Why Data Science is a Science

Why Data Science is a Science
Dr. Christoforos Anagnostopoulos
Founder and Chief Data Scientist, Mentat Innovations
Lecturer in Statistics (on leave), Imperial College London
Mentat Innovations

Credentials
BA Mathematics at Cambridge University
MSc Machine Learning at Edinburgh University
MSc Logic and Computer Science at Athens University
PhD in Machine Learning for Data Streams at Imperial
Postdoc Fellow at Statistical Laboratory, Cambridge Uni.
Lecturer in Statistics at Imperial College
Founder and Chief Scientist of Mentat Innovations

Credentials
PhD in Machine Learning for Data Streams at Imperial
Postdoc Fellow at Statistical Laboratory, Cambridge Uni.
Lecturer in Statistics at Imperial College
Founder and Chief Scientist of Mentat Innovations
Numerous consulting projects in real-time data analysis:
• social media analysis, sensor network telemetry, online
RTB advertising, cybersecurity and fraud, retail banking
• engaged with data journalism on several occasions
(The Independent, The Guardian, BBC, …)
Mentat Innovations is pioneering real-time anomaly
detection on network, application and telemetry data

This talk
This talk has been given around the world
Much of the thinking in this talk comes from colleagues
that I have had the privilege to work with over the years:
Prof. David Hand, OBE (Chairman of Advisory Board of Mentat)
Renowned statistician, twice president of Royal Statistical Society
Authority on pattern recognition and data mining for retail ﬁnance

This talk
Professor Niall Adams, Imperial College London
Machine Learning expert
Data Mining in CyberSecurity pioneer

This talk
Professor David Leslie, Lancaster University
World-wide expert in machine learning within game theory

This talk
George Cotsikis (CEO and co-Founder of Mentat)
Enterpreneur, 17 years experience in quantitative ﬁnance

Data Science: the origins
Courtesy of Cathy O’Neil and Rachel Schutt

Data Mining
Pattern Recognition
Statistical Modelling
Business
Intelligence
Many rediscoveries of data
analysis in the last 20 years
Neural Nets
Knowledge
Discovery

Data Mining
Pattern Recognition
Analytics
Business
Intelligence
Predictive
Analytics
Big Data
Search and
Information Retrieval
Neural Nets
Knowledge
Discovery

Data Mining
Pattern Recognition
Machine Learning
Analytics
Business
Intelligence
Predictive
Analytics
Big Data
Search and
Natural Language
Preocessing
Neural Nets
Deep
Learning
Knowledge
Discovery

Data Mining
Pattern Recognition
Machine Learning
Analytics
Business
Intelligence
Predictive
Analytics
Big Data
Search and
Natural Language
Preocessing
Neural Nets Deep
Learning
Learning
from
Data
Knowledge
Discovery

1970s: Peter Naur introduces “data science” as a
synonym to “computer science”

1997: Jeff Wu claims “statisticians” are “data scientists”.

2001: William Cleveland introduces data science as an
independent discipline, extending statistics.

2001: William Cleveland introduces data science as an
independent discipline, extending statistics.
2008: DJ Patil (LinkedIn) and Jeff
Hammerbacher (Facebook) describe their job
role as that of “Data Scientist”

Term became trending since 2008
38 years

What about Big Data?
Volume SQL
HDFS

Volume SQL
HDFS
Velocity
complex events processing
apache storm
apache spark streaming

Volume SQL
HDFS
Velocity
complex events processing
apache storm
apache spark streaming
Variety
structured semi-structured unstructured
social graphs, system logs,
tweets/blogs, CCTV
many variables, sampling variability
(e.g., spatiotemporal)

Volume
Velocity
Variety
Veracity
Value
Nobody wants data.
Everybody wants data-driven
reliable actionable insights.

Big Data in Science
CERN
1 Petabyte per day
10 GB per second
Astrostatistics
Biomedical
Climatology

Big Data in Science
Models guided by theory
Well formulated questions
Big Data in the Commercial World
Little to no theory
“Needle in the haystack”

Example: car loan provider
Online advertising
Saw an ad
Clicked
Browsed
Converted
Cookie Info

Online advertising
Credit scoring data
Application data submitted
Credit bureau queried
Credit scoring computed
Interest raid tailored
Loan offered

Online advertising
Credit scoring data
Behavioural data
Timely payments for 3 months
Delayed 4th payment
Delayed 5th payment

Online advertising
Credit scoring data
Behavioural data
External data
Social media data
Public info about employer
Demographic data
Macroeconomic data

Online advertising
Credit scoring data
Behavioural data
External data
Collections
Sent letter, no reply
Telephoned, non-cooperative
In-person visit

