Marketplace and Quality Assurance Presentation - Vincent Chirchir
Summary artificial intelligence in practice- part-4
1. Some Impressionistic Take away from the Book of
Bernard Marr & Matt Ward
Artificial Intelligence in Practice
( Part – 4)
( How 50 Successful companies used AI & Machine Learning to Solve problems)
Ramki
ramaddster@gmail.com
2. The Summary of this book is made in 4 parts
due to large coverage of the book .
This is Part – 4
( Read this after Part-1 , 2 & 3)
4. Using AI to Detect Fraud & Improve customer Experience
5. American Express
With 110 million AmEx cards in operation and more than one trillion
dollars in transactions processed.
American Express handles more than 25 percent of U.S. credit card
activity.
The company interacts with people on both sides of transactions –
millions of businesses and millions of buyers – giving American Express a
rich flow of data to leverage.
How American Express uses Big Data in practice
Data analytics, and specifically machine learning, is at the heart of
American Express’s decision making. Two areas where this is evident are
detecting fraud, and bringing merchants and customers closer together.
Credit card fraud detection and prevention now relies heavily on machine
learning algorithms.
AmEx’s goal is to detect fraudulent transactions as quickly as possible to
minimize loss, so they employ a machine learning model that uses a
variety of data sources, including card membership information, spending
details, and merchant information to detect suspicious events, and make
a decision in milliseconds by comparing that event to a large dataset.
6. American Express
This has enabled American Express to detect more fraudulent
transactions and save millions. Using Big Data and Machine Learning
algorithms for fraud prevention has now become commonplace in the
industry.
Visa also uses this technology and checks many hundreds of aspects of
any transaction in near-real-time.
According to Visa’s estimates, this approach has identified $2 billion in
potential annual incremental fraud opportunities – which the company
was able to sort out before any of it was lost.
Interestingly, American Express is increasingly moving away from
focusing on its traditional function of providing credit for consumers and
merchant services for processing transactions, and towards actually
making the connection between the consumer and the business that
wants to reach them.
For consumers, the company is using its vast data flows to develop
apps that can connect a cardholder with products or services.
7. American Express
One app looks at past purchase data and then recommends restaurants in the
area that the user is likely to enjoy.
Another, called Amex Offers, shows real-time coupons relevant to the
individual’s lifestyle and buying habits based on their physical location and
nearby businesses. And this isn’t just a benefit to cardholders that use the
app, but also hopefully an incentive for more businesses to accept American
Express.
On the merchant side, American Express is offering new online business trend
analysis and industry peer benchmarking to help companies see how they’re
doing compared to their competition. The data is anonymized, meaning any
personally identifiable data is stripped out of the transactions, but merchants
are able to see detailed trends within specific niche markets or customer
segments.
The technical details
In 2010, American Express upgraded from traditional database technology to
a Hadoop infrastructure and brought in machine learning algorithms. The
company is so serious about embracing the data side of its business that it
has now opened a tech lab in Palo Alto, California, specifically to focus on Big
Data, cloud computing and mobile infrastructure.
8. Results, Key Challenges, Learning Points & Takeaways
Real time analysing transactions using Machine Learning
algorithms leads to detection of fraudulent transactions.
Less chances of false positives occurring, reducing
inconveniences to customers- Trust building .
Machine learning models for detecting fraud need to constantly
adapt & update themselves in real time, meaning they need a
consistent flow of data to learn from .
Distributed storage and large amount of compute power are
needed to handle the amount of data that is needed to make
accurate predictions in real time.
Overall security improvement through small increases in
efficiency.
9. Using AI to Improve Medical decisions & Scientific Research
10. Elsevier
Global multimedia publishing business – offers 20000 products
for educational & professional science & healthcare
communities.
First stage digital transformation – digitization of huge amount
data published in reports & journals – 140 years of history.
AI being used to draw new insights from this data.
Two patients of same age & gender will present to their primary
healthcare practitioner with the same symptoms, and yet there
will be a huge variation in the outcome & cost of treatment they
receive.
This is due to diagnosing and treating are by different health care
staff with different levels of knowledge & experience.
Developing AI derived pathway from initial presentation to
examination to treatment procedures and prescribing of
medication, patients are more likely to get better quicker, and
reduction in cost.
11. Elsevier-Clinical Decision support
When considering how to make AI work for Elsevier and its customers, as
Its mission here is to build what are described as “clinical decision support”
systems.
