Method for sorting different types of bar chart?

[AndhikaWB]

Question 1

I'm trying to sort the Y axis based on a custom list (early morning, morning, etc), but it doesn't seem to work. However, when using simple ascending/descending sort on the Y axis, the sort does work. Is this a bug?

Code

c1 = []

# I can't use facet since departure and arrival are independent of each other
for col in ['departure_time', 'arrival_time']:
    time_sort = ['Early_Morning', 'Morning', 'Afternoon', 'Evening', 'Night', 'Late_Night']

    c1.append(
        alt.Chart(
            df,
            title = alt.Title(
                f'Most Common {col.title().replace('_', ' ')} by Airline',
                anchor = 'middle'
            ),
            width = 150
        ).mark_bar().encode(
            x = alt.X(col, aggregate = 'count', title = None),
            # Ascending and descending works, but custom list sort doesn't work
            # Also, the legend doesn't have the same order when using descending sort
            y = alt.Y(col, sort = time_sort),
            column = alt.Column('airline', title = None),
            color = col
        )
    )

alt.vconcat(*c1)

Question 2

Also, when using ascending/descending sort on the Y axis, the legend doesn't follow the Y axis sort (I assume this will be the same for custom sort). I already tried using order but it doesn't have any effect, and even if it has, it doesn't support custom list either (only ascending/descending).

Code & Screenshot

c1 = []

for col in ['departure_time', 'arrival_time']:
    time_sort = ['Early_Morning', 'Morning', 'Afternoon', 'Evening', 'Night', 'Late_Night']

    c1.append(
        alt.Chart(
            df,
            title = alt.Title(
                f'Most Common {col.title().replace('_', ' ')} by Airline',
                anchor = 'middle'
            ),
            width = 150
        ).mark_bar().encode(
            x = alt.X(col, aggregate = 'count', title = None),
            # Changed this line to descending instead of custom list
            y = alt.Y(col, sort = 'descending'),
            # Doesn't have any effect and doesn't accept custom list
            order = alt.Order(col, sort = 'descending'),
            column = alt.Column('airline', title = None),
            color = col
        )
    )

alt.vconcat(*c1)

TLDR:

1. How to custom sort Y axis? (SOLVED)
2. How to custom sort legends? (so it can match the Y axis, but ideally should be independent sort (SOLVED)

[AndhikaWB]

Okay, after searching more deeply, I finally got a way to custom sort the color legends by using alt.Color(scale = alt.Scale(domain = custom_list)), but the Y axis remains unsorted.

EDIT2: Using scale parameter on both alt.Y and alt.Color seems to fix the ordering issue. Still, it's very confusing, many different ways just to sort colors/values.

Code:

c1 = []

for i, col in enumerate(['departure_time', 'arrival_time']):
    time_sort = ['Early_Morning', 'Morning', 'Afternoon', 'Evening', 'Night', 'Late_Night']

    c1.append(
        alt.Chart(
            df,
            title = alt.Title(
                f'Most Common {col.title().replace('_', ' ')} by Airline',
                anchor = 'middle'
            ),
            width = 150
        ).mark_bar().encode(
            x = alt.X(col, aggregate = 'count', title = None),
            # Fix order of Y axis using scale
            y = alt.Y(col, scale = alt.Scale(domain = time_sort)),
            column = alt.Column('airline', title = None),
            # Fix order of legends using scale
            color = alt.Color(col, scale = alt.Scale(domain = time_sort))
        )
    )

alt.vconcat(*c1)

[AndhikaWB]

Okay I did some research by comparing Altair, Letsplot, and Plotly, and I must say the legend ordering "bug" also exist on Plotly, and on Letsplot I can't even order the legend because it's missing the reorder function (as_discrete isn't what I want).

Despite my little disappointment about these "bugs" or confusing documentation, Altair is still (objectively) better than those two, so I'm gonna accept the answer for now and hope it can be made more clear/improved in future release.

Thanks for creating this wonderful library.

[ChristopherDavisUCI]

I'm glad you're liking Altair!

