EdenMusk/cartcompare-ai-dataset

CartCompare AI — Exploratory Data Analysis (EDA)

This document is the EDA deliverable for the CartCompare AI synthetic product catalog (cartcompare_products.csv). It covers the full dataset — 11,321 records, 11 columns, generated by prompting Qwen2.5-1.5B-Instruct (Hugging Face, few-shot prompting) across 3 synthetic store chains (GreenBasket, ValueMart, FreshCo) and 10 target product categories.

Why EDA matters here: an AI-generated catalog is not guaranteed to be correct. Before building the embeddings-based Recommendation engine on top of this data, we needed to check it for AI mistakes — wrong categories, hallucinated brands, impossible prices — so the matching engine doesn't silently inherit garbage. EDA is the quality gate between "the AI invented data" and "we trust this data enough to build on it."

Methodology note: every issue below was handled with a flag column (flag_unmapped_category, flag_real_brand, flag_promo_contradiction, flag_messy_size, flag_price_outlier) rather than deleting rows. The assignment requires a minimum of 10,000 records, so deleting rows to "fix" the data risked dropping below that floor. Flagging keeps every correction visible and reversible instead of silently vanishing. In total, 1,151 records (10.2%) carry at least one flag — the other ~90% of the catalog is clean as generated.

All numbers below were computed directly from the full, unsampled dataset.

---

1. Category drift

The model was prompted to use only 10 fixed categories, but the data actually contains 234 distinct category strings. 668 records (5.9%) fall outside the 10 requested categories — some are near-duplicates of a real category (e.g. "Produce" instead of "Fruits & Vegetables"), others are genuinely unrelated to groceries (Safety Devices, Lighting, Toys & Hobbies — apparent bleed from the Baby Products category).

This is the most consequential finding because it directly affects the Recommendation stage: whether to trust the category field for matching, or rely on product_name instead.

2. Hallucinated real-world brand names

Only fictional brand names were supplied in the generation prompt, but the catalog contains real, recognizable brands anyway — Nestlé, Kellogg's, Coca-Cola, PepsiCo, Eggland's Best, Kerrygold, General Mills, Nature Valley, and others. A classic LLM hallucination: the model "recalled" real names from its training data instead of inventing new ones as instructed.

3. Promotion-flag contradictions

85 records are marked on_promotion = False but still carry a full promotion_description — an internal contradiction in the generated data.

4. Messy size field

A subset of records contain free-text reasoning instead of a clean value — e.g. "Not specified but likely 2 packs", "Packaging varies but typically about 6 inches thick". The model "thought out loud" inside a field that should only hold a number/unit.

5. Price outliers

Most high prices (up to ~$199.99) are legitimate — expensive baby gear (strollers, car seats, carriers) genuinely costs $150–300. But several Meat & Fish records are unrealistic hallucinations, e.g. "Pork Shoulder 5 lbs" at $199.90 and "Chuck Roast Beef 10lbs" at $199.90 (realistic range: ~$15–25). The left panel shows the full Meat & Fish price distribution with the IQR outlier fence; the right panel lists the worst offenders.

Bonus — Store × category price heatmap

Average price per store × category, confirming the three chains (GreenBasket, ValueMart, FreshCo) are reasonably balanced in pricing across categories, with Baby Products and Meat & Fish consistently the most expensive categories in every store.

Bonus — Top product tags

Across 34,083 total tag instances (11,640 distinct tags) on the tags field, the same handful of words recur far more than chance would predict: "organic" (589 products, 5.2%), "gluten-free" (422, 3.7%), "healthy" (244, 2.2%), "natural" (183, 1.6%). This style bias shows up even outside food where it makes less sense — e.g. "eco-friendly" (116, 1.0%) sits in the top 20 overall, surfacing on cleaning/household products next to the same "organic"/"healthy" language used for groceries. This isn't a data-quality bug like the five issues above, but a subtle stylistic pattern worth noting about the generation model's behavior: exact counts, not a word cloud, because counts are the thing that's actually being claimed.

---

Summary table

#	Finding	Records flagged	% of dataset
1	Category drift (outside the 10 requested categories)	668	5.90%
2	Hallucinated real-world brand names	245	2.16%
3	Promotion-flag contradictions	85	0.75%
4	Messy size field (free-text reasoning)	133	1.17%
5	Price outliers (Meat & Fish, IQR method)	52	0.46%
—	Any flag (at least one issue)	1,151	10.17%

Note: exact counts for findings 2 and 4 depend on the matching rule used (brand list / hedging-keyword list respectively) — the underlying pattern and example records are the same regardless of the exact threshold chosen.

Checked and ruled out: 791 rows sharing the same store_name + product_name looked like duplicates at first glance, but turned out to be the same generic product name (e.g. "Almond Butter") sold under multiple brands/prices — realistic, not a bug.

---

Generated from cartcompare_products.csv (raw, pre-cleaning). The cleaned, flag-augmented file is cartcompare_products_clean.csv (11,321 rows × 16 columns).