Online advertising
Credit scoring data
Behavioural data
External data
Collections
Data silos
No substantive theory
Often question is unclear (“ﬁshing”)
Data quality low
Not necessarily that Big
Variety of data

Statistical Methodology
Exploratory Data Analysis
Formulate question, get data

Model and Variable Selection
Model Fitting
Model Diagnostics

Model Fitting
Model Diagnostics
Inference Prediction

Model Fitting
Model Diagnostics
histograms
density plots
xy-plots
summary stats

Model Fitting
Model Diagnostics
histograms
density plots
xy-plots
summary stats
variable selection,
dimensionality reduction,
model averaging
(ensembles),
Cross-Validation,
bootstrapping, QQ plots,
outlier detection,…

Model Fitting
Model Diagnostics
histograms
density plots
xy-plots
summary stats
variable selection,
model averaging
(ensembles),
Cross-Validation,
classiﬁcation
regression
forecasting
X,Y,Z have an
effect on W

Model Fitting
Model Diagnostics
histograms
density plots
xy-plots
summary stats
variable selection,
model averaging
(ensembles),
Cross-Validation,
classiﬁcation
regression
forecasting
X,Y,Z have an
effect on W Anomaly /
Change Detection

Bayesian vs Classical
Classical: data are noisy, parameters are ﬁxed but unknown.
We use probability distributions to model the noise.
Bayesian: we use probability distributions to model our
uncertainty about both the data and the parameters

Bayesian vs Classical
Classical: data are noisy, parameters are ﬁxed but unknown.
We use probability distributions to model the noise.
Bayesian: we use probability distributions to model our
uncertainty about both the data and the parameters
In practice:
Bayesians “average” over their uncertainty a lot. This means
they use a lot of numerical integration (recently: Monte Carlo).
Everything has a probability distribution. Some are subjective.
Frequentists usually report “their best guess”. They use a lot of
classical optimisation (gradient descent etc.) - faster. In cases
where the variation is simple/physical, less subjective.

Data Mining and Pattern Recognition
• Focus on pattern extraction rather than inference
• Often no question formulated in advance
Machine Learning
• Focus on prediction (out-of-sample error)
• Largely more automatic, black-box techniques are OK
• Huge success stories in stylised worlds
• Onus on the user to ﬁt their problem into one of only a few
“templates” (classiﬁcation, regression) - carries big risks.
Deep Learning and Cognitive AI
• Aims to replicate human cognition, low to mid-level faculties
such as vision, hearing, natural language understanding.
• Can share methods with statistics/probabilistic modelling,
but is mostly fundamentally different in its approach.

ANALYTICS LEARNINGvs

retrospective summaries generalisation

retrospective summaries generalisation
a matter of resources to
compute the exact answer
(storage, distributed queries,
parallel computation, …)
mathematics
probability theory
numerical optimisation
logic and algorithms
no “exact” answer

Takeaways:
• Black boxes aren’t enough
• More Data != More Information
• Big Data needs Big Models
• Quantity vs Quality vs Homogeneity

Black boxes aren’t enough
Peter Norvig:
Statement largely driven by “quantum step” in machine translation
offered by black-box (neural net) techniques, compared to explicit
grammar models and classical natural language processing tools
Black-box AI is experiencing a second coming. However, it does
rely on (nearly commoditised) natural language preprocessing
tools for keyword extraction, named entity recognition etc.
 
 
Almost never true. Even if generalisation is not needed, there are
always sources of error (measurement, nonresponse), as well as
latent factors (e.g., the effect of X on Y, correlation, causality).

More Data != More Information
20 years worth of credit scoring data, but …
• Only one snapshot of each applicant’s behaviour
• Unknown levels of demographic variability
• Unknown levels of temporal variability
With more data (usually) comes more heterogeneity:
one could say that Big Data = Many Small Datasets
Databases went from ﬂat to relational to noSQL, but
most commodity models are pre-relational!
Models are not as re-usable as people think (for
example, a decision tree might be a good predictor
but a poor customer segmentation tool)

More Data != More Information
The signal sometimes simply isn’t there
Substantive theory (and common
sense) are still needed.
External (unobserved) factors,
inherent inpredictability
Biased sampling (observational vs
prospective - e.g., A/B testing). The
lost art of survey sampling (elections?)