These use information gathered from Elsevier’s archives, combined with
patient medical data and financial claims data, to suggest the best course
of action – known as a “pathway” – for treating specific patients.
The next step of the process is to augment these decision support systems
with machine learning and deep learning.
This should mean more accurate predictions, more efficient treatment and
ultimately better patient outcomes.
The large amounts of data increase predictability and produce more
accurate models with better inferences
You get a much more accurate understanding of individual patients by
bringing together data their clinical history, their claims data, genomic data,
etc.
Elsevier’s technology is now in use at 2,000 oncology centers at healthcare
facilities across the US, where it informs decisions on patient healthcare
every day, and the adherence rate, i.e. the degree at with doctors follow
that advice, is very high.
12. “We were able to create these neural networks of closed loops, and
we trained these predictive models against large patient databases
… so we built an application what generates a differential diagnosis
based on that model.
It gives you a weighted differential, it says, with those symptoms in a
person of this age and gender, there’s a 70% chance it’s this, a 35%
chance it’s that.”
“Neural networks” refers to a technology designed to mimic the
learning characteristics of the neurons in the human brain.
Data is passed between neurons which essentially each ask a
different question about it.
The results of all of those questions are aggregated into a single
output – representing the answer to the question that the neural
network is designed to solve.
.
Elsevier-Neural Network
13. Results, Key Challenges, Learning Points & Takeaways
The adherence rate of 85% among clinical staff to the treatment
pathways suggested by its Via oncology platform.
Elsevier amalgamates patient medical records, insurance claims
& billing data & published medical literature to predict which
treatment pathways are most likely to be effective.
Treatment can be standardized if machines are used to
determine optimal treatment paths dependent on the patient’s
details, medical history and the symptoms they present with.
Standardized treatments lead to better patient outcomes if they
can be optimized according to the data, and also help healthcare
providers to reduce overall cost.
14. Using AI to Combat the US $ 450 B Counterfeit Industry
15. Entrupy
Uses AI to combat counterfeit goods .
Provides a platform as a service to brands to reduce the revenue loss due
to counterfeits.
Focus area counterfeiting through machine learning .
Sales of counterfeit goods is close to half trillion US $ annually.
Diluting brand identity.
Eats into business of genuine resellers and wholesalers.
Entrupy developed scanning technology using machine learning and deep
learning techniques – which detects whether items are genuine- cloths ,
accessories, jewelry, electrical goods and automobile parts.
Service users can use a phone app or dedicated handheld scanner to
check their purchases.
Microscope cameras that are able to record the tiniest details of product’s
construction .
This technology is able to discern even : Super fakers” very high quality
replica that are impossible for human to differentiate from a genuine
product.
16. Technology used
Database of millions of images of products sold by the different brands-
Chanel, Dior, Burberry, Gucci, Louis Vuitton & Prada.
Specialized microscope lenses to capture micro details
The images of the genuine products are used to train convolutional neural
network algorithms to classify images based on texture.
17. Results, Key Challenges, Learning Points & Takeaways
System had a 98.5% rate of correctly identifying counterfeit
merchandize.
Technology has enabled brand guarantee and confidence to the
customers .
AI can parse image data in incredibly high detail far more
quackery than a human eye could, and determine between
counterfeit products and genuine items.
Brands are happy to help with putting this technology into the
hands of customer and resellers if it helps protect their revenue
and perceived value.
19. Experian
One of the largest consumer credit reference agencies.
Businesses, Banks and financial institutions rely on this
company
They 3.6 petabytes of data on people and their spending
habits.
Apply AI to this data to make accurate predictions.
Like all financial services, they are being rapidly changed by
waves of technological innovation sweeping through industry
– none more so than AI & Machine learning .
Mortgage – time consuming & complex. Coordinating
information between –Buyers, Sellers, Surveyors, Estate
agents, Solicitors, Underwriters, Mortgage brokers & lenders.
Buying a property is stressful process.
Work gets duplicated between agencies due to
inconsistencies in the way information is transferred.
20. Experian
AI is used for analyzing thousands of applications to determine
improvement in efficiencies , reducing the duplicates & streamlining
workflows.
System is trained to look at every data, frequency of use .
Practically impossible for human beings to carry out other than small
historic sample data.
Predictive technology for limited credit histories to obtain mortgages
or personal loans.
Assessment by lenders and comparisons of profiles to come out
with trustworthy decisions.
