Sampling: An an often overlooked art in exploratory data analysis

SamplingAn often overlooked art in exploratory data analysis

Eli Bressert @astrobiased Stitch Fix

exploratory data analysis

what to optimize

1

2

What we [data scientists]

do

1. obtain data 2. explore 3. do research/create data product 4. fine tune project and release 5. rinse and repeat

1. obtain data 2.explore 3. do research/create data product 4. fine tune project and release 5. rinse and repeat

basic statistics simple graphics formulate hypotheses assess best models & approaches

graphic simplicity

0etric 00 0etric 01 0etric 02 0etric 03

0etric 04 0etric 05 0etric 06 0etric 07

0etric 08 0etric 09 0etric 10 0etric 11

0etric 12 0etric 13 0etric 14 0etric 15

0etric 16 0etric 17 0etric 18 0etric 19

0etric 20 0etric 21 0etric 22 0etric 23

0etric 24 0etric 25 0etric 26 0etric 27

0etric 28 0etric 29 0etric 30 0etric 31

0etric 32 0etric 33 0etric 34 0etric 35

0etric 36 0etric 37 0etric 38

met

ric 0

0

met

ric 0

1

met

ric 0

2

met

ric 0

3

met

ric 0

4

met

ric 0

5

metric 01

metric 02

metric 03

metric 04

metric 05

metric 06

−0.4

−0.3

−0.2

−0.1

0.0

0.1

0.2

0.3

0.4

−3 −2 −1 0 1 2 3 4

−4

−3

−2

−1

0

1

2

3

Anscombe’s Quartet

10 8.048 6.95

13 7.589 8.81

11 8.3314 9.966 7.244 4.26

12 10.847 4.825 5.68

10 9.148 8.1413 8.749 8.7711 9.2614 8.16 6.134 3.112 9.137 7.265 4.74

10 7.468 6.7713 12.749 7.1111 7.8114 8.846 6.084 5.3912 8.157 6.425 5.73

8 6.588 5.768 7.718 8.848 8.478 7.048 5.2519 12.58 5.568 7.918 6.89

I II III IV

import seaborn as snsfrom scipy.optimize import curve_fit

def func(x, a, b): return a + b * x

df = sns.load_dataset(“anscombe")

df.x.mean()df.y.mean()df.x.var()df.y.var()df.x.corr(tmp.y))popt, pcov = curve_fit(func, tmp.x, tmp.y)

Mean x: 9.0Mean y: 7.5Variance x: 11.00Variance y: 4.13Correlation between x and y: 0.816Linear regression coefficients: y = 3.00 + 0.50x

http://goo.gl/Zuw4Qe

2

4

6

8

10

12

14

y

dataVet I dataVet II

2 4 6 8 10 12 14 16 18 20x

2

4

6

8

10

12

14

y

dataVet III

2 4 6 8 10 12 14 16 18 20x

dataVet IV

dataVetIIIIIIIV

EDA results will affect all that follows

processing speed

faster technology

bigger data

Boundaries

Pushing

You have two options

design your data sample

plan and execute

hit the big red button and wait for the process

to finish

attention span

?

time cost

hit red button

design and sample

explore, hypothesize, model

explore, hypothesize, model

tim

e

hit red button

design and sample

explore, hypothesize, model

explore, hypothesize, model

tim

e

fail frequently

learn fast

tried and truemodels and methods

sampling considerations

what you’re sampling priors that you can assume what operations you will run

?