plan/02-dataset.md

Synthetic Dataset Plan

Purpose

Generate a realistic e-commerce dataset to train and evaluate all recommendation approaches. Synthetic data gives full control over data characteristics (sparsity, user segments, product diversity).

Dataset Components

1. Products (products.csv) — ~500 items

Field	Type	Example
product_id	int	1
name	str	"Samsung Galaxy S24"
category	str	"Electronics"
subcategory	str	"Smartphones"
brand	str	"Samsung"
price	float	799.99
description	str	"Latest Samsung flagship with AI features"
features	str	"6.8in, 256GB, 12GB RAM"
avg_rating	float	4.5
num_reviews	int	234

Categories: Electronics, Clothing, Home & Kitchen, Books, Sports, Beauty, Toys, Automotive

2. Users (users.csv) — ~200 users

Field	Type	Example
user_id	int	1
name	str	"Alice"
age	int	28
gender	str	"Female"
preferred_categories	str	"Electronics,Books"
budget_min	float	100.0
budget_max	float	1500.0
favorite_brands	str	"Samsung,Sony"

User segments: budget-conscious, brand-loyal, category-specific, high-spenders

3. Ratings (ratings.csv) — ~5000 entries

Field	Type	Example
user_id	int	1
product_id	int	42
rating	float	4.0
timestamp	int	1700000000

Sparsity: ~5% density (controlled to allow meaningful CF comparison)

4. Interactions / Purchases (interactions.csv)

Field	Type	Example
user_id	int	1
product_id	int	42
purchased	bool	True
quantity	int	1

Generation Logic

1. Create products with realistic brand-category mappings
2. Create users with varied preference profiles
3. Generate ratings based on:
   • User's preferred categories → higher ratings
   • User's favorite brands → higher ratings
   • Price affinity (if within budget) → higher ratings
   • Random noise (10-15%) for realism
4. Create train/test split (80/20) for evaluation

Cold-Start Simulation

• 10% of users have < 3 ratings (cold-start for CF)
• 10% of products have < 2 ratings (new items)
• This allows meaningful comparison of approaches under different conditions