Hands-on Practical: Basic Data Cleaning Steps

A practical example demonstrates cleaning up a small, messy dataset, preparing it for a machine learning algorithm.

"Imagine we have a dataset representing information about used cars, and we want to predict their price. Data often looks less than perfect, much like this sample:"

Initial Car Dataset

Mileage (km)	Age (years)	Engine Type	Condition	Price ($)
50000	3	Petrol	Good	15000
120000	8	Diesel	Fair	8000
80000	5	Petrol	Good	11000
NaN	10	Diesel	Excellent	9500
65000	4	NaN	Good	13000
150000	12	Petrol	Fair	NaN
95000	6	Hybrid	Excellent	18000

We can immediately spot a few issues algorithms won't like:

• Missing Values: There are NaN (Not a Number) entries in the Mileage, Engine Type, and Price columns.
• Mixed Data Types: We have numerical data (Mileage, Age, Price) and categorical data (Engine Type, Condition). Algorithms typically require all input features to be numerical.
• Different Scales: Mileage has values in the tens or hundreds of thousands, while Age is in single or double digits. This difference in scale can sometimes cause problems for certain algorithms.

Let's tackle these step-by-step.

Step 1: Handling Missing Values

First, we need to deal with the NaN entries. As discussed earlier, common strategies include removing rows/columns or imputation (filling in missing values). Since our dataset is small, removing rows might discard too much information. Let's try imputation.

• Numerical Features (Mileage, Price): We can fill missing numerical values with the mean or median of the column. The median is often preferred if the data might have outliers (extreme values) that could skew the mean. Let's calculate the median for Mileage and Price from the non-missing values:

  • Existing Mileages: 50000, 120000, 80000, 65000, 150000, 95000. Sorted: 50000, 65000, 80000, 95000, 120000, 150000. Median (average of middle two) = (80000 + 95000) / 2 = 87500.
  • Existing Prices: 15000, 8000, 11000, 9500, 13000, 18000. Sorted: 8000, 9500, 11000, 13000, 15000, 18000. Median = (11000 + 13000) / 2 = 12000.
  • We'll fill the missing Mileage with 87500 and the missing Price with 12000.

• Categorical Features (Engine Type): For categorical data, we can fill missing values with the mode (the most frequent value).

  • Existing Engine Types: Petrol, Diesel, Petrol, Diesel, Petrol, Hybrid.
  • Counts: Petrol (3), Diesel (2), Hybrid (1). The mode is 'Petrol'.
  • We'll fill the missing Engine Type with 'Petrol'.

Our dataset after handling missing values looks like this:

Dataset After Imputation

Mileage (km)	Age (years)	Engine Type	Condition	Price ($)
50000	3	Petrol	Good	15000
120000	8	Diesel	Fair	8000
80000	5	Petrol	Good	11000
87500	10	Diesel	Excellent	9500
65000	4	Petrol	Good	13000
150000	12	Petrol	Fair	12000
95000	6	Hybrid	Excellent	18000

Note: The Price column is often our target variable (what we want to predict). Depending on the situation, if the target value is missing, we might choose to simply remove that row entirely, as imputing the target could distort the model's learning. For this example, we've imputed it to demonstrate the technique, but keep this consideration in mind.

Step 2: Encoding Categorical Features

Machine learning algorithms work with numbers. We need to convert our categorical features (Engine Type, Condition) into a numerical format. One common method is One-Hot Encoding.

• Engine Type: This has categories 'Petrol', 'Diesel', 'Hybrid'. One-Hot Encoding creates a new binary (0 or 1) column for each category.

• Condition: This has categories 'Good', 'Fair', 'Excellent'. Again, we create a new binary column for each. Note that 'Condition' might have an inherent order (e.g., Fair < Good < Excellent). While One-Hot Encoding ignores this order, other encoding methods (like Ordinal Encoding) could preserve it if deemed important for the model. For simplicity here, we'll use One-Hot Encoding.

