---
language:
- en
tags:
- recommender-systems
- sequential-recommendation
- amazon-reviews
- side-information
pretty_name: Amazon Reviews 2023 (10 Categories, Post-processed)
license: mit
size_categories:
- 1M<n<10M
---

# Amazon Reviews 2023 (10 Categories, Post-processed)

## Overview
This dataset is a curated and post-processed subset of **Amazon Reviews 2023**. We select **10 product categories** and apply a standard preprocessing pipeline widely used in sequential recommendation research. The resulting dataset provides user interaction sequences along with structured item side information.

- **Original source**: https://amazon-reviews-2023.github.io/
- **Categories**: 10
- **Content**: user interaction sequences + structured item features  
- **Splits**: `train`, `valid`, `test`

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

---

## Included Categories (alphabetical)

1. Arts_Crafts_and_Sewing  
2. Baby_Products  
3. Beauty_and_Personal_Care  
4. Books  
5. Health_and_Household  
6. Industrial_and_Scientific  
7. Musical_Instruments  
8. Sports_and_Outdoors  
9. Toys_and_Games  
10. Video_Games  

### Why these 10 categories?
We chose these categories because after filtering items with missing structured side information, the remaining number of interactions does not drop significantly compared to many other categories.

---

## Post-processing Pipeline

The dataset is processed **per category** as follows:

1. **5-core filtering**  
   Users and items with fewer than five interactions are removed.

2. **Chronological ordering**  
   The remaining interactions are sorted by time to form user behavior sequences.

3. **Leave-one-out split**  
   A leave-one-out strategy is used to construct `train/valid/test` splits from each user’s chronological sequence.

4. **Ensure valid evaluation targets**  
   If a target item in an evaluation split does not appear in the training data, it is added to the training set to ensure valid evaluation.

5. **Structured side information extraction and filtering**  
   For each item, we extract structured side information (store identifier and 1st/2nd/3rd-level category identifiers). Items missing any of these fields are filtered out, and interactions involving them are removed accordingly.

---

## Directory Layout (per category)

Each category has its own folder containing:

- `item_feature_explain.json`  
  Statistics for each item-side feature. For each feature, `id_num` is defined as **(#unique values in this category) + 1**, where the extra `+1` slot is reserved for a **mask token**.

- `item_feature/`  
  A Parquet table storing the structured side information for each item.

- `train/`  
  A Parquet dataset storing training user sequences.

- `valid/`  
  A Parquet dataset storing validation user sequences.

- `test/`  
  A Parquet dataset storing test user sequences.

---

## Intended Use
This dataset is intended for research on:
- Sequential recommendation / next-item prediction
- Methods that incorporate structured item side information

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
```bibtex
@misc{ReSID,
      title={Rethinking Generative Recommender Tokenizer: Recsys-Native Encoding and Semantic Quantization Beyond LLMs}, 
      author={Yu Liang and Zhongjin Zhang and Yuxuan Zhu and Kerui Zhang and Zhiluohan Guo and Wenhang Zhou and Zonqi Yang and Kangle Wu and Yabo Ni and Anxiang Zeng and Cong Fu and Jianxin Wang and Jiazhi Xia},
      year={2026},
      eprint={2602.02338},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2602.02338}, 
}
```