Why data preparation consumes most ML time in ML Python - Why Metrics Matter

Metrics & Evaluation - Why data preparation consumes most ML time

Which metric matters for this concept and WHY

When we talk about data preparation time, the key metric is time efficiency. This means how long it takes to clean, organize, and transform data before training a model. Good data preparation saves time later and improves model quality. So, measuring the time spent on data preparation helps us understand why it takes most of the project time.

Confusion matrix or equivalent visualization

Data Preparation Time Breakdown (Example):

+-------------------------+------------+
| Task                    | Time (hrs) |
+-------------------------+------------+
| Data Collection         | 5          |
| Data Cleaning           | 15         |
| Data Transformation     | 10         |
| Feature Engineering     | 8          |
| Data Validation         | 2          |
+-------------------------+------------+
| Total                   | 40         |
+-------------------------+------------+

Data preparation tasks take 40 hours, which is 80% of a 50-hour ML project.
    

Precision vs Recall (or equivalent tradeoff) with concrete examples

In data preparation, the tradeoff is between speed and quality. If you rush data cleaning, you save time but risk poor model results (low precision and recall). If you spend more time cleaning and transforming data carefully, the model learns better and predicts more accurately.

Example: For a spam filter, if data is messy, the model might wrongly mark good emails as spam (low precision). Spending more time fixing data helps avoid this mistake.

What "good" vs "bad" metric values look like for this use case

Good: Data preparation takes about 60-80% of total project time, ensuring clean, well-structured data that leads to high model accuracy and reliability.

Bad: Data preparation is rushed or skipped, taking less than 30% of time, resulting in noisy data, poor model performance, and more time fixing errors later.

Metrics pitfalls (accuracy paradox, data leakage, overfitting indicators)

• Ignoring data quality: High accuracy can be misleading if data is biased or incomplete.
• Data leakage: Poor preparation can cause information from test data to leak into training, inflating metrics falsely.
• Overfitting: Without proper feature engineering and cleaning, models may memorize noise, hurting real-world performance.
• Underestimating time: Not tracking data prep time leads to unrealistic project plans and rushed work.

Self-check: Your model has 98% accuracy but 12% recall on fraud. Is it good?

No, it is not good. The model misses most fraud cases (low recall), which is dangerous. This often happens if data preparation missed important fraud examples or features. Spending more time preparing data can improve recall and make the model safer.

Key Result

Data preparation often takes 60-80% of ML project time because clean, well-structured data is essential for good model performance.

Practice

1. Why does data preparation usually take the most time in a machine learning project?

easy

A. Because testing the model takes a long time

B. Because training the model requires many calculations

C. Because raw data is often messy and needs cleaning before use

D. Because deploying the model to production is complex

Solution

1. Step 1: Understand the nature of raw data

   Raw data collected from real sources often contains errors, missing values, and inconsistencies.

2. Step 2: Recognize the need for cleaning and organizing

   Before training, data must be cleaned and formatted properly to ensure the model learns correctly.

3. Final Answer:

   Because raw data is often messy and needs cleaning before use -> Option C

4. Quick Check:

   Messy data = More preparation time [OK]

Hint: Messy raw data means more cleaning time [OK]

Common Mistakes:

• Confusing model training time with data cleaning time
• Thinking deployment takes the most time
• Assuming testing is the longest step

2. Which of the following is the correct Python code snippet to check for missing values in a pandas DataFrame named df?

easy

A. df.isnull().sum()

B. df.missing_values()

C. df.check_na()

D. df.nulls_count()

Solution

1. Step 1: Recall pandas method for missing values

   The method isnull() returns a DataFrame of booleans indicating missing values.

2. Step 2: Summing missing values per column

   Using sum() on the boolean DataFrame counts missing values per column.

3. Final Answer:

   df.isnull().sum() -> Option A

4. Quick Check:

   pandas isnull + sum = count missing [OK]

Hint: Use df.isnull().sum() to count missing data [OK]

Common Mistakes:

• Using non-existent pandas methods
• Confusing isnull() with isna() (both work but only isnull() shown)
• Trying to call methods that don't exist on DataFrame

3. Consider this Python code for data preparation:

import pandas as pd

data = {'age': [25, None, 30], 'salary': [50000, 60000, None]}
df = pd.DataFrame(data)
df_filled = df.fillna(df.mean())
print(df_filled)

What will be the output?

medium

A. DataFrame with missing values replaced by zeros

B. DataFrame unchanged with null values still present

C. Error because df.mean() cannot be used with fillna

D. DataFrame with missing values replaced by column means

Solution

1. Step 1: Understand df.mean()

   Calculates mean of each numeric column ignoring null (NaN) values.

2. Step 2: fillna with means

   fillna replaces missing values with the corresponding column mean.

3. Final Answer:

   DataFrame with missing values replaced by column means -> Option D

4. Quick Check:

   fillna(df.mean()) fills missing with mean [OK]

Hint: fillna(df.mean()) replaces missing with average [OK]

Common Mistakes:

• Thinking fillna(df.mean()) causes error
• Assuming missing values remain unchanged
• Believing missing values become zero

4. You have this code snippet to drop rows with missing values:

import pandas as pd
data = {'height': [170, None, 160], 'weight': [70, 65, None]}
df = pd.DataFrame(data)
df_clean = df.dropna(axis=1)
print(df_clean)

What is wrong with this code if the goal is to remove rows with missing data?

medium

A. dropna() requires a threshold parameter to work

B. dropna(axis=1) removes columns, not rows

C. dropna() cannot be used on DataFrames

D. dropna(axis=1) removes rows, not columns

Solution

1. Step 1: Understand dropna axis parameter

   axis=0 drops rows; axis=1 drops columns with missing values.

2. Step 2: Identify the mistake

   Using axis=1 removes columns, but the goal is to remove rows.

3. Final Answer:

   dropna(axis=1) removes columns, not rows -> Option B

4. Quick Check:

   axis=1 means columns, not rows [OK]

Hint: Use dropna(axis=0) to remove rows with missing data [OK]

Common Mistakes:

• Confusing axis=1 with rows
• Thinking dropna always removes rows by default
• Believing dropna needs threshold parameter

5. You have a dataset with many missing values and inconsistent formats. Which two data preparation steps combined will most improve model quality and reduce errors?

hard

A. Filling missing values with mean and converting all text to lowercase

B. Randomly shuffling data and dropping duplicate rows

C. Ignoring missing values and training the model directly

D. Removing all rows with any missing data and normalizing numeric columns

Solution

1. Step 1: Consider filling missing values

   Replacing missing data with mean keeps data size and reduces bias from dropping rows.

2. Step 2: Standardize text format

   Converting text to lowercase fixes inconsistencies and helps model learn better.

3. Final Answer:

   Filling missing values with mean and converting all text to lowercase -> Option A

4. Quick Check:

   Fill missing + text clean = better data [OK]

Hint: Fill missing + clean text improves data quality [OK]

Common Mistakes:

• Dropping all rows loses too much data
• Ignoring missing values causes errors
• Shuffling or dropping duplicates alone won't fix missing data