Applying One-Hot Encoding transforms the dataset structure:

Dataset After One-Hot Encoding

Mileage (km)	Age (years)	Engine_Petrol	Engine_Diesel	Engine_Hybrid	Cond_Good	Cond_Fair	Cond_Excellent	Price ($)
50000	3	1	0	0	1	0	0	15000
120000	8	0	1	0	0	1	0	8000
80000	5	1	0	0	1	0	0	11000
87500	10	0	1	0	0	0	1	9500
65000	4	1	0	0	1	0	0	13000
150000	12	1	0	0	0	1	0	12000
95000	6	0	0	1	0	0	1	18000

Now, all our feature columns are numerical.

Step 3: Feature Scaling

Look at the Mileage and Age columns. Mileage ranges from 50,000 to 150,000, while Age ranges from 3 to 12. Some algorithms are sensitive to such large differences in scale. Let's apply scaling to the numerical features (Mileage, Age). We'll use Normalization (Min-Max Scaling) which scales data to a range between 0 and 1, using the formula:

$$x' = \frac{x - \min(x)}{\max(x) - \min(x)}$$

Let's apply this to Mileage and Age:

• Mileage: Min=50000, Max=150000. Range = 100000.
  • 50000 -> (50000 - 50000) / 100000 = 0.0
  • 120000 -> (120000 - 50000) / 100000 = 0.7
  • 80000 -> (80000 - 50000) / 100000 = 0.3
  • 87500 -> (87500 - 50000) / 100000 = 0.375
  • 65000 -> (65000 - 50000) / 100000 = 0.15
  • 150000 -> (150000 - 50000) / 100000 = 1.0
  • 95000 -> (95000 - 50000) / 100000 = 0.45
• Age: Min=3, Max=12. Range = 9.
  • 3 -> (3 - 3) / 9 = 0.0
  • 8 -> (8 - 3) / 9 ≈ 0.56
  • 5 -> (5 - 3) / 9 ≈ 0.22
  • 10 -> (10 - 3) / 9 ≈ 0.78
  • 4 -> (4 - 3) / 9 ≈ 0.11
  • 12 -> (12 - 3) / 9 = 1.0
  • 6 -> (6 - 3) / 9 ≈ 0.33

The Price column is our target variable. Whether you scale the target variable depends on the specific model and context. Often, it's left unscaled. We will leave it unscaled here. The one-hot encoded features are already binary (0 or 1), so they don't typically require scaling.

Our final, preprocessed feature set (excluding the target Price) looks like this:

Dataset After Scaling (Features Only)

Mileage (Normalized)	Age (Normalized)	Engine_Petrol	Engine_Diesel	Engine_Hybrid	Cond_Good	Cond_Fair	Cond_Excellent
0.0	0.0	1	0	0	1	0	0
0.7	0.56	0	1	0	0	1	0
0.3	0.22	1	0	0	1	0	0
0.375	0.78	0	1	0	0	0	1
0.15	0.11	1	0	0	1	0	0
1.0	1.0	1	0	0	0	1	0
0.45	0.33	0	0	1	0	0	1

Let's visualize the effect of scaling on Mileage and Age.

Comparison of 'Mileage' vs 'Age' before (left, blue) and after (right, orange) Normalization. Notice how the scaled features now occupy a space between 0 and 1 on both axes.

Step 4: Splitting Data (The Next Step)

Remember from the previous section and Chapter 2, the final step before training is typically splitting the data into a training set and a testing set. We would take our fully cleaned and prepared data (the features table above and the corresponding Price column) and divide it, usually reserving about 70-80% for training the model and 20-30% for testing its performance on unseen data.

We won't perform the split here, but it's the immediate next action before feeding the data into an algorithm like Linear Regression or KNN.

Summary

In this hands-on walkthrough, we took a raw dataset with common issues and applied basic cleaning steps:

1. Handled Missing Values: Used median imputation for numerical features and mode imputation for categorical features.
2. Encoded Categorical Data: Transformed 'Engine Type' and 'Condition' into numerical representations using One-Hot Encoding.
3. Scaled Numerical Features: Applied Normalization (Min-Max Scaling) to 'Mileage' and 'Age' to bring them into a common range (0 to 1).

This processed data is now in a much better format for most machine learning algorithms. While these are basic techniques, they form the foundation of data preparation, a significant part of any machine learning project.