week4-Discussion
Jaspereric
Lecture.pptxData Driven Decision Making Week 3 Data Analysis Data Software
1
How to choose what data analysis to do?
Data Analysis: Process
Data preparation
Exploratory Data Analysis
Plotting
Distributions
Correlations
Advanced analytics
Descriptive modeling
Categorization
Predictive Modeling
Forecasting
Recommendation Systems
Business Question/Need
Business Decision
Data collection
Never actually this linear
Any step may be your stopping point
Simple analyses are great!
Will jump in and out of it
Data Preparation
Exploratory Data Analysis
Plotting
Distributions
Correlations
Continuous data, calculated for every year, 1928-2017
Mean return, 12%
Learning Data Methods
For each method you learn consider:
What sorts of data would I apply this method to?
Types of data
E.g. seasonal and frequency analyses would be used for time series data
Data from particular industries or specialties
E.g. Markov chain modeling for financial growth modeling
E.g. Logistics and Supply chain modeling
What do I have to look out for with this analysis?
Who do I know who’s an expert in this?
Data analysis
Machine Learning Artificial Intelligence
Regression and Statistical Modeling
What do you know about regression and statistical modeling?
What is Machine Learning? Artificial Intelligence?
“Formally”
Artificial Intelligence (AI)
Coined in 1956 by John McCarthy
“machines that can perform tasks that are characteristic of human intelligence” - McCarthy
General AI – AI that has all aspects of human intelligence
Narrow AI – AI that expresses some facet(s) of human intelligence, e.g. facial recognition
Machine learning (ML)
Coined in 1959 by Arthur Samuel
“the ability to learn without being explicitly programmed” - Samuel
Saves the billions of lines of code you’d need to procedurally create AI
“Machine learning is simply a way of achieving AI” – McClelland
Narrow AI
McClelland, Calum. (Dec 4, 2017) The Difference Between Artificial Intelligence, Machine Learning, and Deep Learning. Medium.com. Retrieved from https://medium.com/iotforall/the-difference-between-artificial-intelligence-machine-learning-and-deep-learning-3aa67bff5991
What is Machine Learning? Artificial Intelligence?
In usual conversation
Artificial Intelligence = Machine learning
With a computer scientist
Depending on the era they came up in, more likely to use AI or ML, generally more likely to use ML unless talking to potential funders / think pieces / TEDTalks
Press/public thinker
More likely to use AI
Start ups
More likely to use AI
From Rob Tibshirani
http://statweb.stanford.edu/~tibs/stat315a/
What is Machine Learning?
Statistics with different terminology?
From Rob Tibshirani
http://statweb.stanford.edu/~tibs/stat315a/
What is Machine Learning?
Supervised vs. Unsupervised learning
Supervised learning – you know the labels for the data
Regression
Classification
Unsupervised learning – you don’t know the labels
(probability) density estimation
clustering
https://commons.wikimedia.org/wiki/File:Cluster-2.svg
From Rob Tibshirani
http://statweb.stanford.edu/~tibs/stat315a/
What is Machine Learning?
Supervised vs. Unsupervised learning
Supervised learning – you know the labels for the data
Regression
Classification
Unsupervised learning – you don’t know the labels
(probability) density estimation
clustering
In both cases, you’re fitting the weights/parameters of model
You may be familiar with
Using Solver in Excel
Fitting a regression line to data
Forecasting
From Rob Tibshirani
http://statweb.stanford.edu/~tibs/stat315a/
What is Machine Learning?
Once you’ve fit your model, need to see how it performs with new data.
Overfit?
Set aside a subset of your data as a “test set” to see how the model performs on data that you didn’t fit on
From Rob Tibshirani
http://statweb.stanford.edu/~tibs/stat315a/
What is Deep Learning?
What is Deep Learning?
Deep learning is a method of Machine Learning (that is hot right now)
Deep learning is a type of (Artificial) Neural Network
Primary used for Supervised Learning
Aka Regression with A LOT OF WEIGHTS
What is Deep Learning?
(Artificial) Neural Network
Lots of parameters to fit
Deep learning
ANN with many layers
More difficult to train (incl time)
Computer scientists figured out how to train ~2006
Advanced field in many difficult ML tasks, e.g. object recognition
https://commons.wikimedia.org/wiki/File:Colored_neural_network.svg
What is Deep Learning?
(Artificial) Neural Network
Lots of parameters to fit
Deep learning
ANN with many layers
More difficult to train (incl time)
Computer scientists figured out how to train ~2006
Advanced field in many difficult ML tasks, e.g. object recognition
https://commons.wikimedia.org/wiki/File:Colored_neural_network.svg
https://www.dtreg.com/solution/view/21
What is Deep Learning?
(Artificial) Neural Network
Lots of parameters to fit
Deep learning
ANN with many layers
More difficult to train (incl time)
Computer scientists figured out how to train ~2006
Advanced field in many difficult ML tasks, e.g. object recognition
https://commons.wikimedia.org/wiki/File:Colored_neural_network.svg
https://www.dtreg.com/solution/view/21
https://www.dtreg.com/solution/view/21
What is Deep Learning?
