big data conference. 2019 · build a process that uses the power of ml twice to understand and...

22
BIG DATA CONFERENCE. 2019

Upload: others

Post on 19-Oct-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

BIG DATA CONFERENCE. 2019

Page 2: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

91%...

OF CUSTOMER’S QUALITATIVE TEXT FEEDBACK IS BEING WASTED WITHIN COMPANIES

FACT #1

CUSTOMER FEEDBACK IS NOT WELL UTILIZED

BIG DATA CONFERENCE. 2019

Page 3: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

“WHAT IS DRIVING OUR COMPANY’S SALES?”

“WHY ARE CUSTOMERS LEAVING US?”

“WHAT MAKES OUR EMPLOYEES HAPPY?”

FACT #2

COMPANIES STRUGGLE WITH THE QUEST FOR CAUSALITY

BIG DATA CONFERENCE. 2019

Page 4: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

THE IDEA | 1

MAKE USE OF WHAT EVERY COMPANY HAS

BIG DATA CONFERENCE. 2019

THE NET PROMOTER SCORETM

OR ANY OTHER KPIOPEN-ENDED TEXT FEEDBACK BY

CUSTOMERS ON THEIR KPI RATING

1 2

Page 5: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

THE IDEA | 2

BUILD A PROCESS THAT USES THE POWER OF ML TWICE TO

UNDERSTAND AND PREDICT KPIS

UNSTRUCTURED

TEXT FEEDBACK

NLP: CODING,

SENTIMENT

ANALYSIS

CATEGORIES

INCL.

EVALUATION

NEURAL

NETWORKS:

IMPACT

ANALYSIS

EXPLANATION

AND

PREDICTION

OF NPS

Employees

Services

Sentiment

-0.1

0.25

-0.5

0.35

-0.11

0.72

0.3

-0.02

NPS

+28

NLP ML Causal ML

BIG DATA CONFERENCE. 2019

Page 6: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

NLP ML: SEVERAL POSSIBILITIES WERE QUALITY CHECKED

NLP ML

BIG DATA CONFERENCE. 2019

UNSUPERVISED LEARNING SOLUTIONS

MANUAL CODING

SUPERVISED LEARNING SOLUTIONS

CAPLENA/CODIT SOLUTION

Page 7: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

NLP ML: THE PREDICTIVE POWER OF SUPERVISED LEARNING IS

CLOSE TO OR EVEN BETTER THAN MANUAL CODING.

38%

53%

70%75%

Unsupervised

Learning

Open Source

Supervised Learning

Manual coding Supervised Learning

Engine Caplena

HOW CAN AN AUTOMATIC

CODING BE BETTER THAN

MANUAL?

1.) It leverages an own knowledge

database for sentiment codes.

Based on this, it generates

sentiments that are way more

predictive than manually coded

sentiments.

2.) It produces likelihood scores for

codes instead of binary codes.

NLP ML

BIG DATA CONFERENCE. 2019

Page 8: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

NLP ML: CAPLENA ML IS BASED UNSUPERVISED PRETRAINING

AND TRANSFER LEARNING – F1 SCORE TYPICALLY AROUND 70%

NLP ML

BIG DATA CONFERENCE. 2019

Page 9: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

CAUSATION, NOT CORRELATION

“There is nothing more deceptive than an obvious fact.”

Sherlock Holmes

Page 10: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

BIG DATA CONFERENCE. 2019 Careful - fake statistics taken from Social Media. Both show the same but colored picture

MAD COW DISEASE 1992 BREXIT VOTUM 2016

Page 11: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

2016 GOOGLE SEARCHES FOR ERECTILE DYSFUNCTION, HAIR

LOSS, HOW TO GET GIRLS, PENIS ENLARGEMENT, PENIS SIZE,

STEROIDS, TESTOSTERONE, AND VIAGRA

2016 TRUMP VOTERS

BIG DATA CONFERENCE. 2019

Page 12: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

0% 5% 10% 15% 20% 25% 30% 35% 40% 45%

Modern

Smart

Innovative

Can play music from any source I want (Internet music service, direct line-in from phone, Bluetooth streaming, etc.)'