Does this use of an ordered categorical datatype in pandas do what you were looking for more directly? I tried this on a small toy dataset (credit to chatgpt) and it seemed to work:

import pandas as pd
import altair as alt

# Small toy dataset
df = pd.DataFrame([
    # AirAsia
    ["AirAsia",  "Morning",        "Afternoon"],
    ["AirAsia",  "Morning",        "Evening"],
    ["AirAsia",  "Afternoon",      "Night"],
    ["AirAsia",  "Evening",        "Late_Night"],
    ["AirAsia",  "Early_Morning",  "Morning"],
    ["AirAsia",  "Morning",        "Afternoon"],
    ["AirAsia",  "Night",          "Morning"],

    # AirIndia
    ["AirIndia", "Morning",        "Afternoon"],
    ["AirIndia", "Afternoon",      "Evening"],
    ["AirIndia", "Evening",        "Night"],
    ["AirIndia", "Morning",        "Evening"],
    ["AirIndia", "Late_Night",     "Morning"],
    ["AirIndia", "Afternoon",      "Night"],
    ["AirIndia", "Evening",        "Late_Night"],

    # SpiceJet
    ["SpiceJet", "Early_Morning",  "Morning"],
    ["SpiceJet", "Morning",        "Afternoon"],
    ["SpiceJet", "Afternoon",      "Evening"],
    ["SpiceJet", "Evening",        "Night"],
    ["SpiceJet", "Night",          "Late_Night"],
    ["SpiceJet", "Morning",        "Evening"],
    ["SpiceJet", "Afternoon",      "Morning"],
])

df.columns = ["airline", "departure_time", "arrival_time"]

time_sort = ['Early_Morning', 'Morning', 'Afternoon', 'Evening', 'Night', 'Late_Night']

col = 'departure_time'

df[col] = pd.Categorical(
    df[col],
    categories=time_sort,
    ordered=True
)

alt.Chart(df).mark_bar().encode(
    x = alt.X(col, aggregate = 'count', title = None),
    y = alt.Y(col),
    column = alt.Column('airline', title = None),
    color = alt.Color(col)
)

[AndhikaWB]

Reply

Actually, this is a continuation of my previous discussion (#3956) where I use this dataset and Polars dataframe library. I should've clarified it, sorry. I like Polars because of it's expressive syntax, and I guess that's how I find Altair.

And to answer your question:

Does this use of an ordered categorical datatype in pandas do what you were looking for more directly?

Sadly, Polars doesn't have categorical/ordinal data type. It does have enum but it's not exactly supported by PyArrow. It resulted an error (on PyArrow side) when I tried to plot the data using enum data type (string type works fine).

---

Slightly off topic

I have made several edits to the post actually (and the title too, but forgot to change it again once I cleaned up the post). Originally, I included this table as result of my self experimentation:

Table

Chart	Item	Sort (Asc/Desc)	Sort (Custom List)
Normal bar	Axis	✅ x/y sort by value (x/y) ☑️ x/y sort by alphabet (asc/desc)	✅ x/y sort ☑️ x/y scale domain
	Color	⚠️ color sort
	Legend	✅ color sort by value (x/y) ☑️ color sort by alphabet (asc/desc)	✅ color sort ☑️ color scale domain ☑️ color legend values
Grouped bar	Bar	✅ (x/y)Offset sort by value (x/y) ☑️ (x/y)Offset sort by alphabet (asc/desc)	✅ (x/y)Offset sort ☑️ (x/y)Offset scale domain
Stacked bar	Stack	⛔ order sort by value (x/y) ✅ order sort by alphabet (asc/desc)	✅ order sort ⛔ order scale domain (workaround)

Glossary:

• ✅ = The right way
• ☑️ = Alternative way (if the right way doesn't work)
• ⚠️ = Color can be sorted, but we should think from the perspective of the legend instead of the color. Sorting the color is useless most of the times, while sorting the legend is more meaningful
• ⛔ = Will not work or not available yet

Notes/rationales:

• x/y doesn't have legend, so always use scale to sort it (if sort = [list] doesn't work)
• color has both scale and legend, so use whether you like
• (x/y)Offset is unique to grouped bar (to order the bars per group), while order is unique to stacked bar/area chart (to order the stack)
• Use -x, -y to sort by value descendingly, and descending to sort by alphabet descendingly

But decided to remove it since it doesn't really answer my initial question.

Still, I think having a cheatsheet like that on Altair's docs would be nice, and resolve_XXX too. I originally wondered what's the difference between resolve_axis(x = 'independent') and resolve_scale(x = 'independent') so I had to write my own note.