Big Data needs Big Models
With enough data, everything is signiﬁcant
This assumes the model is right and the data i.i.d.
• Bigger data typically means more sources of variation
• Model complexity should grow with the data (Kolmogorov)
−5 0 5 10 15
−2000200400
Small Data
Attribute
Response
●
●
●
●
●
●
●
●
●
●
Truth
Complex model
Simple model
−5 0 5 10 15
−2000200400
Bigger Data
Attribute
Response
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●●
●
●
●
●
●
Truth
Complex model
Simple model

Personally a big fan of Bayesian non-parametrics.
Zoubin Ghahramani thinks it’s
“the rise of the automated statistician”

Fat Data vs Tall Data
Sometimes bigger means more features for the same
examples: curse of dimensionality. Modern techniques for
sparse learning (p >> n) are a great aid (e.g., Lasso)
ID Age Income Tweet Tweet Tweet ...
1
2
3
4
...
ID Age Income
1
2
3
4
5
6
7
8
...

Fat Data vs Tall Data
Consider recommender systems. As data grows:
• more items, more users
• each user ranks a ﬁxed number of items: sparser matrices

Temporal homogeneity: the hidden bottleneck
At one extreme, one could ignore all past data as irrelevant
At the other one could assume the future is like the past
Solutions in the middle include dynamic modelling (very
complicated and computationally expensive), and exponential
filters of various specifications (my field of expertise)
−4 −2 0 2 4
0.00.20.40.60.81.0
X
Density
Prior
Posterior
Posterior with power prior
Posterior with flat prior

Sometimes there is nothing to do
●
●●
●
●
●
●
●
●
● ●
●
● ●
●
●
●
● ●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
−4 −2 0 2 4
−4−2024
X1
X2
●
●
●
● ●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
Class 1
Class 2
●
● ●
●
●
●
●
●
●
●
●
●
●●●
● ●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
−4 −2 0 2 4
−4−2024
X1
X2
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
Class 1
Class 2

What looks like drift for one model might not be for another,
especially when the population, not the concept, is drifting
●
●
●
●
●
●
●
●● ● ●●
●
●●
●
● ●●
●
●
●
●
●
●
●
● ●
●
●
●
●
● ●
●●
● ●●●
●
●
●
●
●
●
●
●
● ●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
−3 −2 −1 0 1 2 3
−10−50510
X
y
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
old data
new data

Robustness
Important to have built-in guarantees. Robustness and model
diagnostics is the unsung hero in classical statistics.
Complicating the assumption set sometimes leads to overly
complex models. Robustness is often the expedient solution.

Do not torture the data
The Wall Street Journal:
“Big Data Unveils Some Weird Correlations”
• orange used cars are more reliable
• taller people are better at repaying loans
−4 −2 0 2 4
0.00.20.40.60.81.0
X
Density
Prior
Posterior
Posterior with power prior
Posterior with flat prior
• http://www.tylervigen.com

Streaming data
Exact answers are sometimes possible (e.g., running mean)
But sometimes they are not (e.g., top-K, median)
Streaming approximate algorithms are fast, and can be very
accurate, but they can be complicated (e.g., hyperloglog).
Keep constant memory footprint.
Keep up (do not queue)

Streaming data
However, in Machine Learning, there is no “exact” answers.
Will batch always outperform streaming (more resources)?
• Temporal heterogeneity (drift)
• Simulated annealing
• Overﬁtting (prequential learning)
www.ment.at/blog.html
Keep constant memory footprint.
Keep up (do not queue)

Infrastructure
I haven’t discussed infrastructure as much. It’s critical.
If you are late, sometimes you might as well give up.
Parallelisation (e.g., GPUs), distribution (e.g., HDFS),
streaming (e.g., Spark Streaming), λ-architectures …
Algorithms often need to be designed from scratch.
Great progress in this direction. Keep working on it!

datastream.io
additional deployment options

How to manage data scientists
Treat negative results like you treat positive results
Encourage lab reports: data analysis is a process.
Do not overﬁt. Do not ﬁsh for p-values. Do not torture the data.
Specify hypotheses in advance whenever possible. Then test.
Black box solutions are great for prediction. Only.
Do not silo data scientists. Incorporate expert knowledge
whenever possible. Explicit prior beliefs are not a bias risk.

Conclusions
• Knowledge is power. Knowledge relies on data.  
• The process of extracting knowledge from data has
become more efﬁcient and more powerful than ever –
but it’s still far from automatic (we are working on it ...)  
• Big Data needs Big Models  
• More Data != More Information  
• A Data Scientist is a team, not an individual

Afterthought
What about strong Artiﬁcial Intelligence?
Machines are outperforming
humans in an increasingly broad
array of cognitive tasks.
Last time this happened we had the
Industrial Revolution.
Data Science is at the cusp of this wave. This is an
exciting time, but it also carries a lot of responsibility.

Afterthought
If machines replace us, there will only be one profession left
AI programmers and Data Scientists

Why Data Science is a Science

More Related Content

What's hot

Similar to Why Data Science is a Science

Recently uploaded

Why Data Science is a Science