lorentzenchr/notebooks

notebooks's Issues

Fix colsample_bynode

In the new blog post with code

• https://github.com/lorentzenchr/notebooks/blob/master/blogposts/2021-11-13%20random%20forests%20monotone.ipynb
• https://github.com/lorentzenchr/notebooks/blob/master/blogposts/2021-11-13%20random%20forests%20monotone.r

the formula for colsample_bynode might need correction.
Python int(np.sqrt(m)) / m -> np.sqrt(m) // m
R floor(sqrt(m)) / m -> floor(sqrt(m) / m)

@mayer79 ping

Random forest feature subsampling error in charts

Thank your for your interesting review. The code in seaborn seems buggy (some dataset's names are missing) and the visualisation seems poor to me. I'd prefer using ggplot in R like this:

df <- arrow::read_parquet('./data/random_forests_max_features.parquet')
library(tidyverse)
# fill na with zero
df[is.na(df)] <- 0
df %>%
    mutate(max_features = factor(max_features, levels = c("p/3", "sqrt(p)", "0.9p", "p"))) %>%
    ggplot( aes(x=max_features, y=mse, fill=max_features)) + 
    geom_bar(stat = "identity")+
    geom_errorbar(aes(ymin=mse-mse_std, ymax=mse+mse_std), width=.2) +
    scale_fill_brewer(palette = "Set1")+
    facet_wrap(facets = vars(dataset), scales = 'free_y')+
    theme(legend.position="none")

df %>%
    mutate(max_features = factor(max_features, levels = c("p/3", "sqrt(p)", "0.9p", "p"))) %>%
    ggplot( aes(x=max_features, y=fit_time, fill=max_features)) + 
    geom_bar(stat = "identity")+
    scale_fill_brewer(palette = "Set1")+
    facet_wrap(facets = vars(dataset), scales = 'free_y')+
    theme(legend.position="none")

I usually use python and I'm not very confident in R but I love ggplot (unfortunately altair doesn't catch up yet even though it's better than seaborn - I hate that "ci" in seaborn does not accept a field as value - but I find it sometimes tricky). Give it a try:

import altair as alt
d = {}
for i,data in enumerate(df.dataset.unique()):
    tmpdf = df[df['dataset']==data]
    
    # the base chart
    base = alt.Chart(tmpdf).transform_calculate(
    ymin="datum.mse-datum.mse_std",
    ymax="datum.mse+datum.mse_std"
    )
    
    # Add BarChart
    barchart = base.mark_bar().encode(
    x=alt.X(field='max_features', type='nominal',
           sort=["p/3", "sqrt(p)", "0.9p", "p"], title=data),
    y=alt.Y('mse:Q', title='mse'),
    color=alt.Color('max_features:N',legend=None, scale=alt.Scale(scheme='set1'))    
    ).properties(width=200)
    
    # Add errorbars
    errorbars = base.mark_errorbar().encode(
    x=alt.X(field='max_features', type='nominal',
           sort=["p/3", "sqrt(p)", "0.9p", "p"]),
    y=alt.Y("ymin:Q", title="mse"),
    y2="ymax:Q"
    )
        
    # Assign value
    d[f"bars_{i}"]  = barchart + errorbars

# Concat and display  on 3 columns
alt.concat(
    *d.values(),
columns=3
)

There are some weird values like in Brazilian_houses where the std is huge and a couple of missing values (colleges is NA for p).