Skip to content

Azure Machine Learning

Azure Machine Learning#

Table of Contents

Azure Machine Learning#

Sources#

https://github.com/MicrosoftLearning/Data-Science-and-ML-Essentials/tree/master/Labs

Requirement#
  1. Anaconda

OR

  1. Spyder (or IPython console)
  2. scikit-learn
  3. matplotlib
  4. numpy

Module 1: Intro to Data Science#

Introduction#
  • Evolving subject, no single definition
  • Requires a range of skills

Exploration and quantitative analysis of all available structured or unstructured data to develop understanding, extract knowledge, and formulate actionable results.

Data --> Decisions --> Actions

Data ---> What happened? --> Why did it happen? --> What will happen? ---> Decision#

Accidents like plane crashes etc

Areas of interest: Automatic Trading, Bidding

Steps#
  • Finding data sources
  • Acquiring data
  • Cleaning and transforming data, Reshaping (99% work)
  • Relationship finding
  • Decision
Types of Analytics#
  • Retrospective
  • Real-time
  • Predictive (Most ML falls under)
  • Prescriptive
  • Intelligent Saas apps (Cortana, ..)
Predictive vs Prescriptive#
  • Predictive analysis calibrated on past data, tells us what to expect
  • Prescriptive analysis tells what actions to take
Historical Notes#
  • Big Data by astronomers Cox & Ellsworth in 1997
  • By CCC in 2012
  • By KDD in 1996
Big Data Process#
CCC#
  • A
KDD#
  • A

Module 1#

Chapter 4: Regression#

Intro#
Simple Linear Regression#
Ridge Regression#
Support Vector Machine Regression (SVM)#
Cross-Validation#
Nested Cross-Validation#
  • Popular evalution technique in ML
  • Divide data set into 10 folds, pich one for test, reserve 1 for validation, and rest 8 as test data.

Chapter 5: Classification#

Intro#
  • Prediction of labels/predictable data - X (true/false or 1/-1) using independent variable/Feature/ - Y..
Decision Boundary#
Classification Error#
Loss Functions#

Different ML Techniques & LFs#

Logistic Regression#
SVM Regression#
AdaBoost Regression#
Decision Tree#
Boosted Decision Tree#
Imbalanced Dataset#
Minority Class Data (Excess amount, Weight)#
ROC (Receiver Operating Characteristic) Curve#
FPR & TPR (False Positive Rate & True Positive Rate)#

Chapter 6: Clustering#

Intro#
  • Unsuperwised label prediction
Unsuperwised Learning#
  • Means training data has no ground truth labels to learn from

K- Means Clustering#

  • Input K = number of clusterss
  • Randomly initialize centers
  • Assign all the points to the closest centers
  • Repeat till convergence

Hierarchical Agglomerative Clustering#

  • Start with each point in its own cluster
  • Repeatedly merge the clusters of the closest two points
Distance metrics are important#
  • Large impact on the solution
  • Some algos uses "Adaptive" distance measures

Chapter 7: Recommender Systems & Matrix Factorization#

Intro#
\left(\begin{array}{cc} 5 & * & 1 & 1\\ 5 & * & 1 & 1 \end{array}\right)
Example:#

Netflix contest

Options#
  • User-Based Collaborative Filtering
  • Item-Based Collaborative Filtering
Matrix-Factorization#
Carmen_1 \left (\begin{array}{cc} 5 & 1 \end{array}\right)

Chapter 8: Intro to Data Science Technologies#

Why Azure ML?#
  • Easy to deploy services on production
Supports?#
  • Sql
  • R
  • Python
Cortana Analytics Suite#
  • https://www.microsoft.com/cortanaanalytics
  • Preconfigures Solutions
  • Dashboard & Visualization
  • Machine Learning & Analytics
  • Azure Bigdata (Hadoop Implementation)
  • Information Management
Azure ML Studio#
  • https://account.azure.com
  • Experiments contain workflow
  • Experiments constructed of modules
  • Modules:
    • Transform Data
    • Compute Models
    • Score Models
    • Evaluate Models
  • Create custom modules with SQL, R & Python

Module 2: Working with Data#

Chapter 9#

Chapter 10#

Chapter 11: Data Sampling and Quantization#

Azure ML Table Data Types:#
  • Numeric: Integer, Floating points
  • Boolean
  • String
  • Date time
  • Time span
  • Categorial
  • Image
Continuous Vs Catergorial Variables#

Continuous: Countable, e.g. Time, Temperature, Counts* Categorial: Classifiable, e.g. Gender, Type, City

* descrete continuous

Quantization#

A range with sampled data.

