aslan-ng/amazon_products_2025

Dataset Summary

A compact, hand-curated dataset of Amazon top-selling products (2025).
Each record contains product title, description, average_rating, rating_number, main_category, and a derived product_quality_score that uses a Bayesian-adjusted rating formula to estimate the true quality of products.

This dataset is designed for rapid prototyping of:

• product topic classification (e.g., infer main_category from text)
• rating prediction / quality regression
• ranking / recommendation experiments
• lightweight product text understanding benchmarks

Note: The dataset is small (~500 samples) for clarity, reproducibility, and teaching use.

Supported Tasks

• Text Classification (title, description → main_category)
• Text Regression (title, description → product_quality_score)
• Item Ranking (based on Bayesian-adjusted product quality)

Languages

• English (en)

Dataset Structure

Data Instances

{
  "title": "Wireless Noise-Cancelling Headphones",
  "description": "Over-ear design, 30 hours of playtime, adaptive sound control.",
  "average_rating": 4.7,
  "rating_number": 20431,
  "main_category": "Electronics",
  "product_quality_score": 0.918
}

Data Fields

Field	Type	Description
title	string	Product title as listed by the marketplace.
description	string	Short marketing text or summary of product features.
average_rating	float64	Average user rating (0–5).
rating_number	int64	Total number of user ratings.
main_category	string	Coarse product category label.
product_quality_score	float64	Bayesian-adjusted product score, normalized to [0, 1], combining average_rating and rating_number.

Calculation of product_quality_score

product_quality_score is computed using a Bayesian average formula to reduce bias toward products with few ratings.
The score combines each product’s observed average (R) and the global mean (C) weighted by the number of votes (v) and a prior constant (m):

score = (v / (v + m)) × R + (m / (v + m)) × C

where:

• ( R ) = product’s average rating
• ( v ) = number of ratings
• ( C ) = mean rating across all products
• ( m ) = minimum votes threshold (empirically chosen, e.g. 100)

The resulting score is then normalized to [0, 1] for consistency across tasks.

This approach favors well-rated products that also have sufficient review volume, mitigating small-sample noise.

Example Python Implementation

def bayesian_score(R, v, C, m=100):
    """
    Compute Bayesian-adjusted product quality score.

    Args:
        R (float): product average rating
        v (int): number of ratings
        C (float): mean rating across all products
        m (int): minimum votes threshold (prior weight)
    Returns:
        float: normalized Bayesian quality score in [0, 1]
    """
    bayes = (v / (v + m)) * R + (m / (v + m)) * C
    return bayes / 5.0  # normalize to [0,1] assuming rating scale 0–5

Limitations

• Small scale; not representative of the entire Amazon catalog.
• Single-snapshot dataset (no time evolution).
• Derived Bayesian prior constant (m) is approximate.
• Some descriptions truncated to fit Hugging Face tokenization limits.

Maintainers

• Maintainer: aslan-ng
• Contributions: PRs welcome for new categories, Bayesian hyperparameter tuning, or evaluation splits.