Improve automatic bin determination for histograms via start, end, and step attributes

[micahtyong]

In this PR

Closes #265 and #217 by modifying Histogram.py and providing a start, end, and step bin for the Altair renderer. The bin width, step, is stored in the Vis object during PandasExecutor.py#execute_binning().

Changes

• Modifications in vislib/altair/Histogram.py and executor/PandasExecutor.py
• Additional instance variable, .__bin_size, in Vis.py

Example Output

Here is an updated screenshot for histogram binning along the x-axis. As an example,

df = pd.read_csv("https://github.com/covidvis/covid19-vis/blob/master/data/interventionFootprintByState.csv?raw=True",index_col=0)
df['dateBefore'] = pd.to_datetime(df['dateBefore'], format='%Y-%M-%d')
df

This yields the following output for Distribution histograms:

Then, we can specify intent to view Filter histograms.

df.intent = ["severityScore"]
df

Here is another screenshot to show histogram binning along the y-axis, which remains unchanged.

[micahtyong]

Hi @domoritz, thank you for your feedback so far! In my most recent commit (e.g. 9216dba), I explore the latter approach from an earlier discussion where markbar is no longer needed. Feel free to comment with your thoughts!

[domoritz]

you're very welcome.

Your latest changes definitely produce better results but require Vega-Lite to do the data transformation. It should work well for small to medium-sized data.

[domoritz]

Just to note, the last commit reverts the improvement (for potentially better performance).

[dorisjlee]

Thanks @micahtyong! There is some custom bar size determination that we had to do since Lux delegate the binning to the executor side. The new implementation looks good for Filter type actions where the xmin/xmax extent is fixed and the filter changes. However, it causes some issues when the extent range is very different, in the case of Distribution action. Our earlier implementation worked better in this case.

We could look into reimplementing something along the lines of Vega-Lite's bin.js. Thanks for pointing this out @domoritz!

[domoritz]

We could look into reimplementing something along the lines of Vega-Lite's bin.js. Thanks for pointing this out @domoritz!

As a first step, I would suggest passing the step size explicitly to Vega-Lite so that the bars are sized correctly. Then as a next step, you could improve how Lux determines the bin offset and step size.

[micahtyong]

Example Test Suite

The following are examples for reviewers to try. There are also additional tests in test_vis.py which validate bin width. To set up, be sure to include

import lux
import pandas as pd
lux.config.default_display = "lux"
lux.config.plotting_backend = "vegalite"

Olympic Dataset

df = pd.read_csv("https://github.com/lux-org/lux-datasets/blob/master/data/olympic.csv?raw=True")
df

Now, let's add a filter.

df.intent=["Height"]
df

Car Dataset

df = pd.read_csv("https://github.com/lux-org/lux-datasets/blob/master/data/car.csv?raw=True")
df['Year'] = pd.to_datetime(df['Year'], format='%Y')
df

df.intent=["Acceleration"]
df

COVID Dataset

df = pd.read_csv("https://github.com/covidvis/covid19-vis/blob/master/data/interventionFootprintByState.csv?raw=True",index_col=0)
df['dateBefore'] = pd.to_datetime(df['dateBefore'], format='%Y-%M-%d')
df

df.intent = ["severityScore"]
df

[domoritz]

The results look good to me. I left a few small comments for improvements.

[dorisjlee]

Thanks @micahtyong! The results look much better now for histograms in both Distribution and Filter.

[dorisjlee]

Thanks @micahtyong, the new histograms look great! I think the cutoff that you included for the bin width really helped with the low number of datapoints filter examples.
Special thanks @domoritz for all the helpful feedback!

[domoritz]

Thank you @micahtyong for the improvements to binning.

[micahtyong]

Thank you @domoritz for all the helpful feedback, both technical and from your HCI expertise! I really enjoyed the initiated conversations from this PR.

Thank you @dorisjlee for the feedback, review, and pointer to the executor! I learned a lot from this enhancement.