Selecting Columns and Rows

Access a single column by name using bracket notation df['col'], which returns a Series. You can also use dot notation df.col when the column name is a valid Python identifier with no spaces. Bracket notation is always safe; dot notation fails for names that clash with DataFrame methods like count or mean.

To select multiple columns, pass a list of column names inside the brackets: df[['col1', 'col2']]. Note the double brackets — the outer pair is the indexing operator, the inner pair creates a Python list. The result is a DataFrame, not a Series. This is commonly used to extract a feature matrix for machine learning.

df.loc[row_label, col_label] selects data by label. Provide a single label, a list of labels, or a slice for both rows and columns. df.loc[:, 'score'] selects all rows of the 'score' column. df.loc['r1':'r3', ['a', 'b']] selects rows r1 to r3 (inclusive) for columns a and b.

df.iloc[row_pos, col_pos] selects by integer position ignoring labels. Both arguments follow Python slice conventions with exclusive stops. df.iloc[:, 0] selects the first column as a Series. df.iloc[0:2, 1:3] selects a rectangular 2×2 sub-DataFrame from the top-left.

Pass a boolean Series (with the same index as the DataFrame) to .loc to filter rows. Build the boolean Series from a column condition. You can combine conditions with & and |. This is the standard filtering pattern in Pandas data analysis and is far more expressive than SQL WHERE clauses for complex multi-column conditions.

Multiple conditions must each be in parentheses and combined with & (AND) or | (OR). Using Python's and/or keywords on Series raises an error. To negate a condition, use ~ (tilde) instead of not. Always test conditions individually before combining to isolate the source of unexpected results.

df.iloc[n] returns the n-th row as a Series. df.iloc[n:m] returns rows n through m-1 as a DataFrame. df.iloc[[0, 2, 4]] selects specific rows by position using fancy indexing. These are equivalent to NumPy array row selection and are often used in train/test splitting before applying machine learning models.

Both .loc and .iloc accept two arguments: df.loc[row_selector, col_selector]. This enables precise rectangular selection in a single expression. A colon : alone selects all rows or all columns. This pattern is essential for extracting a feature matrix with specific columns for only the training rows of a dataset.

When you pass a single scalar label to df.loc[label], the result is a Series where the index is the column names. This is useful for inspecting a specific record. To get a single-row DataFrame instead, wrap the label in a list: df.loc[[label]]. The distinction matters when passing results to functions expecting a DataFrame.

df.at[row_label, col_name] and df.iat[row_pos, col_pos] retrieve or set a single scalar value with lower overhead than .loc/.iloc. They are designed for loops that update individual cells. Always prefer vectorized operations over cell-by-cell updates in production code, but use at/iat when you genuinely need to iterate.

Chained indexing like df['col'][condition] or df[mask]['col'] = val can produce a SettingWithCopyWarning because Pandas may operate on a copy instead of the original. Always use a single .loc expression for any modification: df.loc[mask, 'col'] = val. This is the correct, warning-free pattern.

Test your understanding of selecting columns and rows from this lesson.

In this lesson you learned: .loc selects rows and columns by label with inclusive slice stops, .iloc selects by integer position with exclusive stops like Python slices, and boolean conditions applied through .loc filter rows without the SettingWithCopyWarning pitfall of chained indexing. Next up we add new computed columns, rename existing ones, and remove unwanted columns with drop().