Is Machine Learning still a buzz word or can we easily put it to use to gain actionable insights from our data? In this demo-heavy, hands-on session, Ram will explain how to find hidden patterns in the data, find outliers and predict results using real-world datasets.
By the end of this session, you will be familiar with Machine Learning concepts (regression, classification, over-fitting, cross-validation etc.) and you should be able to build, deploy and consume Machine Learning models with ease.
About the Presenter
Based in Brisbane, Ram Katepally is a Microsoft Certified Solutions Expert and Data Analytics consultant at WARDY IT Solutions. As a consultant, Ram has significant experience working with companies of all sizes across Australia, empowering them to make data-driven business decisions. He’s passionate about Machine Learning, Internet of Things, Office 365 and Power BI. In his free time, he is an avid player of chess.
2. Agenda
Machine Learning
What is Machine Learning
Types of Machine Learning
Azure Machine Learning
Licensing, Modules, Algorithms etc.,
Building and operationalizing ML models
Questions
3. What is Machine Learning
Machine Learning is teaching computers to learn from past experiences.
Or
It is the process of making better decisions by detecting the hidden patterns and
trends within the data.
Source : kdnuggets.com
4. Types of Machine Learning
Supervised Learning
Regression – predict a continuous or discreet value (Income Prediction, Stock
Prices)
Classification – assign a category or class (Pattern recognition, Spam Filtering,
Facial Recognition, weather prediction)
Unsupervised Learning
Clustering – Grouping a set of objects such that objects in the same group
(cluster) are more similar to each other than to those in other groups
Reinforcement Learning
Feedback, games – relationship among variables
7. AML - Steps to get you started
Head to free Azure Trial Subscription
Valid for a month with $200 credit
Need a live id or any Microsoft account
Login to portal and create a sample ML Work Space
Login to ML Studio to start building experiments
8. Life Cycle of Azure ML Experiments
Source Data
Transform or Cleanse Data
Split Data for Train and Test
Pick ML Algorithm
Train Model
Score Model
Evaluate Model
9. Building ML Predictive Models
Define a business problem
Source and cleanse the data
Develop model (determine ML algorithm to use or BYO algo)
10. Demo – Build Model
Predict Titanic Survivors
Input parameters of Class, Gender, travelling with Siblings etc.,
Supervised Learning – we know past survivors for training data
Regression – Predicting a value
11. Building ML Predictive Models(cntd..)
Deploy Model
Finally consume in Applications using REST APIs in multiple
programming languages (R, python, C# etc.,)
12.
13. Consuming AML web service
AML experiments provide a REST API which is consumed by web
applications, mobile applications, custom desktop applications
and even from within Excel.
• Provides code to call web service in R, C#, and Python
• can be consumed in two ways, either as :
– request-response service (RRS) - low latency, highly scalable where application
expect a response in real-time
– batch execution service (BES) – high volume, asynchronous scoring from sources
such as blobs, SQL azure, HDInsight where the scoring is done at scheduled
intervals.
14. Consuming AML web service(cntd..)
• The web services can be called with any programming
language and from any device that satisfies three criteria:
– Has a network connection
– Has SSL capabilities to perform HTTPS requests
– Has the ability to parse JSON (by hand or support libraries)
15.
16. Demo – Building and operationalizing ML models
• Consuming the Predictive Model in Power BI
• Deploying the web service to Azure Template web app
• Log Mining
17. But…
• 2016 US Presidential Election results
– Nate Silver gave trump ~30% change for trump victory, NYT estimated at
~15%, Huffington Post at ~2% but still Trump won
– Root causes : non-representative voters sample, underestimated
republican vote in few states
• Uber surge Pricing – Sydney Lindt Café Incident
– Machine saw a spike in demand and increased price to meet demand. It
looked like it was cold profiteering from the tragedy.
– Root causes : It didn’t reason out why, may need human moderation for
extreme events
18. Conclusion - How to improve ML Results ?
• Understand the domain of data
• Be wary of overfitting and generalization in training data
– Overfitting - Justifying sample observations as possible by coming up with an overly
complex (and possibly incorrect) hypothesis.
• Be wary of outliers in the data sources
– Depending on the context, they either deserve special attention or should be completely
ignored
• Question the assumptions, examine data quality, beware of personal bias
• Repeat till satisfied
19. Azure ML Studio Components (Quick Recap)
Datasets
Data is sourced using data sets from a local file
Experiments
Modules can be dragged and dropped on the canvas
Train, Score and Evaluate models are key modules
Trained Models
Once the model is finalized it can be saved for re-use in other experiments
Web Services
Create online or batch predictions using web service
Sample code is shown for accessing end points using POST requests