Cartwheel’s order_items table has a quantity, a unit price and a discount on each of its 3,991 lines, but no revenue. Revenue is something you calculate, and how you calculate it has to be written down where everyone can see it. That is what a computed column does: it turns a business rule into one line of code applied to every row.
Assigning to a new column name creates it:
\nimport pandas as pd\n\nitems = pd.read_csv(\"order_items.csv\")\nitems[\"gross\"] = items[\"quantity\"] * items[\"unit_price\"]\nitems[\"revenue\"] = items[\"gross\"] * (1 - items[\"discount\"])\nitems[\"discount_amount\"] = items[\"gross\"] - items[\"revenue\"]\nprint(items.head().to_string())\nprint(round(items[\"revenue\"].sum(), 2), round(items[\"discount_amount\"].sum(), 2))Each line works on whole columns, row by row, with no loop. The result: 471,940.08 dollars of revenue across all lines, after 11,658.52 dollars of discounts. (That total still includes cancelled orders; you will remove them in Section 4 once you can join order_items to orders.)
assign() does the same but returns a new DataFrame instead of changing the existing one, which lets you chain steps:
import pandas as pd\n\nitems = (\n pd.read_csv(\"order_items.csv\")\n .assign(revenue=lambda d: d[\"quantity\"] * d[\"unit_price\"] * (1 - d[\"discount\"]))\n .assign(is_discounted=lambda d: d[\"discount\"] > 0)\n)\nprint(items[[\"revenue\", \"is_discounted\"]].head(3))\nprint(items[\"is_discounted\"].sum(), \"discounted lines\")lambda d: ... is a tiny unnamed function that receives the DataFrame at that point in the chain. You need it when a step uses a column created earlier in the same chain. Both styles are fine; plain assignment is easier to read when you are starting out, and chains shine in the cleaning function you will write in Lesson 3.5.
A two-way label is a job for NumPy’s where, which pandas is built on:
import numpy as np\nimport pandas as pd\n\nitems = pd.read_csv(\"order_items.csv\")\nitems[\"revenue\"] = items[\"quantity\"] * items[\"unit_price\"] * (1 - items[\"discount\"])\nitems[\"line_size\"] = np.where(items[\"revenue\"] >= 100, \"large\", \"small\")\nprint(items[\"line_size\"].value_counts())Read it as “where revenue is at least 100, large; elsewhere, small”. For three or more categories, start with a default and overwrite subsets with .loc:
import pandas as pd\n\nitems = pd.read_csv(\"order_items.csv\")\nitems[\"discount_band\"] = \"none\"\nitems.loc[items[\"discount\"] > 0, \"discount_band\"] = \"small\"\nitems.loc[items[\"discount\"] > 0.10, \"discount_band\"] = \"big\"\nprint(items[\"discount_band\"].value_counts())Order matters exactly as it did with if/elif in Lesson 1.4, but reversed: later assignments overwrite earlier ones, so go from the broadest condition to the narrowest. Discounts of 5% and 10% end up small; 15% and 25% end up big.
Columns full of codes (mobile_app, ids, status flags) often need readable labels. map() with a dictionary translates each value:
import pandas as pd\n\norders = pd.read_csv(\"orders.csv\")\nlabels = {\"web\": \"Website\", \"mobile_app\": \"Mobile app\", \"marketplace\": \"Marketplace\"}\norders[\"channel_label\"] = orders[\"channel\"].map(labels)\nprint(orders[\"channel_label\"].value_counts())\nprint(orders[\"channel_label\"].isna().sum(), \"unmapped\")Any value missing from the dictionary becomes NaN, silently. The last line is the check that catches it: zero unmapped values means every channel had a label.
Sometimes the values are right but the type is not. products.csv stores is_active as 0 and 1, which pandas reads as integers. Summing works, but a reader of your code cannot tell a flag from a count, and filters read awkwardly. astype() converts a whole column:
import pandas as pd\n\nproducts = pd.read_csv(\"products.csv\")\nproducts[\"is_active\"] = products[\"is_active\"].astype(bool)\nprint(products[\"is_active\"].dtype, products[\"is_active\"].sum(), \"active products\")\nprint(products.loc[~products[\"is_active\"], \"name\"].tolist())Three products are discontinued, and ~products["is_active"] reads as “not active”. Use astype when every value can be converted cleanly; Section 3 shows what to do when some cannot, such as totals with dollar signs.
