The mean (average) is the sum of all values divided by how many there are. It's the single most-used summary statistic in ML — and also one of the easiest to misuse, because it hides how the data is actually spread out.
What the Mean Actually Represents
Geometrically, the mean is the balance point of your data — if you placed every data point as a weight on a number line, the mean is exactly where you'd need to place a fulcrum for the line to balance perfectly. This is a more useful mental model than "just add and divide," because it explains, intuitively, why one extreme value can drag the mean so far.
Formula — Population Mean vs Sample Mean
\(x_i\) is the \(i\)-th value, \(N\) is the size of the entire population, \(n\) is the size of the sample you actually have. The formulas are mechanically identical — the distinction matters because in almost every real ML project, you only ever have a sample (your training data) and are using \(\bar{x}\) as an estimate of the true, unknown \(\mu\) of the underlying process generating that data.
Graphical Intuition — The Balance Point
The mean is the point where the data "balances" — it factors in every value's exact position, not just rank.
Step-by-Step Numerical Example
Seven students' exam scores: \(62, 68, 68, 72, 75, 80, 85\).
| Step | Calculation |
|---|---|
| 1. Sum all values | \(62+68+68+72+75+80+85 = 510\) |
| 2. Count the values | \(n = 7\) |
| 3. Divide | \(\bar{x} = 510 / 7 \approx 72.86\) |
The average score is about 72.86 — notice this isn't any actual student's score, which is completely normal for a mean.
Second Example — Where the Mean Gets Misleading
Five people's monthly incomes (in rupees): \(30000, 32000, 35000, 40000, 500000\).
The mean, ₹127,400, is higher than four out of five actual incomes in the dataset — a single high earner pulled the average far above what's "typical." This is exactly why the mean alone can be a misleading summary for skewed data, and why it's almost always reported alongside the median in practice.
Python Implementation
# From scratch
scores = [62, 68, 68, 72, 75, 80, 85]
mean_from_scratch = sum(scores) / len(scores)
print(mean_from_scratch) # 72.857142857...
# NumPy / Pandas — what you'd actually use
import numpy as np
import pandas as pd
print(np.mean(scores)) # 72.857142857...
print(pd.Series(scores).mean()) # 72.857142857...
# On a full DataFrame, per column
df = pd.DataFrame({"score": scores, "income": [30000, 32000, 35000, 40000, 500000, 45000, 38000]})
print(df.mean(numeric_only=True))Mean vs Median vs Mode — At a Glance
| Mean | Median | Mode | |
|---|---|---|---|
| Uses every value's exact position? | Yes | No (only rank) | No (only frequency) |
| Sensitive to outliers? | Very | Barely | No |
| Works on categorical data? | No | No (needs order) | Yes |
| Best summary for | Symmetric, outlier-free numeric data | Skewed numeric data | Categorical / discrete data |
Why This Matters for ML
- Mean imputation is one of the most common ways to fill missing numeric values
- Standardization centers every feature around its mean (subtracting \(\mu\)) before scaling by \(\sigma\)
- A model's average error across all predictions (like MSE or MAE) is literally a mean, applied to per-sample errors instead of raw data values
- Baseline models often start by predicting the mean for every input — a useful "can my real model beat just guessing the average?" sanity check
Advantages
- Uses every data point, not just rank or frequency, so it reflects the full dataset's magnitude
- Has convenient mathematical properties (e.g. it's what minimizes total squared distance to every point) that make it useful inside optimization and loss functions
Limitations
- Extremely sensitive to outliers — a single extreme value can shift it dramatically, as the income example shows
- Not meaningful for purely categorical data (you can't "average" city names)
- Can misrepresent a skewed distribution as if it were centered, when most of the actual data sits elsewhere
Common Mistakes
- Reporting only the mean for a skewed feature (like income or house prices) without checking the median alongside it.
- Mean-imputing missing values in a skewed or outlier-heavy column, silently injecting unrealistic values — see Missing Value Imputation.
- Computing the mean of a categorical column that's been label-encoded as if the resulting number were meaningful — it usually isn't.
- Confusing population mean (\(\mu\)) and sample mean (\(\bar{x}\)) notation in a way that misleads about whether you're describing your whole dataset or estimating a broader truth from a sample.
Interview Relevance
Q: "When would you prefer the median over the mean?" When the data is skewed or contains outliers — the mean gets pulled toward extreme values while the median stays representative of a "typical" observation, as shown directly by the income example above.
Q: "Why is mean imputation risky for a skewed feature?" Because the mean of a skewed feature isn't representative of most actual values — filling missing entries with it systematically biases the distribution toward the extreme that's pulling the mean, distorting whatever pattern a model would otherwise learn.
Practice Question
Given the dataset \([12, 15, 14, 10, 90]\), compute the mean by hand. Then explain, without recomputing, why the median would likely give a more "typical" summary of this data.
