coverage by weighting features according to the user-item context.Source: §5.2 · p. 8, Table 4 · p. 8
5.3 Multi-sentence Generation Performance
- Generating more than one sentence improves ROUGE recall on all measures, while ROUGE precision is slightly lower than in the one-sentence case.The experiments varied the number of sentences from one to three on Beauty.Source: §5.3 · p. 8
5.4 Case study
- The hierarchical GRU helps generate multiple sentences by using a context GRU to capture long dependencies and shorten each sentence-generation sequence.The review sentences are not length-aligned, and their number and lengths vary.Source: §5.4 · pp. 8–9
- Sparse data can leave an item vector undertrained, resulting in an explanation that incorrectly describes the item.The case study includes fluent explanations, feature-containing but less fluent sentences, and incorrect item descriptions.Source: §5.4 · p. 9, Table 7 · p. 9
6 CONCLUSIONS AND FUTURE WORK
The HSS framework combines rating prediction with personalized explanation generation. The authors report rating results against three baselines and describe future work on wrong descriptions, robustness, and knowledge-enhanced generation.
6 CONCLUSIONS AND FUTURE WORK
- HSS outperforms BiasedMF and SVD++ on rating prediction and achieves a result comparable to DeepCoNN.The authors describe DeepCoNN as state of the art.Source: §6 · p. 9
- The explanation module uses a hierarchical GRU with feature-aware attention to generate personalized explanation sentences.Source: §6 · p. 9
- Auto-denoising is designed to reduce the effect of unrelated sentences during training.Source: §6 · p. 9
- Future work includes addressing wrong descriptions, testing robustness on other datasets, and exploring knowledge-enhanced explanation generation.Source: §6 · p. 9