Machine Learning Exam 2 (with examples)

Data Understanding

Data Understanding

- examining key summary characteristics
- find problems (invalid values, missing values, unexpected distributions, outliers)
- visualize data

Single variable summaries

mean and standard deviation

mean

simple average of all the values

standard deviation

spread of the distribution, most often needed in the normal distribution

Descriptive statistics

organize, describe, and summarize data to help you better understand it

examples of descriptive statistics

frequency, min/max, centra tendency (mean, median,mode), dispersion or variability (range, variable, SD)

What descriptive statistics are good for

- finding unusual data
- screening the range and shape of the data
- determining the central tendency
- drawing preliminary conclusions

FREQ procedure

proceeds a one-way frequency table for each variable named in the tables statement

TABLES statement

defines the variable in a proc FREQ procedure

If the TABLES statement is omitted, a one-way frequency table is produced for every variable in the data set. This can produce a large amount of output and is seldom preferred.

FREQ procedure outputs

The default output includes frequency and percentage values, including cumulative statistics.

2 options to suppress statistics in a FREQ procedure (used in the TABLES statement)

nocum
nopercent

suppress statistics

use options in the tables statement to suppress the display of selected default statistics
(tables variable(s) / options;)

nocum

suppress the cumulative statistics

- cumulative frequency
- cumulative percent
<b>both are standard with FREQ procedure</b>

nopercent

surpasses the percentage display
<b>standard with REQ procedure</b>

BY statement

used to request separate analyses for each BY group

- the data set must be sorted or indexed by the variable(s) named in the BY statement

<b>goes in a proc freq</b>

cross-tabulation table

an asterisk between two variables generates a two-way frequency table, or cross tabulation table

3 ways to say cross-tabulation table

1. cross-tabulation table
2.contingency table
3.. two-way frequency table

cross-tabulation table output

a single table with statistics for each distinct combination of values of the selected variables.

4 standard statistics on cross-tabulation table output

1. Frequency
2. Percent
3. row pct
4. col pct

Frequency

the numbers that meet the criteria for the given box

Percent

the percentage of the total that meet the criteria for the given box

row pct

the percentage of the row that meet the criteria for the given box

col pct

the percentage of the column that meet the criteria for the given box

4 options to suppress statistics in a cross-tabulation table

norow
nocol
nofreq
nopercent

norow

suppresses the display of the row percentage

nocol

suppresses the display of the column percentage.

nofreq

suppresses the frequency display

nopercent

suppresses the percentage display

2 options to change the look of the cross-tabulation table output

1. LIST
2. CROSSLIST

<b>Included in TABLES statement</b>

LIST

Makes the output look like a list

CROSSLIST

Displays the crosstabulation results in a segregated form

FREQ procedure with multiple variables (in TABLES statement) but no asterisks (*)

lists all discrete values for a variable and reports missing values

FREQ procedure with multiple variables (in TABLES statement) but no asterisks (*) output

Like single proc FREQ but doubled

order = freq

displays the results in descending frequency order

<b>included in proc FREQ statement</b>

if-then statement

executes a SAS segment for observations that meet a specific condition

- defines a condition
- statement can be any executable SAS statement
- if expression is true, then statement executes

Fixing the data

must fix all problems in the data set and have a reason to fix them

means procedure

summary reports with descriptive statistics

PROC MEANS output

Includes descriptive stats (mean, std dev, min, max)

analysis variables

numeric variables for which statistics are to be computed

var statement

identifies then analysis variable (or variables) and their order in the output

Data Preparation

Hardest most time consuming part
Takes 60-90% of the time
Most important
How models are wrong

Variable Cleaning

- incorrect variables
- outliers (remove, transform, bin, leave them)
- missing values

3 Types of missing values

1.MCAR
2.MAR
3.MNAR

MCAR

Missing Completely At Random
- no way to determine what it should be

MAR

Missing at random
- not completely random

MNAR

missing not at random
- can probably determine what it should be
- something someone doesn't want to specify but is implied

