Skip to content

04-amazon-reviews.md

Latest commit

 

History

History
64 lines (55 loc) · 3.59 KB

04-amazon-reviews.md

File metadata and controls

64 lines (55 loc) · 3.59 KB
Raw

4. Amazon Customer Reviews (Amazon Reviews '23, McAuley Lab)

From the dataset (verbatim from amazon-reviews-2023.github.io)

"Large-scale Amazon Reviews dataset, collected in 2023 by the McAuley Lab, contains 571.54M reviews, 245.2M users, 48.19M items, and 30.14B review tokens. Covers May 1996 to September 2023."

Files & Schema (from dataset documentation)

Two parallel files per category, gzipped JSONL:

Reviews (<Category>.jsonl.gz):

Field Type Description
rating float 1.0 – 5.0 stars
title str Review title
text str Review body
images list Embedded image URLs
asin str Item ID
parent_asin str Group ID (variants share parent)
user_id str Reviewer ID
timestamp int Unix milliseconds
verified_purchase bool Verified buyer flag
helpful_vote int "Was this helpful" upvotes

Metadata (meta_<Category>.jsonl.gz):

  • main_category, title, average_rating, rating_number, features, description, price, images, videos, store, categories, details, parent_asin, bought_together

Scale (from documentation)

  • 33 product categories: All_Beauty, Amazon_Fashion, Books, Electronics, Home_and_Kitchen, etc.
  • 571.54M reviews, 48.19M items, 54.51M users, ~30B tokens of text.
  • Categories range from <1M (All_Beauty) to >50M reviews (Books, Electronics).

Representative sample row (General knowledge — schema-faithful)

{
  "rating": 5.0,
  "title": "Worth every penny",
  "text": "Sound is surprisingly full for the size. Alexa picks up my voice from the next room.",
  "asin": "B07FZ8S74R",
  "parent_asin": "B07FZ8S74R",
  "user_id": "AGKHLEW2SOWHNMFQIJGBECAF7INQ",
  "timestamp": 1697822400000,
  "verified_purchase": true,
  "helpful_vote": 27,
  "images": []
}

How a Retail PM uses this (General knowledge)

  1. Voice-of-customer themes per category — topic modeling (BERTopic / LDA) on review text → roadmap signal ("what do shoppers complain about for kitchen knives?").
  2. Aspect-based sentiment — extract aspect–opinion pairs ("battery life: short", "fit: runs small") to feed PDP improvements and supplier-quality conversations.
  3. Competitor benchmarking — pull reviews for competing SKUs (same parent_asin grouping or same category) and compare complaint clusters quantitatively.
  4. Review-quality / fake-review detection — verified_purchase=False reviews with extreme polarity are the easy training set; production models add temporal burst and user-graph signals.
  5. Image-text alignment — review images + text → multimodal model fine-tuning for visual search.
  6. Pricing perception — match price in metadata to review sentiment over time; quantify the "price-up complaint" effect.
  7. Star-rating drift — average_rating is point-in-time; reconstruct true rating curves from review timestamps to detect quality regressions.

Gotchas

  • No PII, but user_id is consistent across reviews — re-identification risk if cross-joined with other Amazon data. Don't ship to third parties.
  • Item metadata is scraped at a single point in time; prices and titles may be stale relative to the reviews.
  • Books and Electronics categories dominate. Be careful generalizing tone/length stats across categories.
  • The 2023 version supersedes the older 2018 release; columns are similar but not identical.