Machine learning is used for processing of data across WF & builds
models about what data is valuable, what is surplus at each stage of
the process.
Experian has built a platform –Analytical Sandbox- allows to
produce on demand data for insights.
Also uses Cloudera’s Enterprise platform to enable quick access to
big data.
21. Results, Key Challenges, Learning Points & Takeaways
Approval mortgage applications will be reduced to day from
weeks and months.
Data driven decisions.
Credit reference agencies are positioned to streamline
workflows across complex procedures .
AI examines every aspect of the workflow , tracks in details
across instances .
Smart businesses are learning that repackaging data,
processing it with ML & offering it as a service is a great way of
diversifying their range of services in the age of AI.
Technology not ahead of security of data.
23. Harley-Davidson
An US Manufacturer of motorcycles- Sells around 1,50,000 bikes
per year globally.
Licenses its iconic brand for use of clothing, homeware &
accessories.
When we think about effective marketing for a Harley-Davidson
dealership, the first thing that pops into our mind is probably not the
ways you can use artificial intelligence to ramp up your results. It’s a
good thing the owner of New York City’s Harley dealership, Asaf
Jacobi decided to give AI a try, because it increased the dealership
leads by 2,930% in just 3 periods.
That is a remarkable number for a start-up, but for an established
brand such as Harley-Davidson, that was extraordinary.
Although Jacobi had started researching options to boost sales at
his dealership in the off-season and came across some AI tools for
marketing and advertising, it was his chance meeting with Or Shani,
the CEO of AI firm Algorithm, which had an AI-driven marketing
platform called Albert that convinced Jacobi to give it a try.
24. Harley-Davidson-How Albert works
The first test of Albert was a weekend promotion called "48 Bikes in 48
Hours."
They sold 15 motorcycles that weekend, nearly doubling the summer
sales record of eight bikes sold in one weekend.
Albert used business logic, the KPIs available for Harley-Davidson NYC
and past campaign performance to identify unknown audiences, the
best budget allocation across digital channels and even evaluate the
performance of different word choices or colors on the creative.
Albert processed the data it had been given to figure out trending
behavior. It continued to optimize the marketing and ad performance as
new data continued to come in.
Albert executes digital ad campaigns autonomously and adjusts them
automatically based on performance.
It can figure out the audiences that are most likely to convert, compare
platforms and implement learning across platforms, and discover what
creative worked better.
AI can do this exponentially faster than humans, in virtually real-time.
25. Harley-Davidson-How Albert works
Here’s what Albert was able to do for Harley-Davidson NYC:
The dealership credits Albert with 40% of its motorcycle sales over a six-
month period.
In just three months, they had an increase of 2,930% in leads, with 50% of
those being from lookalikes, prospects with similar buying patterns and
preferences as those likely to purchase Harley-Davidsons. This insight
opened up an entirely new audience that they had previously not marketed
to.
Albert also discovered that Facebook ads converted 8.5 times higher than
other Harley ads, so they focused their advertising efforts only on platforms
that worked.
New factory optimized for more expensive workforce
Unlike many manufacturers, when you walk into the Harley-Davidson plant
in York, Pa., you will still find human workers assembling the iconic
motorcycles by hand.
Although the factory was redesigned, it still has people everywhere.
Workers operate in teams of five or six to build each bike. Since virtually
every bike is one-of-a-kind, humans are uniquely qualified to adjust on the
fly when it's necessary to create these customized bikes.
26. Results, Key Challenges, Learning Points & Takeaways
Data on customer behaviour – Predictions are more accurate.
Automated segmentation & targeting of customers can often
result in uncovering entire demographics that a business has
never considered marketing to, but in fact make great
customers.
Identify more effective channels- email, social media, display
advertising – assigning of resources where probability says
they will provide the best return.
AI boosted selling is no longer the preserve of the big tech
companies, thanks to new generation of “ as-a-service”
platforms for targeting & selling to new customers.
28. Hopper
Mobile app-based platform that uses Machine learning & uses huge
volumes of historical flight data to predict the best time to buy
flights.
Launched in 2015 & in 2017 $ 1 million worth of flights everyday .
Sales approaching one Billion a year.
Scanning price comparison sites to find the best price for holiday
flight, weekend deals.
Removed the middle man from the process.
Replaced old human travel agent with an AI travel agent.
Users tell it where they want to travel to with rough idea of date-
Hopper gives the best prices it can find.