2 ways to deal with missing values

1.Listwise delete them
2. Impute a value

listwise deletion

- delete missing values all together
- risk of biasing data

impute a value

Use a/the (constant, mean, median, distribution, calculation) to fill in missing values

feature engineering

adds to the data by creating a variable from something that already exists
- variable creation
- dummy coding
- binning (bucketing)
- calculation

dummy coding

Numeric "1" or "0" coding where each number represents an alternate response such as "female" or "male" / "yes" , "no"

Binning

the exact value doesn't matter
- histograms

normalization

converting the range of values into a standard range
- increases the speed of learning
- reduces the computing power

Standardization

- like normalization except uses the standard normal distribution

- Mean of 0 standard deviations of 1

- use for unsupervised learning models, if its close to a bell curve, if there are extreme outliers

Feature engineering drawbacks

more variables are not always better
- risk of overfitting data
- risk of use of dimensionality (need more data points)

Proc means OTHER descriptive statistics available

- Median
- Mode
- Q1
- Q3
- Range
- Qrange

median

middle value when all the values are ordered

mode

the most repeated value

Q1

the value that 25% of all values are at or below

Q3

the value that 75% of all values are at or below

range

the largest minus the smallest

qrange

the Q3 value minus the Q1 value (essentially reduces impact of outliers)

nonobs

suppresses the N Obs column

maxdec

specifies the number of decimals to display
- maxdec =1 will have one decimal place

adding statistics in the MEANS procedure

request statistics after the dataset to override the default stats (includes the OTHER descriptive statistics referred too earlier)

class statement

identifies variables whose values define subgroups for the analysis

classification variables

- character or numeric
- typically have few discrete values
- the data set does not need to be sorted or indexed by the classification

n obs

the number of observations with each unique combination of class variables

n

the number of observations with non missing values of the analysis variable(s)

skewness

- tendency of data to be more spread out on one side of the mean than the other
- asymmetry of data

left skewed

skewness < 0

right skewed

skewness > 0

Kurtosis

tendency of data to be concentrated toward the tails or middle of the data
- concentrated toward the mean

negative kurtosis (platykurtic)

less peaked
k < 0

Positive kurtosis (leptokurtic)

more peaked
k > 0

Moderate Kurtosis (mesokurtic / normal)

Middle peaked
k=0

Normal distribution

a bell-shaped curve, describing the spread of a characteristic throughout a population

1 Standard deviation

68%

2 standard deviations

95%

3 standard deviations

99%

What = normal

-mean and median are similar
- skewness is close to 0
- kurtosis is close to 0

What is used in SAS to determine normality

Proc Univariate

proc univariate

displays extreme observations, missing values and other statistics for the variable(s) included in the var statement

If the VAR statement is omitted, PROC UNIVARIATE analyzes all numeric variables in the data set

proc univariate output includes

-extreme observations section
-moments section
-tests for location section
-missing values section section
-basic statistical measures section
-Quantities section

extreme observations section

includes the five lowest and five highest values for the analysis variable and the corresponding observation numbers

moments section

includes things like skewness and kurtosis

basic statistical measures section

includes things like mean median and mode

Tests for location section

includes p value

Quantities section

includes Q1 & Q3

Missing values section

displays the number and percentage of observations with missing values fro the analysis variable

obs

is the observation number, not the count of observations with that value

ID statement

displays the value of the identifying variable (or variables) in addition tot he observation number

histogram (proc univariate)

specifies that you want a histogram, and what variables to include
(histogram salary / normal;)

inset (proc univariate)

specifies statistics you might want to include on that plot (ex. skewness, kurtosis)

test for normalcy

want a p-value > 0.05 for normal distribution

2 tests for normalcy

1. Kolomogrov-Smirnov
2.Anderson- Darling

-Shows in a goodness of fit output after running a proc univariate with a histogram or probplot

probplot

specifies what variables you want a plot for
options specify how you want the line drawn
(probplot variables </ options>;)

probplot output

line of best fit plus datapoints

one-dimension data visualization

Histograms (binning)
- creating bins for a variable
- see how the variable looks

Box plots

graphical representation of the quartile statistics