The data is a Delhi story
India is 90.6% of all restaurants and Delhi NCR alone is 83%. Outside NCR the file holds only about 20 restaurants per city.
Key takeaways
Four things to know before the charts.
Rating follows popularity and price
Votes (ρ = 0.68) and price tier (3.24 → 3.89) track rating far more than any service feature.
Table booking is a proxy for price
The +0.24 rating gap in India reverses inside the upscale tiers.
Data quality changed the story
Treating 2,148 "0" ratings as missing lifts the mean from 2.67 to 3.44.
Sample bias limits comparisons
India is 91% of rows, so city and country rankings compare samples, not markets.
Data cleaning
Nothing was deleted: impossible values became missing, and every change is logged in the notebook.
| Issue | Found | Fix |
|---|---|---|
| Encoding damage (�) in names, cities, addresses | 428 cells | 84% restored (Café, São Paulo, İstanbul); rest stripped |
| "Not rated" stored as rating 0 | 2,148 rows (22.5%) | Set to blank + is_rated flag. Mean rating 2.67 → 3.44 |
| Cost in 12 currencies | every row | Country-based currency + approximate USD columns |
| Wrong currency label (Philippines = "Botswana Pula") | 22 rows | Corrected |
| Average cost = 0 | 18 rows | Set to blank |
| Coordinates = 0 or outside the country | 500 rows | Set to blank + coords_valid flag |
| Missing cuisines | 9 rows | Filled with "Unknown" |
| Constant / redundant columns | 3 columns | Dropped. No duplicate restaurants found, no rows deleted |
Findings
Rating analyses use rated restaurants only; service comparisons use India only.
Ratings are bell-shaped, centred near 3.4
Mean 3.44. Only 301 restaurants are "Excellent" (4.5+). 22.5% have no rating at all (cleaned from a hidden 0).
More votes, higher ratings
Spearman ρ = 0.68. Mean rating rises from 2.98 (≤10 votes) to 4.20 (1,000+ votes). 32% of restaurants have under 10 votes, so many ratings are noisy.
Pricier restaurants rate higher
Mean rating climbs from 3.24 (budget) to 3.89 (fine dining), Kruskal-Wallis p < 0.001. In India, cost per person vs rating: ρ = 0.33.
Table booking's "rating boost" is mostly a price effect
In India, booking restaurants average 3.55 vs 3.31, but inside the Upscale and Fine-dining tiers they rate lower (3.59 vs 3.79, 3.62 vs 3.94). Online delivery shows no consistent effect (3.37 vs 3.34).
Which cuisines win in India
Mediterranean (3.98), European (3.91) and Asian (3.85) rate highest; Mithai (3.14) and Raw Meats (3.07) lowest. North Indian is the most common primary cuisine (2,992).
Wider menus, slightly higher ratings
Mean rating goes from 3.40 (one cuisine) to 3.66 (five or more), but the correlation is weak (ρ ≈ 0.12) and cost rises alongside.
Cities and localities
Ghaziabad, Faridabad (3.10) and Noida (3.16) rate lowest; London (4.54) and Rio (4.49) highest, though on ~20 restaurants each. Connaught Place is the busiest locality (122 restaurants, avg 3.78).
Countries are not comparable like-for-like
India averages 3.35 versus 3.6–4.5 elsewhere, but foreign samples are small and 99–100% rated, so this is not a quality ranking. Online delivery exists only for India (28%) and the UAE (47%).
Delhi NCR: dense, with mixed quality
Grey dots are unrated restaurants. Coordinates were validated against country bounds; 500 unusable points were excluded.
What moves with rating
Votes 0.68, cost per person 0.45, price range 0.40, table booking 0.13, cuisines 0.12, online delivery −0.04 (Spearman, rated restaurants).
Big chains rate below average
Domino's 2.93, Cafe Coffee Day 3.00 and Subway 3.00 sit well under the 3.44 mean; Starbucks (3.81) and Barbeque Nation (4.35) stand out.
Recommendations
Show vote counts next to ratings
32% of restaurants have fewer than 10 votes, and rating rises with votes. A rating from 5 votes should be labelled provisional.
Fix the data capture
Store "not rated" as missing, not 0, fix the currency labels and text encoding, and validate coordinates at entry.
Next: model and deep-dive
Predict rating from cost, votes, cuisines, services and location, and analyse Delhi NCR localities, the only area with enough data.
Method & caveats
Methods: Spearman correlation, Kruskal-Wallis and Mann-Whitney tests (ratings are not normally distributed), stratified comparisons by price tier to control for confounding.
Caveats: USD costs use fixed, approximate 2018 exchange rates. About 70 damaged text cells could not be restored. The sample is not random, so results are associations, not causes.