adityasuyal/cafe-sales-cleaned

☕ Cafe Sales Dataset (Cleaned)

A cleaned and preprocessed version of a synthetic cafe sales transaction dataset. Originally a dirty, real-world-style dataset riddled with missing values, corrupt entries, and wrong data types — now fully cleaned and ready for analysis or ML tasks.

---

---

📋 Dataset Overview

Property	Value
Original Rows	10,000
Cleaned Rows	9,540
Columns	8
Time Period	January 2023 – December 2023
Domain	Retail / Food & Beverage

---

📁 Files

File	Description
cleaned_cafe_sales.csv	Final cleaned dataset, ready for use
dirty_cafe_sales.csv	Original raw dataset with all issues intact
cafe_data_cleaning.ipynb	Full cleaning pipeline notebook (step-by-step)

---

🗂️ Column Description

Column	Type	Description
Transaction ID	string	Unique identifier for each transaction (e.g., TXN_1961373)
Item	string	Menu item purchased (Coffee, Tea, Cake, Cookie, Sandwich, Salad, Juice, Smoothie)
Quantity	float	Number of units purchased (1–5)
Price Per Unit	float	Price of a single unit in USD (1.0–5.0)
Total Spent	float	Total transaction amount = Quantity × Price Per Unit
Payment Method	string	Payment type used (Cash, Credit Card, Digital Wallet)
Location	string	Where the order was placed (In-store, Takeaway, Unknown)
Transaction Date	string	Date of transaction in YYYY-MM-DD format

---

🧹 Data Cleaning — What Was Wrong & How It Was Fixed

The raw dataset had several real-world data quality issues. Here's a complete breakdown:

1. Corrupt String Values ("ERROR" and "UNKNOWN")

Problem: Multiple columns contained "ERROR" and "UNKNOWN" strings instead of actual values or proper nulls. These were scattered across Item, Quantity, Price Per Unit, Total Spent, Payment Method, and Location.

Fix: All "ERROR" and "UNKNOWN" strings were replaced with NaN (proper missing values) using df.replace(["ERROR", "UNKNOWN"], np.nan).

Impact after replacement:

Column	Missing Before Replace	Missing After Replace
Item	333	969
Quantity	138	479
Price Per Unit	179	533
Total Spent	173	502
Payment Method	2,579	3,178
Location	3,265	3,961
Transaction Date	159	460

---

2. Wrong Data Types

Problem: All numeric columns (Quantity, Price Per Unit, Total Spent) were loaded as object/str dtype because of the corrupt string values mixed in.

Fix: Converted all three to float64 using pd.to_numeric(..., errors='coerce'), which automatically turned any remaining non-numeric values into NaN.

---

3. Missing Values in Categorical Columns

Problem: Item, Payment Method, and Location had significant missing data (up to ~32–40%).

Fix:

• Item → filled with mode (most frequent value)
• Payment Method → filled with mode (most frequent value)
• Location → filled with "Unknown" (a distinct, interpretable label rather than imputing a false value, since ~40% was missing — too high to impute reliably)

---

4. Missing Values in Numeric Columns

Problem: Quantity (479 missing) and Price Per Unit (533 missing) had nulls after type conversion.

Fix: Both filled with their respective medians (both were 3.0), which is robust against skew and the limited range of values (1–5).

---

5. Missing Total Spent — Derived Column Recovery

Problem: 502 rows had missing Total Spent.

Fix: Since Total Spent = Quantity × Price Per Unit was a deterministic relationship (verified — 0 mismatches existed in rows where all three were present), missing values were recalculated directly: df["Total Spent"] = df["Total Spent"].fillna(df["Quantity"] * df["Price Per Unit"]).

---

6. Missing & Invalid Transaction Date

Problem: 460 rows had missing or invalid dates (including "ERROR" strings parsed as NaT).

Fix: Parsed the column to datetime64 using pd.to_datetime(..., errors='coerce'), then dropped all rows where the date was null. Dates are a critical anchor for time-series analysis — imputing fake dates would corrupt temporal patterns.

Result: 10,000 → 9,540 rows after dropping 460 undated rows.

---

7. Total Spent Consistency Recheck

Problem: After all imputations, a final audit revealed 696 rows where Total Spent didn't match Quantity × Price Per Unit (due to median-filled values for Quantity/Price Per Unit changing the expected total).

Fix: Recomputed Total Spent from scratch for all rows: df["Total Spent"] = df["Quantity"] * df["Price Per Unit"]. This guarantees 100% internal consistency.

---

8. Duplicate Check

Problem: Potential duplicate rows or duplicate Transaction IDs.

Result: ✅ 0 duplicate rows. ✅ 0 duplicate Transaction IDs. No action needed.

---

9. Whitespace in Categorical Columns

Fix: Applied .str.strip() to Item, Payment Method, and Location to remove any leading/trailing spaces.

---

✅ Final Dataset Stats

Shape: (9540, 8)
Missing values: 0 in all columns
Duplicate rows: 0
Duplicate Transaction IDs: 0
Total Spent mismatches: 0

Data Types:

Transaction ID       str
Item                 str
Quantity             float64
Price Per Unit       float64
Total Spent          float64
Payment Method       str
Location             str
Transaction Date     str (YYYY-MM-DD)

---

📊 Value Distributions (Cleaned Data)

Item:

Item	Count
Juice	2,051
Coffee	1,123
Salad	1,099
Cake	1,082
Sandwich	1,075
Smoothie	1,048
Cookie	1,035
Tea	1,027

Payment Method:

Method	Count
Digital Wallet	5,212
Credit Card	2,170
Cash	2,158

Location:

Location	Count
Unknown	3,779
Takeaway	2,889
In-store	2,872

Numeric Ranges:

Column	Min	Max	Mean	Median
Quantity	1.0	5.0	3.02	3.0
Price Per Unit	1.0	5.0	2.95	3.0
Total Spent	1.0	25.0	8.92	8.0

---

💻 Usage

import pandas as pd

df = pd.read_csv("cleaned_cafe_sales.csv", parse_dates=["Transaction Date"])
df.head()

---

🛠️ Cleaning Notebook

The full cleaning process is documented step-by-step in cafe_data_cleaning.ipynb. It covers:

• Initial EDA and dirty data exploration
• Corrupt value detection and replacement
• Type casting and null handling strategies
• Derived column consistency verification
• Final quality audit

---

📜 License

This dataset is released under the MIT License. Free to use for personal, educational, and commercial purposes.

---

👤 Author

Aditya Suyal
HuggingFace Profile