What?#

Continuous variables must be sampled

Sampling?#

Digitizing the domain. * Time stamped * Precision

Example#
  • Temperature every minute
  • Count over 1 hour
Quantization of Continuous Variable#

Convert continuous variables into categorial using binning/categorizing.

Binning: Allocating each value into one category/bin.

Example: * Small, Medium & Large

Module to use: Quantize Module

Extra#

Metadata Editor

Chapter 12: Data Cleansing and Transformation#

(Data Munging)

  • Deals with
    • Missing & repeated values
    • Outliers and errors
    • Scaling
    • Filtering with custom code
  • Iterative process
  • Example: Forest-Fire Data

Missing & Repeated Values#

  • are common
  • many ML algos don't deal with missing values
  • repeated values bias results, so
    • search for them
    • make estimation
    • treat them
Clean Missing & Repeated values#
  • remove rows
  • substitute a specific value
  • Interpolate values - Linear/polynomial on the basis of growth/trend of the data
  • forward/backword fill
  • With Azure ML Module: Clean Missing Data, Remove Duplicate Rows
  • With R
    • Missing data: is.na()
    • Repeated data: duplicated()
  • With Python
    • Missing data: pandas.isnull()
    • Repeated data: DataFrame.drop_duplicates()

Errors & Outliers#

  • can bias model training, so
    • search for them
    • validate
    • treat them
Visualizing Outliers#
  • Scatter plot matrix
    • R - pairs plot
    • Python - pandas.tools.plotting.scatter_matrix
  • Bar chart or graph
  • histogram
Clean Errors & Outliers#
  • Error treatment
    • Censor: remove entire row
    • Trim: trim the value inbetween a range
    • Interpolate: Linear or polynomial on the basis of growth/trend of the data
    • Substitute
  • With Azure ML Module: Clip values (select column--> set lower/upper threshold)
  • With R
    • data.frame = data.fram[filter.expression,]
  • With Python
    • frame1 = frame1[(frame1["col1"] > 40.0) & (frame1["col2"] < 30.0) & (frame1["col3"] < 23.0)]

Scaling Data#

(aka Normalization, Transformation) * Why: * to put all the numerical data into same range line -1 to 1 or 0 to 10 other than a:0-1, b:0-100, c:500:1000 * not doing so: * will make adverse effect on training model * will get biased training model

  • What:
    • looking at numerical features/columns
    • numerical features/variable/columns needs similar scale
    • Scaling methods:
      • zero mean & unit variance
      • min-max: all numeric values in range 0 to 1
      • logrithmic: does distributional changes (good for classification)
      • LogNormal:
      • Hyperbolic tangent scaling: distribution transformation
    • ordered data like time-series may need to de-trend
    • scale after treating outliers
  • How:
    • Azure ML Module: Normalize Data
    • R:
    • Python:
  • Doubts:
    • How to make such transformations?

Module 3: Visualizing Data & EXploring Models#

Chapter 13: Data Exploration & Visualization#

Exploratory Data Analysis#

  • What:
    • Explore the data with visualization
    • Understand the relationships in the data
  • How:
    • Create multiple views of data
    • Data conditioning: Poweful plotting method to project multiple dimension on two dimension page/screen
View of data#
  • Relationships in data can be complex
  • Data exploration requires multiple views
  • Conditioned (aka faceted, trellis, lattice) plots are ideal
    • project multiple dimension onto two
    • plots of subsets (group by)
Types of plots#
  • Scatter and line plots
  • Bar:
    • like histogram but
    • Used for categorical & factor data like disease, blood grp
    • Types: ordered, un-ordered
  • Histogram:
    • used for continuos variable like time, temp
    • density or count are plotted on vertical axis
    • widely used
  • Violin
  • Q-Q
  • Box:
    • Shows 4 quartiles, i.e.
      • a box divided in two half (by median),
      • one upper vertical line, one lower and
      • dot as outliers
  • Line: connecting dot--> Polynomial regression--> curve
1


1