Users also get a prediction – it will also tell if they can wait for a
better price.
It is like walking into a shop, being persuaded not to buy yet, but to
wail until prices come down.
Predictive model gives Hopper the competitive advantage.
29. Hopper-Technology & Tools
Built and trained predictive algorithms using data in from global
distribution operators.
This data was generally considered less valuable, it was able to
negotiate a good price.
Prices likely to change based on flow of demand.
Hopper augmented the data with information about customers-
Close proximity to more than one airport and potential saving if they
fly out of different airports.
Alternate destinations to planned destinations.
For e.g. – Someone searching for flights to Rome, do they really
just want to visit Italy ?
Suggestion thrown up for Milan or Naples mixed in with their results
30. Results, Key Challenges, Learning Points & Takeaways
Fourth largest downloaded travel app after Uber, Lyft and
Airbnb with over 20 million users.
Able to predict the cheapest time for its users to buy flights
anywhere in the world with 95% accuracy – average saving of
US $ 50 every flight.
20% of the $ 500 million it had taken bookings came from
selling flights that customers had not even searched for directly.
User friendly search criteria
AI can replace many of the old-fashioned “ middle men” roles –
travel agents, doing the same work at much larger scale and
reduced costs.
Machine learning prediction can accurately find cheaper flights
& reduce the stress and fear of missing out inherent in price
comparison site searches.
32. Infervision
Chinese computer vision specialist has applied the technology to potentially
save millions of lives from life threatening diseases.
Use AI to interpret visual data.
Image recognition technology (the same sort of thing used by Facebook for
facial recognition or Google in image searches) is one of those tasks that’s
ideally suited to AI, particularly Deep learning .
Now, the technology is quickly progressing to a point where loftier ambitions,
like saving lives, are being realized.
Lung cancer is the leading cause of death in China, claiming the lives of more
than 600,000 people every year, largely because of air pollution.
With lung cancer, CT scan images are examined by radiologists to spot signs of
cancer as early as possible. However, in a country with a real dearth of doctors,
especially qualified radiologists, this can mean radiologists wading through
hundreds of scans each day.
This is time-consuming, laborious, and, frankly, quite tedious. Simple human
error, often caused by fatigue, means mistakes being made and important
diagnoses being missed.
This problem inspired Chen Kuan, founder of medical image diagnostics
startup Infervision, to focus his work with deep learning and image recognition
on the world of medicine.
33. Results, Key Challenges, Learning Points & Takeaways
Partnership with over 200 hospitals around the world.
Technology is being currently being used to analyse 20000
scans a day.
AI will lead to a digital shift in traditional medical imaging,
requiring AI & people to work together to meet the challenges of
the medical industry.
Deep neural enable computer algorithms to become
increasingly efficient at sorting images.
New value can be extracted from old data with cutting edge
technology such as deep learning .
It will not replace doctors, but rather enable them to work far
more quickly & efficiently than they previously could
35. MasterCard
Processed Billion of transactions – crucial link between thousands
of banks and millions of merchant establishments.
Year 2017 acquired Brighterion to complete its mission of rolling out
AI technology across its network.
Having a card transaction declined at the checkout can be a
frustrating and embarrassing occurrence.
So much so that it can seriously damage brand loyalty – according
to research by MasterCard, a third of us have withdrawn our custom
from a retailer due to our cards being refused.
Often this is due to the transaction being incorrectly flagged as
fraudulent in some way – the algorithms which make the call on
whether a payment is valid have erred on the side of caution, and
sometimes they get it wrong.
Aside from the inconvenience, it causes us, the cost to businesses
and the wider economy of these false declines is around $118 billion
– an amount 13 times higher than the cost of actual card fraud.
36. MasterCard – Real time Analytics
The quantum leap in the ability to both detect fraud and reduce false
declines has come about through its acquisition of California-based artificial
intelligence specialists Brighterion.
Technology developed with Brighterion has enabled it to move to analyzing
data in real time.
Machine learning algorithms must be incredibly efficient to handle the 75
billion transactions per year happening at 45 million global locations, which
are processed by the MasterCard network.
Today, the decisions of whether or not to decline a transaction are based on
a constantly flowing stream of data, and self-teaching algorithms, rather
than a static sample dataset and fixed rules, which has had impressive
results.
The artificial intelligence systems, because they are self-learning, are
always current and there is no longer a learning lag happening.