(Artificial) Neural Network
Lots of parameters to fit
Deep learning
ANN with many layers
More difficult to train (incl time)
Computer scientists figured out how to train ~2006
Advanced field in many difficult ML tasks, e.g. object recognition
https://commons.wikimedia.org/wiki/File:Colored_neural_network.svg
https://www.rsipvision.com/exploring-deep-learning/
https://www.dtreg.com/solution/view/21
Deep learning: Facial recognition example
Exploring Deep Learning & CNNs - RSIP Vision. (2015). RSIP Vision. Retrieved 13 November 2018, from https://www.rsipvision.com/exploring-deep-learning/
Each layer combines features from previous layer
ML Interpretability
ML often uses high-dimensional non-linear transformations of data
Can we build a model based on the ML results? Or is it a black box?
Why do we care?
Easier to motivate
Easier to identify biases
Statistical Diagnostics
The following is based on Randy Bartlett’s book: A Practitioner’s Guide to Business Analytics, Chapter 8, Statistical Diagnostics
Evaluating the analysis
Statistical Diagnostics measure the quality of the data analysis. They provide five broad benefits:
Detecting mistakes or weaknesses—foibles.
Measuring the accuracy of an analysis.
Measuring the reliability of an analysis.
Providing insight into interpreting the results.
Providing insight into potential improved solutions.
Uwe Hohgrawe
Diagnostics families: Tool sets and Themes
Diagnostics families: Tool sets and Themes
External numbers
Numbers outside the analysis
Test forecasting model against what happens
Diagnostics families: Tool sets and Themes
Juxtaposing Results
By Method
By Quant
By Approach
By Repetition
Diagnostics families: Tool sets and Themes
Cross-Validation
Homework video
Diagnostics families: Tool sets and Themes
Statistical tests on your model
Simulation/stress testing
Performance Measurement
Test Statistical Assumptions
Test Business Assumptions
Intervals and Regions
Diagnostics families: Tool sets and Themes
Design of Samples
Design of Experiments
Covered earlier in the lecture
Data Software
Quick overview, for more information (as you need it) see this week’s supplemental materials readings on the subject and Bartlett Chapter 11
Programming languages for Data Science
This week’s reading!
Articles put it better and more concisely than I could.
Read (optional): "Which Languages Should You Learn For Data Science?" by Peter Gleeson
Overview of most commonly used programming languages in Data Science and their pros and cons
Read (optional): "What programming language should aspiring data scientists learn?" by Derrick Mwiti
Includes plots from Kaggle survey data about what programming language data scientists would recommend others learn first, what methods/tools they are interested in learning next, etc.
What language – consider:
Frameworks/libraries: Does a particular framework or library exist, which fits your needs? Do you have to use a free library, or can you spend money on a commercial product?
Performance: Are there performance restrictions which require a compiler language? Are there perhaps "soft" performance expectations by the customer?
Language-task fit: Do you have a task which can more simply or more expressively be solved in one language or another?
Language features: Do you need specific language features (e.g., regular expressions), which one language supports better than the other?
Industry standards: What languages or tools are regularly used in your company or your industry?
Last not least - personal experience. Do you need excessive ramp-up time in one language, maybe because your knowledge in one language is significantly better than in the other?
Modified from Uwe Hohgrawe
Data visualization software
Examples
Tableau
Qlikview
Can integrate with programming languages and/or data warehouse
We’ll discuss data visualizations more in Week 5
Best Coding Practices
Learn to write good code.
Even if you’re using tool like Excel or Tableau, you are writing code
Why?
Limit mistakes – at the moment and in the future
Collaboration with yourself and others
“All code has at least one collaborator and that is future you.” – Hadley Wickham
Your Github repo can be a part of your virtual resume
Basics:
Don’t hard code parameters. Especially not globals.
Keep things organized and neat
File, function, and variable names
Break problems down into smaller components
Document
Test
Use version control
More (optional) supplemental material
Better Science Code slides by Eric Denovellis
Many points from Better Science Code by Eric Denovellis
0 5
10 15 20
-0. 6
-0. 5
-0. 4
-0. 3
-0. 2
-0. 1 0 0.1 0.2 0.3 0.4 0.5 0.6
M ore
Fr eq
ue nc
y
Bin
Histogram of S&P Annual Return
Frequency
0
5
10
15
20
-
0
.
6
-
0
.
5
-
0
.
4
-
0
.
3
-
0
.
2
-
0
.
1
0
0
.
1
0
.
2
0
.
3
0
.
4
0
.
5
0
.
6
M
o
r
e
F
r
e
q
u
e
n
c
y
Bin
Histogram of S&P Annual Return
Frequency
1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 2020 2030
Pe rc
en t A
nn ua
l R et
ur n
Year
S&P 500
192019301940195019601970198019902000201020202030
P
e
r
c
e
n
t
A
n
n
u
a
l
R
e
t
u
r
n
Year
S&P 500
Glossary
Machine learning Statistics
network, graphs model
weights parameters
learning fitting
generalization test set performance
supervised learning regression/classification
unsupervised learning density estimation, clustering
large grant = $1,000,000 large grant= $50,000
nice place to have a meeting: nice place to have a meeting:
Snowbird, Utah, French Alps Las Vegas in August
1
