When Reviews Are Scarce: Structure-Aware Review Imputation for Recommendation in Realistic Sparse Review Settings
Authors: Leyao Wang, Xutao Mao, Qinwen Ge, Xuhui Zhan, Yuying Zhao, Bo Ni, Ryan A. Rossi, Nesreen K. Ahmed, Tyler Derr
Organizations: Yale University New Haven, CT, USA · Vanderbilt University Nashville, TN, USA · Adobe Research San Jose, CA, USA · Cisco AI Research Santa Clara, CA, USA
Textual reviews enrich recommender systems with fine-grained preference signals and enhanced explainability. However, in real-world scenarios, users rarely leave reviews, resulting in severe sparsity that undermines the effectiveness of review-aware recommenders under partial observability. A natural remedy is to synthesize the missing reviews to restore the textual signal these recommenders rely on. Yet, conventional imputation techniques, such as matrix completion and LLM-based augmentation, either lose contextualized semantics by embedding texts into vectors or overlook structural dependencies on the user-item graph. To address these shortcomings, we propose TWISTER (ToWards Imputation on Sparsity with Textual Edge Graph Representation), a unified framework for sparse-text recommendation that synthesizes missing reviews by jointly modeling semantic and structural signals. Specifically, we represent user-item interactions as a Textual-Edge Graph (TEG), treating reviews as edge attributes. To capture relational context, we construct line-graph views and employ an LLM as a graph-aware aggregator. For each interaction lacking a textual review, our model aggregates natural-language representations from neighboring interactions to generate a coherent and personalized review. Experiments on multiple datasets show that TWISTER consistently outperforms traditional numeric, graph-based, and LLM baselines, delivering higher-quality imputed reviews and, more importantly, improved downstream recommendation. In summary, TWISTER is a drop-in, structure-aware imputer that recovers the textual signal that review-aware recommenders need when reviews are scarce.
Figures & tables
Figure 1. Recommendation performance averaged over Amazon_Video_Games and Goodreads_Children . We compare Non-Review , Blank-Fill , interaction-only LightGCN , rating-only GINE (GINE-Rating), TWISTER (TWISTER-Q), and an Oracle with ground-truth reviews for all ratings. TWISTER achieves the strongest performance on all three reported metrics, motivating structure-aware review imputation.
Figure 2. Pipeline for review imputation. (1) Construct a bipartite textual-edge graph from user–item interactions and reviews. (2) Build user-side and item-side line graphs to capture relational context. (3) Use an LLM to aggregate neighborhood information into textual representations. (4) Prompt the LLM with these textual representations, ratings, and metadata to generate missing reviews.
Figure 3. Recommendation performance under the Cold Start setting ( 50% of users have no reviews in the training set).
Figure 4. LLM-as-Judge scores (1–5) from DeepSeek-R1 on Authenticity and Helpfulness. Structure-aware variants ( TWISTER -L, TWISTER -Q) consistently outperform. Full four-dimension results are in Appendix C .
Figure 5. Heatmap illustrating the semantic similarity of imputed and original texts. With the exception of the over-smoothed KNN and Mean baselines, our method ( TWISTER -L and TWISTER -Q ) achieves notably higher semantic fidelity.
Figure 6. Normalized Dirichlet energies on weighted user-side line-graph views (lower is smoother). Excluding the over-smoothed KNN and Mean baselines, our methods ( TWISTER -L and TWISTER -Q) attain the lowest energy.
Figure 7. LLM-as-Judge heatmaps. DeepSeek-R1 scores (1–5) for synthetic reviews on Specificity and Readability across methods/datasets. Structure-aware LLM variants ( TWISTER -L , TWISTER -Q ) consistently outperform the baselines.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Datasets
# Reviews
# Item
# User
Videos
2,565,3629
71,982
1540,618
Music
1,512,530
11,2222
90,330
Toys
8,194,101
624,792
4,204,994
Comics
542,015
89,311
59,347
Children
734,640
123,946
92,667
Appendix
Table 3. Statistics of the raw data.
