Plotting facetgrid plots in seaborn with smoothing - pandas

Question

I have a pandas dataframe a snippet of which is shown below:-
I wish to recreate the graphs shown below in Seaborn. These graphs were created in R using ggplot, but I am working with pandas/matplotlib/seaborn.
Essentially the graphs summarize the variables(mi,steps,st...) grouped by sensor id, with hours to the event on the x-axis. Additionally and most importantly, there is smoothing performed by stat_smooth() within ggplot. I have included a snippet of my ggplot code.
step.plot <- ggplot(data=cdays, aes(x=dfc, y=steps, col=legid)) +
ggtitle('time to event' +
labs(x="Days from event", y='Number of steps') +
stat_smooth(method='loess', span=0.2, formula=y~x) +
geom_vline(mapping=aes(xintercept=0), color='blue') +
theme(legend.position="none")

Answer 1

here is how I would do it. Bear in mind that I had to make assumptions about the structure of your data, so please review what I did before applying it.
Creating some simulated data
subject = np.repeat(np.repeat([1, 2, 3, 4, 5], 4), 31)
time = np.tile(np.repeat(np.arange(-15, 16, 1), 4), 5)
sensor = np.tile([1, 2, 3, 4], 31*5)
measure1 = subject*20 + time*(5-sensor) - time**2*(sensor-2)*0.1 + (time >= 0)*np.random.normal(100*(sensor-2), 10, 620) + np.random.normal(0, 10, 620)
measure2 = subject*10 + time*(2-sensor) - time**2*(sensor-4)*0.1 + (time >= 0)*np.random.normal(50*(sensor-1), 10, 620) + np.random.normal(0, 8, 620)
measure3 = time**2*(sensor-1)*0.1 + (time >= 0)*np.random.normal(50*(sensor-3), 10, 620) + np.random.normal(0, 8, 620)
measure4 = time**2*(sensor-1)*0.1 + np.random.normal(0, 8, 620)
Putting it in a long form dataset for plotting
df = pd.DataFrame(dict(subject=subject, time=time, sensor=sensor, measure1=measure1,
measure2=measure2, measure3=measure3, measure4=measure4))
df = pd.melt(df, id_vars=["sensor", "subject", "time"],
value_vars=["measure1", "measure2","measure3", "measure4"],
var_name="measure")
Creating the plot, without smoothing
g = sns.FacetGrid(data=df, col="measure", col_wrap=2)
g.map_dataframe(sns.tsplot, time="time", value="value", condition="sensor", unit="subject", color="deep")
g.add_legend(title="Sensor Number")
g.set_xlabels("Days from Event")
g.set_titles("{col_name}")
plt.show()
Plotted data, before smoothing
Now let's use statsmodels to smooth the data.
Please review this part, this is where I made assumptions about the sampling unit (I assume that the sampling unit is the subject, and therefore treat sensors and measure types as conditions).
from statsmodels.nonparametric.smoothers_lowess import lowess
dfs = []
for sens in df.sensor.unique():
for meas in df.measure.unique():
# One independent smoothing per Sensor/Measure condition.
df_filt = df.loc[(df.sensor == sens) & (df.measure == meas)]
# Frac is equivalent to span in R
filtered = lowess(df_filt.value, df_filt.time, frac=0.2)
df_filt["filteredvalue"] = filtered[:,1]
dfs.append(df_filt)
df = pd.concat(dfs)
Plotted data, after smoothing
From there you can tweak your plot however you like. Tell me if you have any question.