feat: churn-risk redesign

[mturoci]

Churn risk app redesign

Previous:

https://user-images.githubusercontent.com/64769322/112501274-633e5f00-8d89-11eb-852e-e37ad2e240a0.mov

New:

https://user-images.githubusercontent.com/64769322/112501331-6fc2b780-8d89-11eb-90da-96196ec4d171.mov

This PR:

• improves performance of the app by doing almost all the time-consuming tasks before app startup
• Uses flex layout instead of box - removed a few scrollbars for both analysis and code view
• Replaces image plots with native Wave plots
• cleans up the code
• formats docs markdown files
• removes unnecessary dependencies

Thanks @mtanco for DS support!

Future work:

• Replace picker with dropdown when https://github.com/h2oai/wave/pull/303 ready.
• Replace initial "empty" card with something nicer when ready. E.g. tall_info_card

Open questions:

• The last 2 plots can be divided into 4 separate ones. Wdyt?
• I noticed there are some unit/e2e tests, but not sure how usable they are. Can they be removed or are they part of some long-term goal?

[mtanco]

Visually this looks beautiful! I haven't had a chance to go through code yet, but as an idea you could do the following on the "home page" before someone chooses a Phone Number ( I don't know that it is needed, but it is common to do Global vs. Local explanations):

1. Shap Explanation - > Global Shap

   temp = contributions_df.mean(axis=0).to_numpy()
   shap = [(contributions_df.columns[i], temp[i]) for i in range(len(contributions_df.columns))]
   shap.sort(key=lambda e : e[1])
   display(shap)

2. Churn Rate - > Average Churn Prediction
3. Total Charges -> Average Total Charges
4. Total Charges Breakdown -> Average Charges Breakdown
5. and 6. Feature Contribution to Retention & Churn --> Global PDP

   col = choose min or max column from Global Shapley
   pdp = model.partial_plot(
   df, 
   cols=[col],
   plot=True,  # change to false, just for debugging
   nbins=20 if not df[col].isfactor()[0] else 1 + df[col].nlevels()[0],
   )

[mturoci]

Added a dark theme as well:

https://user-images.githubusercontent.com/64769322/112615508-da740180-8e22-11eb-840b-d9ee90c8870f.mov

[geomodular]

After adding the theme switcher, the input box stands out more. I think it's better.

[mturoci]

Thanks for the valuable review @geomodular! Comments addressed.

[mturoci]

Sounds good @mtanco! Will also add a global explanation.

[mturoci]

Comments addressed except of the last one (waiting for workaround).

[mtanco]

@mturoci I added in a check on the column's data type and made a histogram using the bins from the first plot if the data is numeric. This should fix the grouping problem caused by the original pseudo code I sent you :)

Before:

After:

[mturoci]

Thanks @mtanco!