Dataset
Model
Type
Count
Total
Average
Tokens
Time (s)
Tokens
Time (s)
Children
Llama
User
70
11,889
59.4
169.8
0.849
Item
1,326
115,004
575.0
86.7
0.434
Qwen
User
70
9,273
46.4
132.5
0.662
Item
1,326
94,303
471.5
71.1
0.356
Comics
Llama
User
70
12,118
60.6
173.1
0.866
Appendix
Table 4. Aggregation Token Counts and Generation Time
Dataset
Model
# Review
Total Metrics
Average Metrics
Imputation
Imputation
Imputation
Imputation
Tokens
Time (s)
Tokens
Time (s)
Children
Llama
3,714
272,902
1,364.5
73.5
0.367
Qwen
3,714
260,770
1,043.1
70.2
0.281
Comics
Llama
2,310
327,866
1,639.3
141.9
0.710
Qwen
2,310
188,131
752.5
81.4
0.326
Appendix
Table 5. Imputation Generation Token Counts and Time
Figure 8. Normalized Dirichlet energies on user-, item-, and weighted user-side line-graph views (lower is smoother). Excluding the over-smoothed KNN and Mean baselines, our methods ( TWISTER -L and TWISTER -Q) attain the lowest energy.
Method
Amazon_Video
Amazon_Music
Amazon_Toys
Goodreads_Comics
Goodreads_Children
ACC
AUC
MRR
NDCG
ACC
AUC
MRR
NDCG
ACC
AUC
MRR
NDCG
ACC
AUC
MRR
NDCG
ACC
AUC
MRR
NDCG
LightGCN
1.8
56.4
7.5
23.4
0.5
51.5
5.4
21.6
2.9
51.0
9.2
25.1
1.8
59.2
7.4
23.6
0.8
58.3
6.3
22.5
TWISTER-L
3.7
53.7
11.7
28.8
7.0
68.5
17.1
32.6
28.9
80.7
40.2
51.9
1.8
52.8
6.9
23.1
2.3
56.2
7.1
23.8
TWISTER-Q
5.9
59.7
12.5
27.9
9.6
69.6
19.0
34.1
6.2
59.1
19.1
34.9
2.6
54.2
7.3
24.1
2.0
53.8
7.5
24.7
Appendix
Table 7. Recommendation performance under the Uniform Masking setting. LightGCN is an interaction-only baseline evaluated with the TWISTER-compatible in-batch negative protocol.
Amazon Product Review Evaluation Prompt
You are evaluating the quality of an Amazon product review for a {category} .
Product ASIN: {book_id}
User ID: {user_id}
Rating: {rating}/5.0
Review Text: “ {review_text} ”
Please rate the review on a 5-point scale (1 = very poor, 5 = excellent):
Appendix
Table 8. Amazon Product Review Evaluation
Dataset
Variant
ACC
AUC
MRR
NDCG
Uniform Masking
Comics
GINE-Rating
1.5±0.4
54.7±2.9
6.6±0.8
22.5±0.9
GINE-ZeroFill
1.3±0.8
60.1±1.5
7.3±0.9
23.5±0.8
GINE- TWISTER -L
1.6±0.3
60.2±0.8
7.6±0.4
23.8±0.4
GINE- TWISTER -Q
1.6±0.5
59.7±0.4
7.5±0.7
23.6±0.6
Children
GINE-Rating
1.6±1.1
56.2±1.6
7.4±0.8
23.5±0.7
Appendix
Table 9. GINE recommendation performance with an alternative downstream backbone. Results are mean ± std over seeds, multiplied by 100. Best results within each dataset and metric are bolded.