It allows me to stream music to multiple speakers throughout my home

Inventive

It allows me to stream TV sound to multiple speakers throughout my home

Creative

It can be controlled from a mobile device

Popular

Is a Proud brand

It realistically reproduces TV and movie audio

Efficient

Outgoing

Energetic

It sounds good no matter where I place it in my house

It has a great looking design

Is a Trustworthy brand

Is a Hopeful brand

The music sounds unproduced and natural, as the artist would have intended'

It has a timeless design that will not look out of date in a few years

It is easy to set up and install

Friendly

It always sounds amazing, rich and crystal clear'

It has products that fit easily and anywhere in my home

It offers a range of speakers (different sizes, performance levels) to meet all my needs'

Curious

Is Calm

It creates an immersive listening experience

Approachable

Controlling a speaker or the whole system is easy

Is changing the home audio listening experience for the better

It reproduces music with all the detail of the original recording

Is a Happy brand

Is Relaxed

It provides access to a wide variety of music services and/or internet radio stations

Sensitive

Easy going

Is an Exciting brand

Is a Fun brand

It is simple to add more speakers to the system over time

It is expensive but worth it

It always has a stable wireless connection

Is a Courageous brand

It provides software updates so that the speakers get better over time

It is a good value for the money

Offers the best portable audio products

It provides excellent customer support

Offers the best home audio products

Is a brand for me

CORRELATION ANALYSIS DOESN’T HELP – AN EXAMPLE

BIG DATA CONFERENCE. 2019

Page 13: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

CAUSAL ML: INDIRECT EFFECTS NEED TO BE CONSIDERED

PURCHASE

INTENTION

Identification

with Brand

Sound

Experience

Software

updates

Multi-Room

„Best Friend“

Brand

Personality

Relax

Emotion

„Leader“

Brand

Personality

BIG DATA CONFERENCE. 2019

Page 14: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

14BIG DATA CONFERENCE. 2019

CAUSAL ML: NONLINEARITIES NEED TO BE CONSIDERED

Page 15: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

CAUSAL ML:

PROCESS OF UNIVERSAL STRUCTURE MODELING TECHNIQUE NEUSRELTM

CONNECTING

THE DOTS INCL. LV‘S

QUANTIFYING

CONNECTIONS (NN)

UNDERSTANDING

CONNECTIONS

BIG DATA CONFERENCE. 2019

Causal ML

Page 16: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

CAUSAL ML: THE ML WE USE EXPLAINS UP TO TWO TIMES BETTER

WHY CUSTOMERS ARE LOYAL OR WILLING TO RECOMMEND.

BIG DATA CONFERENCE. 2019

Causal ML

WHY IS CAUSAL ML WITH NEURAL NETWORKS SO

MUCH BETTER THAN LINEAR REGRESSION?

1.) It prevents spurious relationships and unveils

unexpected nonlinearities and interactions

2.) It considers indirect causal effects that work

through sentiments. This indirect effect adds

additional explanation power on top.

0.70

▲(indirect

effects)

Explanation

power of

traditional

linear

regression

USM’s total

explanation

power incl.

sentiment

analysis

+40%

+41%

EXPLANATION POWER R2

ON THE NPS (ON AVERAGE)

0.50

xUp to 2x

USM’s direct

explanation

power

Page 17: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

BIG DATA CONFERENCE. 2019

DASHBOARD AND TOOL TO STANDARDIZE THE PROCESS

Page 18: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

EXAMPLE

Page 19: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

SENTIMENT ANALYSIS HELPS TO INCLUDE INDIRECT AND

EMOTIONAL ASPECTS

BIG DATA CONFERENCE. 2019

Predictive Promoter Score Cockpit

Causal Graph Output of Sentiment Analysis

Page 20: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

EXAMPLE: INSURANCE COMPANY

Great consulting

Great personal customer

service

Professionality

Trustworthy, Honest,

Reliable, Fair

Bad advise/ bad

information/ lack of

information

I was lied to/

Promises were not

kept

Obvious

Hidden Hidden

FREQUENCY OF MENTIONS

POSITIVE IMPACTNEGATIVE IMPACT

by

Page 21: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

REFERENCES: SOME EXAMPLES OF WHO IS USING OUR TOOL

BIG DATA CONFERENCE. 2019

Page 22: BIG DATA CONFERENCE. 2019 · build a process that uses the power of ml twice to understand and predict kpis unstructured text feedback nlp: coding, sentiment analysis categories incl

BIG DATA CONFERENCE. 2019

SUMMARY