Research
ReviewGraph: knowledge graphs for review rating prediction
A research framework that turns customer reviews into a knowledge graph and combines graph embeddings with sentiment features to predict review ratings. Peer-reviewed and presented at ICKG 2025.
The paper
- Title
- ReviewGraph: A Knowledge Graph Embedding Based Framework for Review Rating Prediction with Sentiment Features
- Authors
- A.J.W. de Vink, Natalia Amat-Lefort, Lifeng Han
- Published in
- 2025 IEEE International Conference on Knowledge Graph (ICKG), 13–14 November 2025, Limassol, Cyprus, pp. 43–50
- DOI
- 10.1109/ICKG66886.2025.00013
- Preprint
- arXiv:2508.13953 (submitted 19 August 2025, revised 15 November 2025)
- Founder of Vink Intelligence
- Bert de Vink is the first author, published as A.J.W. de Vink.
What it does
ReviewGraph transforms textual customer reviews into a knowledge graph by extracting (subject, predicate, object) triples and associating sentiment scores with them.
Graph embeddings (Node2Vec) and sentiment features are then used by machine learning classifiers to predict the review rating.
The paper compares this with traditional NLP baselines (Bag of Words, TF-IDF, Word2Vec) and with large language models, on the HotelRec dataset.
What the paper reports
- Performance similar to the best-performing model in the literature it compares with, at lower computational cost (without ensemble).
- Predictive performance comparable to large language models, and better than the baselines on agreement-based metrics such as Cohen's Kappa.
- Added advantages in interpretability, visual exploration and potential integration into retrieval-augmented generation (RAG) systems.
These are the authors' reported results, summarised from the paper's abstract. Read the paper for the full method, data and limitations.
Experimental setup
- Data: the first 10,000 reviews of the HotelRec dataset (TripAdvisor), covering 59 hotels. Five-star ratings dominate, so sampling strategies were compared.
- Graph: triples extracted from review text, stored in Neo4j, with average, minimum and maximum relationship sentiment computed per review. Node2Vec embeddings were tested at several dimensions.
- Comparison: Random Forest, Logistic Regression and a neural network on graph features, against Bag of Words, TF-IDF and Word2Vec baselines and a GPT-4o rating-prediction model. Metrics: accuracy, MAE, RMSE and Cohen's Kappa.
Findings from the paper
- The best ReviewGraph configuration used 5 embedding dimensions, oversampling and the full sentiment features (average, minimum, maximum). Random Forest performed best on ReviewGraph features, but not on the baselines.
- Adding minimum and maximum sentiment improved accuracy and especially Cohen's Kappa. Node2Vec embeddings alone already carried substantial predictive signal.
- The authors report no substantial difference between the best baseline, the best ReviewGraph model and the LLM model, with a Kappa similar to the LLM and a slightly lower accuracy.
Limitations and next steps
- Triple extraction was crude and produced many low-quality triples. The authors expect a fine-tuned LLM for extraction to help.
- Node2Vec must be retrained when a node is added. GraphSAGE and other graph neural networks are named as follow-up work, as are GraphRAG summaries and human evaluation of the visualisation.
Why it matters to us
ReviewGraph is the clearest published example of how we approach knowledge graphs: structure that makes information connected, interpretable and usable by machine learning. The same thinking applies to other domains, such as the biomedical graphs in Vink Drug.
Our knowledge graph services →Cite this work
de Vink, A. J. W., Amat-Lefort, N. and Han, L. (2025). ReviewGraph: A Knowledge Graph Embedding Based Framework for Review Rating Prediction with Sentiment Features. In 2025 IEEE International Conference on Knowledge Graph (ICKG), pp. 43–50. doi:10.1109/ICKG66886.2025.00013
BibTeX
@inproceedings{devink2025reviewgraph,
title = {ReviewGraph: A Knowledge Graph Embedding Based Framework for Review Rating Prediction with Sentiment Features},
author = {de Vink, A. J. W. and Amat-Lefort, Natalia and Han, Lifeng},
booktitle = {2025 IEEE International Conference on Knowledge Graph (ICKG)},
year = {2025},
pages = {43--50},
doi = {10.1109/ICKG66886.2025.00013}
}