Here is the rule that saves you hours of confusion: to change values in a DataFrame, assign to it in a single step, with a new column or with .loc[rows, column]. Never select and then assign in two steps:
# Wrong: chained assignment. Under Copy-on-Write this never updates items.\nitems[\"discount\"][items[\"discount\"] > 0.2] = 0.2\n\n# Right: one .loc call names the rows and the column together.\nitems.loc[items[\"discount\"] > 0.2, \"discount\"] = 0.2The first form selects the discount column, which produces a new object, and then assigns into that object. Whether items itself changed used to depend on pandas internals, and pandas 2 prints a ChainedAssignmentError warning or a SettingWithCopyWarning when it spots the pattern. With Copy-on-Write (the default from pandas 3.0, available in pandas 2 through pd.options.mode.copy_on_write = True), the rule becomes simple and strict: every selection behaves like a copy, so chained assignment never updates the original. Write it the .loc way and your code behaves identically on every version.
That completes your pandas foundations: load, inspect, select, filter, sort and compute. Section 3 points these tools at the messy export, where blanks, typos, money strings and ambiguous dates are waiting.
\n","toc":[{"id":"sec-1","text":"Calculated columns","level":2},{"id":"sec-2","text":"Conditional columns","level":2},{"id":"sec-3","text":"Translating codes with map","level":2},{"id":"sec-4","text":"Fixing a column’s type with astype","level":2},{"id":"sec-5","text":"Changing values: the one rule","level":2}],"takeaways":["df["new"] = expression adds a column computed for every row at once; assign() does the same and returns a new DataFrame.","Change existing values with one .loc[row_mask, "col"] = value call, never with chained brackets like df["col"][mask] = value.","np.where(condition, a, b) builds a two-way column; .loc assignments in sequence or np.select handle more categories.","map() with a dictionary translates codes into labels, and values missing from the dictionary become NaN.","Under Copy-on-Write, every selection behaves like a copy, so the only way to change a DataFrame is to assign to it directly."],"further":[{"title":"Copy-on-Write (CoW)","url":"https://pandas.pydata.org/docs/user_guide/copy_on_write.html"},{"title":"pandas.DataFrame.assign","url":"https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.assign.html"},{"title":"numpy.where","url":"https://numpy.org/doc/stable/reference/generated/numpy.where.html"}],"quiz":[{"q":"You want to cap every discount above 0.2 at 0.2. Which line does it correctly in pandas 2 and 3?
","options":[{"text":"items["discount"][items["discount"] > 0.2] = 0.2","why":"This is chained assignment. It selects a column first and then assigns into that intermediate object; under Copy-on-Write the original table never changes."},{"text":"items[items["discount"] > 0.2]["discount"] = 0.2","why":"Also chained. The filter creates a new table, the assignment changes that temporary table, and items is left as it was."},{"text":"items.loc[items["discount"] > 0.2, "discount"] = 0.2","why":"Correct. One .loc call names the rows and the column, so pandas assigns directly into items."},{"text":"items["discount"] == 0.2","why":"This compares and returns a boolean Series; == never assigns."}],"answer":2},{"q":"orders["channel_label"] = orders["channel"].map({"web": "Website", "mobile_app": "App"}). What happens to marketplace orders?
"marketplace".","why":"map replaces every value; anything not found in the dictionary does not pass through unchanged."},{"text":"Their label becomes NaN.","why":"Correct. Values missing from the mapping dictionary become missing. Check channel_label.isna().sum() after every map."},{"text":"pandas raises a KeyError.","why":"map with a dictionary never raises for unknown values; it fills them with NaN, which is exactly why it is easy to miss."}],"answer":1},{"q":"What does np.where(items["quantity"] >= 3, "bulk", "regular") produce?
