Handling categorical variables in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - Handling categorical variables

Which metric matters for handling categorical variables and WHY

When working with categorical variables, the main goal is to improve model accuracy and reliability. Metrics like accuracy, precision, and recall matter because they show how well the model understands categories after encoding. Poor handling of categories can cause wrong predictions, so these metrics help check if the encoding method helps the model learn correctly.

Confusion matrix example

Suppose we have a model classifying fruits based on encoded color categories. Here is a confusion matrix from predictions:

      |                 | Predicted Apple | Predicted Orange | Predicted Banana |
      |-----------------|-----------------|------------------|------------------|
      | Actual Apple    | 50 (TP)         | 5 (FP)           | 0                |
      | Actual Orange   | 3 (FN)          | 45 (TP)          | 2 (FP)           |
      | Actual Banana   | 0               | 4 (FN)           | 46 (TP)          |
    

Totals: TP=141, FP=14, FN=14. This shows how well the model predicts categories after encoding.

Tradeoff: Precision vs Recall with categorical variables

Imagine a model classifying customer types (e.g., 'new', 'returning', 'VIP').

• High precision means when the model says a customer is VIP, it is usually correct. This avoids wrongly giving VIP perks.
• High recall means the model finds most VIP customers, even if some non-VIPs are included. This avoids missing real VIPs.

Choosing encoding methods affects this tradeoff. For example, one-hot encoding might help precision but hurt recall if categories are many and sparse.

Good vs Bad metric values for categorical variable handling

Good: Accuracy above 85%, precision and recall balanced above 80% for each category. This means the model understands categories well.

Bad: Accuracy below 70%, or very low precision/recall for some categories (e.g., 40%). This suggests poor encoding or ignored categories causing confusion.

Common pitfalls in metrics when handling categorical variables

• Ignoring rare categories: Leads to poor recall on those classes.
• Data leakage: Using target information in encoding (like target mean encoding without care) inflates accuracy falsely.
• High cardinality: Too many categories can cause overfitting or sparse data, hurting generalization.
• Accuracy paradox: High accuracy but poor precision/recall on minority categories due to imbalance.

Self-check question

Your model has 98% accuracy but only 12% recall on a rare category after encoding categorical variables. Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means the model misses most instances of that rare category, which could be critical depending on the use case. High accuracy is misleading because the rare category is small compared to others.

Key Result

Balanced precision and recall after proper categorical encoding ensure the model correctly understands all categories.

Practice

1. Why do we need to convert categorical variables into numbers before using them in machine learning models?

easy

A. Because most machine learning algorithms only understand numbers, not words.

B. Because numbers take less memory than words.

C. Because categorical variables are always continuous data.

D. Because converting to numbers makes the data look nicer.

Solution

1. Step 1: Understand machine learning input requirements

   Most machine learning models work with numerical data and cannot process text or categories directly.

2. Step 2: Reason why conversion is necessary

   Converting categories to numbers allows models to perform calculations and learn patterns from the data.

3. Final Answer:

   Because most machine learning algorithms only understand numbers, not words. -> Option A

4. Quick Check:

   Machine learning needs numbers = D [OK]

Hint: Remember: ML models need numbers, not words [OK]

Common Mistakes:

• Thinking categorical variables are continuous data
• Believing conversion is just for memory saving
• Assuming conversion is for aesthetics

2. Which of the following is the correct way to apply one-hot encoding using pandas on a column named color?

easy

A. pd.encode_onehot(df.color)

B. df['color'].label_encode()

C. df.one_hot_encode('color')

D. pd.get_dummies(df['color'])

Solution

1. Step 1: Recall pandas one-hot encoding method

   Pandas provides get_dummies() to convert categorical columns into one-hot encoded columns.

2. Step 2: Match correct syntax

   The correct syntax is pd.get_dummies(df['color']). Other options are not valid pandas methods.

3. Final Answer:

   pd.get_dummies(df['color']) -> Option D

4. Quick Check:

   Use pd.get_dummies for one-hot encoding [OK]

Hint: Use pd.get_dummies for one-hot encoding in pandas [OK]

Common Mistakes:

• Using non-existent pandas methods
• Confusing label encoding with one-hot encoding
• Calling methods on dataframe incorrectly

3. Given this code snippet:

from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
df['size_encoded'] = le.fit_transform(df['size'])
print(df['size_encoded'].tolist())

If df['size'] contains ['small', 'medium', 'large', 'medium'], what will be the printed output?

medium

A. [0, 1, 2, 1]

B. [2, 1, 0, 1]

C. [1, 2, 0, 2]

D. [0, 2, 1, 2]

Solution

1. Step 1: Understand LabelEncoder mapping

   LabelEncoder assigns integers in alphabetical order: 'large'->0, 'medium'->1, 'small'->2.

2. Step 2: Map each value in the list

   Input list: ['small', 'medium', 'large', 'medium'] maps to [2, 1, 0, 1].

3. Final Answer:

   [2, 1, 0, 1] -> Option B

4. Quick Check:

   Alphabetical order encoding = A [OK]

Hint: LabelEncoder assigns numbers alphabetically [OK]

Common Mistakes:

• Assuming encoding follows input order
• Mixing up label encoding with one-hot encoding
• Forgetting zero-based indexing

4. You tried to apply one-hot encoding on a categorical column but got an error: ValueError: Input contains NaN, infinity or a value too large for dtype('float64'). What is the most likely cause and fix?

medium

A. The column has missing values; fill or drop them before encoding.

B. The column is numeric; convert it to string first.

C. The one-hot encoder only works on integers; convert categories to integers first.

D. The dataset is too large; reduce its size.

Solution

1. Step 1: Analyze the error message

   The error mentions NaN (missing) or infinite values, which cause encoding to fail.

2. Step 2: Identify fix for missing values

   Fill missing values with a placeholder or drop rows before encoding to avoid errors.

3. Final Answer:

   The column has missing values; fill or drop them before encoding. -> Option A

4. Quick Check:

   Handle missing data before encoding [OK]

Hint: Check for missing values before encoding [OK]

Common Mistakes:

• Ignoring missing values in categorical columns
• Trying to encode numeric columns as categories
• Assuming dataset size causes this error

5. You have a dataset with a categorical feature education_level having values: ['high school', 'bachelor', 'master', 'phd']. You want to use this feature in a regression model where order matters. Which encoding method should you choose and why?

hard

A. One-hot encoding, because it creates separate columns for each category.

B. Random encoding, to avoid bias in the model.

C. Label encoding, because it preserves the order of categories as numbers.

D. No encoding needed, use categories as strings directly.

Solution

1. Step 1: Understand the nature of the feature

   Education levels have a natural order: high school < bachelor < master < phd.

2. Step 2: Choose encoding that preserves order

   Label encoding assigns numbers reflecting order, suitable for regression models that consider magnitude.

3. Final Answer:

   Label encoding, because it preserves the order of categories as numbers. -> Option C

4. Quick Check:

   Ordered categories use label encoding [OK]

Hint: Use label encoding for ordered categories [OK]

Common Mistakes:

• Using one-hot encoding for ordered categories
• Ignoring order importance in encoding
• Trying to use strings directly in models