Graph Neural Networks (GNNs) and Large Language Models (LLMs) have each advanced recommendation systems by modeling structural and semantic signals, respectively. However, integrating their complementary strengths remains challenging, particularly in sparse settings where maintaining semantic precision is critical. We propose TRWH (Text-driven Random Walk Heterogeneous Graph Neural Network), a novel framework that fuses LLM-generated textual profiles with heterogeneous graph structures through strategic random walk augmentation. TRWH consists of three core components: (1) Embedding Creation, which produces user and item representations using both Word2Vec and LLM-based profiling; (2) a Heterogeneous Graph Neural Network (HeteroGNN) that propagates information across multi-relational edges; and (3) Random Walk-based Path Construction, which enriches sparse graphs with second-order user-user and item-item links. Experiments on the Amazon-2023 Fashion (2M users, 825K items) and Beauty (631K users, 112K items) datasets demonstrate that TRWH achieves substantial performance gains over state-of-the-art methods, including 80.0% RMSE and 52.6% MAE reductions on Fashion, and 25.7% and 10.8% improvements on Beauty. Notably, while random walks improve performance with traditional embeddings, they can dilute the nuanced representations learned by LLMs, underscoring the importance of adaptive integration strategies.
He Ma, Chen Liu
The University of Sydney Camperdown, NSW, 2006, Australia · Nankai University Tianjing, China
Incorporating textual reviews into a Recommender System has become a prominent strategy for enriching collaborative signals with semantic information. However, the actual contribution of review-derived representations remains an open question, particularly when strong collaborative baselines are employed. In this work, we systematically investigate the impact of textual information on Matrix Factorization by introducing and comparing three enrichment strategies over a common collaborative backbone. First, we propose a learnable gating mechanism that adaptively balances collaborative and textual signals during training. This mechanism is applied to two distinct review representations: (i) aggregated topic profiles extracted from user and item histories, and (ii) full text embedding representations derived from reviews. Additionally, we explore a cross-attention mechanism that identifies and emphasizes the most informative dimensions of the textual representation before fusion with collaborative factors. We evaluate six variants: pure, enriched with topic profiles and text via gating; enriched with topics and text via gating; and enhanced with cross-attention over textual features. Experiments across multiple review-based datasets reveal that although adaptive fusion mechanisms improve representation flexibility, the marginal contribution of textual signals remains limited compared to the collaborative backbone. These findings suggest that, under typical rating-prediction settings, collaborative information continues to dominate performance, raising important considerations for the effective integration of semantic review signals into recommendation models.
Eduardo Ferreira da Silva, Mayki dos Santos Oliveira, Joel Machado Pires Denis Dantas Boaventura +1
Instituto de Computac¸˜ao – Universidade Federal da Bahia (UFBA) Salvador – BA – Brazil
Recommendation systems help users recommend relevant items from a large collection of choices. Present work on transformer-based sequential recommendation learns user preferences from interaction logs, but it mostly focuses on item identifiers and doesn't fully use the semantic meaning of items. This limitation becomes a major challenge in sparse and cold-start scenarios where historical interaction data is limited. To solve this problem, we introduce SISA-Rec (Semantically Integrated Sequential Recommendation), a transformer-based framework that embeds semantic context directly into sequential modeling. Our approach fuses item ID embeddings with BERT-based text embeddings via a gated fusion module, injects semantic similarity into the self-attention mechanism, and leverages an attention-based aggregation module to construct comprehensive user representations. Finally, a joint learning objective which combines Bayesian Personalized Ranking (BPR) and contrastive alignment loss, aligns the underlying behavioral and semantic spaces. Experiments were conducted on the two highly sparse Amazon Beauty and Amazon Toys & Games datasets, both having 99.93% sparsity. The results show that SISA-Rec outperforms state-of-the-art baseline models across all evaluation metrics. Compared with the BERT4Rec \cite{petrov2022systematic}, SISA-Rec improves HR@10 by 16.6% and NDCG@10 by 10.3% on Amazon Beauty, and HR@10 by 23.1% and NDCG@10 by 17.9% on Amazon Toys & Games. Cold-start analysis further shows that the proposed model achieves the largest improvements for users with limited interaction historical records. This showcases the value of semantic information when user behavior data is scarce. Overall, the results demonstrate that integrating semantic information into the attention mechanism leads to more accurate and reliable recommendations.
Soohan Abbasi, Shahid Munir Shah, Rafia Shaikh +1
Department of Computer Science, DHA Suffa University, Karachi, Sindh, Pakistan · Department of Computer Science, SZABIST, Pakistan · National University of Computer and Emerging Sciences (FAST), Karachi, Sindh, Pakistan +1