PersonaShop: Personalized Shopping Questions and User Profiles

Evaluation data for personalized shopping assistants. Each example pairs a reviewer's preference profile with a target product and a set of decision-point questions whose best answer depends on that profile, so an assistant that ignores the user's history under-performs.

The data covers 33 Amazon product categories and is derived from Amazon Reviews 2023.

It accompanies the paper Personalization Matters: Long-Horizon Conversation Agent with User-Centric Information in Online Shopping Interactions (accepted at AACL-IJCNLP 2026).

Subsets

Config	Rows	Unique user ids	Contents
questions (default)	327	231	Target product, user profile and 6–8 personalized questions per row (2,415 questions in total)
user_profiles	329	232	One preference profile per (category, user)

Both subsets have a single test split; the data is intended for evaluation, not training. Each category contributes 8–10 rows.

from datasets import load_dataset

questions = load_dataset("PersonalizedAgents/PersonaShop", "questions", split="test")
profiles = load_dataset("PersonalizedAgents/PersonaShop", "user_profiles", split="test")

Fields

questions

Field	Type	Description
category	string	Amazon Reviews 2023 category name, e.g. All_Beauty
user_id	string	Reviewer id (see User ids below)
target_product.product_meta	struct	Item metadata from Amazon Reviews 2023 (see below)
target_product.rating	int	The user's star rating of the target product (1–5)
target_product.review_title	string	Title of the user's review of the target product
target_product.review_text	string	Text of the user's review of the target product
user_profile	struct	The user's profile; identical to the matching row in user_profiles
generated_questions	list of string	Personalized decision-point questions about the target product, written from the user's point of view

product_meta contains main_category, title, store, parent_asin, subtitle, author, features, description, categories, average_rating, rating_number, price and details.

The user's own rating and review of the target product are included as reference; they should not be shown to the system under evaluation.

user_profiles

Field	Type	Description
category	string	Category the profile was built for
user_id	string	Reviewer id
Preferences	string	What the user tends to like and rate highly
Style	string	Inferred style preferences
Budget	string	Inferred price sensitivity
Size	string	Size or fit preferences, when the reviews mention them
Quality_Priorities	string	Quality aspects the user cares about
Dislikes	string	What the user rates poorly or avoids
Pain_Points	string	Recurring problems the user reports
Additional_Notes	string	Other observations

Profiles are built per category, so a user who appears in several categories has a separate profile in each. 36 user ids appear in more than one category.

How the data was built

1. User profiles. For each sampled user in a category, GPT-4o summarizes the user's review history in that category into the eight profile fields.
2. Target product. A product the user reviewed is used as the target, together with the user's rating and review of it.
3. Questions. GPT-4o writes questions a shopper with that profile would ask when deciding on the target product. Questions typically refer to the user's history implicitly ("given my past choices…") rather than restating it.

Profiles and questions are model-generated.

Changes from the source files

The release files are produced from the per-category JSON files by a conversion script. Content is unchanged except for the following:

• The 33 per-category files of each subset are merged into one JSONL file and a category column is added.
• user_profiles is flattened from a {user_id: profile} mapping to one row per user.
• product_meta.price is a float; missing or non-numeric prices (NaN, "None", "—") are null. 191 of 327 rows have a price.
• product_meta.details is a JSON-encoded string, because its keys differ from product to product. Decode it with json.loads.
• product_meta.bought_together is dropped; it was empty for every product.
• NaN values in other fields are null. subtitle and author are only populated for book-like products, and author is the string form of the source author record.

User ids

user_id values are the reviewer ids published in Amazon Reviews 2023. A small number of ids carry a numeric suffix such as _1 or _2; these are kept exactly as they appear in the source files and match between the two subsets.

One profile in user_profiles has no corresponding row in questions.

Limitations

• The benchmark is small: 327 user–product pairs.
• Profiles are an LLM's reading of review text and may over-generalize from a few reviews or state preferences the user never expressed.
• Reviews, and therefore profiles and questions, are in English and reflect Amazon's US marketplace.
• Product metadata is a snapshot from Amazon Reviews 2023; prices and ratings may be out of date.

License and source data

The dataset has two kinds of content, licensed separately:

• Generated content — the user profiles (the eight profile fields in user_profiles and in questions.user_profile) and generated_questions — is released by the authors under CC BY 4.0.
• Source content — user_id, category, and everything under target_product (product metadata, ratings, review titles and review text) — comes from Amazon Reviews 2023 and remains subject to the terms of that dataset. No license is granted here for this material.

Citation

If you use this data, please cite the paper:

@article{gao2026personalization,
  title={Personalization Matters: Long-Horizon Conversation Agent with User-Centric Information in Online Shopping Interactions},
  author={Gao, Rena and Dai, Yue and Guan, Hao and Gao, Shengxiang and Wu, Wangyang and Shen, Yixin and Lau, Jey Han},
  journal={arXiv preprint arXiv:2610.11375},
  year={2026},
  note={Accepted at AACL-IJCNLP 2026}
}

Please also cite Amazon Reviews 2023:

@article{hou2024bridging,
  title={Bridging Language and Items for Retrieval and Recommendation},
  author={Hou, Yupeng and Li, Jiacheng and He, Zhankui and Yan, An and Chen, Xiusi and McAuley, Julian},
  journal={arXiv preprint arXiv:2403.03952},
  year={2024}
}