9. Aggregations

The axis parameter (crucial!)#

axis=0 collapses rows (result has one value per column). axis=1 collapses columns (one value per row).

m = np.array([[1, 2, 3],
              [4, 5, 6]])
m.sum()            # 21   (all elements)
m.sum(axis=0)      # [5 7 9]     (down columns)
m.sum(axis=1)      # [ 6 15]     (across rows)
m.mean(axis=0)     # [2.5 3.5 4.5]
np.percentile(m, 50)        # 3.5 (median)
np.percentile(m, [25, 75])  # [2.25 4.75]

keepdims for broadcasting-friendly results#

col_sum = m.sum(axis=1, keepdims=True)   # shape (2,1) instead of (2,)
m / col_sum                              # row-normalize (broadcasts cleanly)

📌 Remember: axis=0 → "for each column, go down the rows." Think of the axis as the one being eliminated.

NaN-aware reductions (real data has holes)#

A single NaN poisons a normal reduction (np.mean([1, np.nan]) → nan). Use the nan* variants to skip missing values:

a = np.array([1.0, 2.0, np.nan, 4.0])
np.mean(a)        # nan   ⚠️ one NaN poisons the whole result
np.nanmean(a)     # 2.333 (ignores NaN)
np.nansum(a)      # 7.0
np.nanmax(a), np.nanstd(a)          # NaN-safe extremes / spread
np.isnan(a).sum()                    # 1  — count the holes
a[~np.isnan(a)]                      # drop NaNs before other ops

⚠️ Common Mistake: np.std uses population std (ddof=0) by default; Pandas .std() uses sample std (ddof=1). Set ddof=1 in NumPy to match statistics conventions.

⭐ Interview Question: What does axis=1 mean? Aggregate along columns → collapse each row to a single value.