PIIR/KGD-dataset

Amazon Reviews 2023 (8 Categories, KGD Preprocessed)

Paper | Code | Original Source

Overview

This dataset provides the preprocessed data used to train and evaluate KGD (Knowledge–Geometry Decoupling) with ManCAR as the backbone. It is a curated subset of Amazon Reviews 2023 covering 8 product categories, and ships not only the interaction splits but also the auxiliary signals required by KGD:

• ready-to-train user interaction sequences (train/valid/test);
• collaborative and semantic item embeddings used by BMTP pretraining;
• a Manifold-Constrained swing graph used by ManCAR.

This dataset is derived from Amazon Reviews 2023 and is intended for research use.

---

Included Categories (alphabetical)

1. Arts_Crafts_and_Sewing
2. Beauty_and_Personal_Care
3. CDs_and_Vinyl
4. Cell_Phones_and_Accessories
5. Office_Products
6. Software
7. Toys_and_Games
8. Video_Games

---

Post-processing Pipeline

The dataset is processed per category as follows:

1. Chronological ordering & leave-last-out split Interactions are sorted by time into user behavior sequences and split into train/valid/test with a leave-last-out strategy.

2. Vocabulary-consistency filtering The item vocabulary is built from the training split; validation/test rows whose target item or history contains unseen items are removed to prevent train-test leakage.

3. Collaborative embedding LightGCN embeddings (dimension 256) trained on a user-item graph whose edges come from the training split only.

4. Semantic embedding Qwen3 text embeddings (dimension 32) encoded from each item's title and hierarchical categories.

5. Swing graph An item-item swing similarity graph providing the Manifold-Constrained supervision for ManCAR.

---

Directory Layout (per category)

Each category has its own folder under the three top-level directories (using Software as an example):

• processed_llo_graph/Software/
  • Software.train.csv, Software.valid.csv, Software.test.csv — user interaction sequences.
  • Software.item.csv — item table with remapped item IDs and side information.
  • graph/swing.parquet — Manifold-Constrained swing graph (ManCAR).
• graph_emb/Software/graph_emb.csv — collaborative (LightGCN) item embeddings (BMTP).
• text_emb/Software/text_emb.csv — semantic (Qwen3) item embeddings (BMTP).

---

Intended Use

This dataset is intended for research on:

• Sequential recommendation / next-item prediction
• Pretrain-then-transfer methods that leverage collaborative and semantic item embeddings

It is not intended for user identification or any high-stakes decision-making.

---

Licensing & Attribution

This dataset is derived from Amazon Reviews 2023. Please refer to the original dataset page for licensing/usage terms and attribution requirements:

• https://amazon-reviews-2023.github.io/

If you use this processed dataset, please cite the original dataset and clearly state that you used a post-processed subset with the pipeline described above.

---

Citation

@misc{KGD,
      title={Knowledge-Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation}, 
      author={Zixuan Wang and Yuhong Chen and Yuxuan Zhu and Guidong Lei and Zhiluohan Guo and Yu Zhao and Kun Wang and Bangyang Hong and Kangle Wu and Yabo Ni and Anxiang Zeng and Cong Fu and Hui Li},
      year={2026},
      eprint={2608.02738},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2608.02738}, 
}