"bulk" or "regular" for every row, which you can assign as a new column","why":"Correct. np.where picks from the second argument where the condition is true and from the third elsewhere, row by row."},{"text":"The rows where quantity is at least 3","why":"Filtering is done with items[mask]. np.where with three arguments builds values, it does not drop rows."},{"text":"A single string, depending on the first row","why":"It evaluates every row, returning one label per row."}],"answer":0}],"exercise":{"kind":"python","files":["order_items.csv"],"starter":"import pandas as pd\n\nitems = pd.read_csv(\"order_items.csv\")\n\n# 1. Column \"revenue\": quantity * unit_price * (1 - discount)\n\n# 2. Column \"discount_band\": \"none\" when discount is 0,\n# \"small\" when it is above 0 and at most 0.10, \"big\" when above 0.10\n\n# 3. Column \"is_bulk\": True when quantity is 4 or more\n\n# 4. Total revenue of all lines, rounded to 2 decimals\ntotal_revenue = None\n","solution":"import pandas as pd\n\nitems = pd.read_csv(\"order_items.csv\")\n\nitems[\"revenue\"] = items[\"quantity\"] * items[\"unit_price\"] * (1 - items[\"discount\"])\n\nitems[\"discount_band\"] = \"none\"\nitems.loc[items[\"discount\"] > 0, \"discount_band\"] = \"small\"\nitems.loc[items[\"discount\"] > 0.10, \"discount_band\"] = \"big\"\n\nitems[\"is_bulk\"] = items[\"quantity\"] >= 4\n\ntotal_revenue = round(items[\"revenue\"].sum(), 2)\nprint(items[\"discount_band\"].value_counts())\n","checks":[{"id":"revenue-col","test":"'revenue' in items.columns and round(float(items.loc[3, 'revenue']), 2) == 270.9"},{"id":"revenue-discounted","test":"round(float(items.loc[items['discount'] == 0.25, 'revenue'].sum()), 2) == round(float((items['quantity'] * items['unit_price'] * 0.75)[items['discount'] == 0.25].sum()), 2)"},{"id":"bands","test":"'discount_band' in items.columns and items['discount_band'].value_counts().to_dict() == {'none': 3367, 'big': 393, 'small': 231}"},{"id":"bulk","test":"'is_bulk' in items.columns and items['is_bulk'].dtype == bool and int(items['is_bulk'].sum()) == 126"},{"id":"total","test":"round(float(total_revenue), 2) == 471940.08"},{"id":"rows-kept","test":"len(items) == 3991"}],"title":"Add revenue and labels to order lines","prompt":"Add three columns to items: revenue (quantity × unit price × (1 − discount)), discount_band ("none" when the discount is 0, "small" when it is above 0 and at most 0.10, "big" when it is above 0.10), and is_bulk (a boolean, True when quantity is 4 or more). Then set total_revenue to the sum of revenue, rounded to 2 decimals. Keep all 3,991 rows.
items["revenue"] = ..., creates the column for every row.","For the bands, set items["discount_band"] = "none" first, then overwrite with items.loc[mask, "discount_band"] = "small" and again for "big", broadest condition first.","A comparison already returns booleans: items["is_bulk"] = items["quantity"] >= 4."],"explanation":"Column arithmetic computes every line’s revenue at once, and the total, 471,940.08, matches the lesson. The bands follow the default-then-overwrite pattern: every row starts as none, discounted rows become small, and rows above 10% are overwritten to big, giving 3,367 / 231 / 393. np.select with a list of conditions would also work. is_bulk needs no np.where at all: a comparison is already a boolean column. Every change was a direct assignment to items, which is the Copy-on-Write-safe way to modify a DataFrame.
revenue exists and line 4 (index 3) is 270.90","revenue-discounted":"Discounts are applied in revenue","bands":"discount_band splits lines into none 3,367 / small 231 / big 393","bulk":"is_bulk is a boolean column with 126 bulk lines","total":"total_revenue is 471,940.08","rows-kept":"items still has all 3,991 rows"}},"words":735,"readMinutes":4}