Data Science Quiz 5

Standardization (Z-Score)

• z(x) = (x - upside down h) / select o

• transform N(upside down h, select o) → N(0,1)

• aggregate stats (mean, s.dev)

• gaussian distribution

• inverse formula: z^-1(x) = z(x) * select o + upside down h

Normalization (Min-Max Scaling)

• n(x) = (x - xmin) / (xmax - xmin)

• transform [xmin, xmax] → [0,1]

• range stats (min, max)

• uniform distribution

• inverse formula: n^-1(x) = n(x) * (xmax - xmin) + xmin

Standardization & Normalization

• reign in data variance

• commonly used in ML

• good to apply to input features

• easily invertible

• adds interpretability

How to Properly Apply Standardization or Normalization

1. split off your test data and set aside

2. calculate the necessary statistics on the train labels (either mean/s.dev or max/min)

3. apply appropriate transform

4. learn the model using transformed training labels

5. when testing, perform inference on test input data then apply inverse transform to the predictions before comparing to the test set labels

Scikit Learn Tool For Standardization

StandardScalar

Scikit Learn Tool For Normalization

MinMaxScaler

Outlier

data points that are far away from other data points

What causes outliers?

• human error

• natural disasters

Why are outliers bad?

• MSE/L2 loss heavily vulnerable

• abnormality

How to detect outliers?

1. Z-score standardization

2. interquartile range (IQR) Q3 - Q1

3. box plots

One-Hot Encoding

useful for input features that themselves are distinct classes/categories

• introduces a binary feature for each category

Caveats of One-Hot Encoding

• increases model complexity, might need regularization

• you must know all the categories you want to consider

• 
  

Feature Pipelines

allows you to wrap a sequence of transforms and models together as a single python object

Data Imputing

scikit learn’s take on handling missing/NaN values

SVM

• support vector machine

• decides the margin classifier

• margin should be fat (linear separability)

Kernel Transform

• mapping we apply to data to make it easier to learn from

• data that is not linearly separable can be converted to be

• specifically meant for SVM

Kernel Functions

• polynomial

  • takes existing input features and apply increasing exponents

• radial basis function

  • new feature is based on distance from fixed point x’

  • many variants depending on how you calculate y’

  • use L1 distance rather than L2

Save File Types

• PNG

• PDF

• JPEG

• SVG

PDF

used for placing in word docs / powerpoints

PDF

use for embedding in LaTeX documents

JPEG

file size, png is mostly better though

SVG

for specific use cases HiDPI & later editing

Matplotlib Interfaces

• MATLAB-Style

• Object-Oriented Style

MATLAB-Style Interface

references the plt object directly to create suplots

Object-Oriented Style

splits plt into fig and ax objects to create subplots

Color Considerations

• pastel-like shades over high-contrast colors

• visual comfort and fatigue

• more professional and better with printing

• colorblindness

Boilerplate

• titles for plot and axis

• plot range limits

Linestyles

• helps identify different lines in grayscale

• information is tied to the legend

• helps colorblind readers

plt.scatter() basic info

• controls individual points

• less code than plt.plot()

• not as efficient for multiple large sequences

plt.plot()

• line plots

• continuous functions

• model-over-time scenarios

• correlation plots

• receiver operating characteristic (ROC) curve

• regression prediction vs. targets

plt.scatter() Pro’s

• visualizes raw data

• unsupervised learning

• clustering

• dimension reduction

Color Maps

set of one or more colors

Heatmap

• 3D plot where X and Y are a discrete map

• made with seaborn

Seaborn

• fancier package that builds atop of matplotlib

• new plot functions, overwrites some of matplotlib

• better at implementing plot cues (such as color)

Histogram Arguments

• bins: how many vertical columns

• normed: present as probability pdf plot instead of frequency

• histype: style of bars

Heatmaps (vs. 2D Histograms)

• bins for specific ranges/categories

• 1cm unit of length

• categorical paper heatmap

• values in bins are arbitrary

  • correlation metrics

  • no restriction on type

2D Histograms (vs. Heatmaps)

• bins are arbitrarily decided

• values in bins are integers

• exception: converting to probability distribution

Legend loc

• alternative: bbox_to_anchor for numerical coordinates

• loc basically allows you to put where legend where you want it

Legend ncol

• allows you to format the legend as a row, column, or matrix

Legend frameon

• allows you to add a shadow, borderpad (control the legend border)

Subplots

• arranged in rows and columns

• ordered from 1 to n

• indexed from L → R then top → bottom

Steps to create Inset (magnifying glass)

1. add a matplotlib rectangle object to the desired region

2. create the separate inset axes object & position it as you would a legend object

3. re-do plot functions for inset plot, though it will only display inside zoomed range

3D Plotting

• adding z-value data

• changing plot() → plot3D()

• changing scatter() → scatter3D()

Annotating a plot requires what?

• code to decide font, color, location, math formatting

• trial and error to make sure the text and arrows are in the right spot

Tick Bloat

• caused by default parameters of subplots

plt.Locator()

can trim up number of ticks for the x and y axis