“What it does is goes through billions of transactions and figures out what is
the propensity of the transaction being fraudulent, and it gives this advice to
the bank in the system, when the transaction goes through for
authorization.
37. MasterCard – Real time Analytics
The system uses a real time stream of transactional data,
along with external data including anonymized and aggregated
customer information, and geographical information.
Geographical information is highly useful because not only
does it give an overview of the types of transactions which are
“normal” for a particular area, it also reveals what patterns of
fraudulent activity are associated with it. Again, all of this
information is aggregated in real time as it happens.
This means that patterns of fraud – which is often carried out
at large scale by organized gangs, who will target businesses
in a particular location, or attempt to “cash out” at ATMs
spread across a city - can be detected, tracked and stopped.
38. MasterCard – The Challenges of AI
Building smart, automated systems has been a core strategy at
MasterCard for many years, but the acquisition of Brighterion and
the incorporation of its technology into MasterCard systems has
been a move towards “pure” AI.
Many areas of its business, from customer service to anti-money-
laundering measures, are set to benefit from an AI overhaul.
One key challenge has been ensuring a consistently high quality of
data – as errors in transaction records or other data stores will
inevitably lead to even the smartest machines making bad
decisions.
Company’s success with this down to the more than 50 years’
experience it has at generating and verifying transactional records –
“We have been doing it for many, many years”.
A second challenge is determining the priorities when it comes to
making decisions on where in the business to deploy potentially
costly AI infrastructure.
39. Results, Key Challenges, Learning Points & Takeaways
Detecting the fraud detection has increased by 3 folds.
False positives has been reduced by 50%.
Decision based on static datasets using fixed rules is not
sufficient for fast, hassle-free fraud verification over a network
of MasterCard's scale.
Datasets & predictive models that update in real time allow for
far more accurate predictions about the legitimacy of a
transaction, meaning fewer false declines.
Talent acquisition became a challenges when the decision was
made to implement AI . By acquiring Brighterion this challenge
was mitigated.
Data quality is utmost importance – inaccurate data would lead
to a potentially even greater number of false positives, or
fraudulent transactions being incorrectly approved.
40. How AI Helps Businesses understand their Customers
41. Salesforce
Salesforce – one of the world’s leading CRM solution provider.
Product & services address the business growth and track
relationships with customers.
Founded in 1999- Concept of software as-a-service ( Saas)
over internet.
Business challenges – Maintaining customer relationship – old
fashioned mail shots to social media & chatbots, acquire and
retain customers across different part of the globe.
Salesforce offers its customers –Einstein platform, which it calls
the world’s AI solution for CRM.
Cloud hosted CRM solution.
42. Salesforce
There are many ways that companies can use machine learning in
their sales process. Here are just a few of the possibilities:
Interpret customer data: ML helps make sense of the data we
collect about our customers. Research shows how important it is
to have a “data-driven understanding” of our customers.
Even though many organizations have systems and spent
resources to gather and store customer data, it’s the machine
learning that will now help us make effective use of that data in
ways that relying on humans alone could not.
Improve sales forecasting: When you gather data on your
prospect (company size, stakeholders, solutions they want) and
then through machine learning have the ability to compare it to
historical sales efforts, you can connect the dots and better predict
what solutions would be effective and the likelihood of the deal
closing and how long it will take. This insight helps sales
management better allocate resources and predict sales
projections.
43. Salesforce
Predict customer needs: Business success relies on how well we provide what
our customers need. Machine learning can improve how responsive and
proactive we are to anticipate the needs of our customers. The better we are in
sales at addressing our clients’ needs before they get escalated and at
suggesting a solution that could help make their life better and easier, the
stronger our relationship will be. Machines won’t forget to follow-up or be too
busy to proactively share solutions.
Efficient transactional sales: According to HBR, by 2020, customers will manage
85% of their interactions with an organization without interacting with a human.
Having machines step in to handle certain sales efforts quickly and effectively
can free up the human sales force to focus on the relationship.
Sales communication: There will most likely be dramatic changes to sales
communication as a result of machine learning. If business communication
mimics the transformation of consumer communication, the business equivalent
of short-form communication such as tweets and text messages will be AI
responses. Machines can quickly and easily answer queries about pricing,
product features or contract terms. Within the next decade, virtual reality would
allow prospects to tour a factory, “join” in to conferences and meetings with your
entire team and see products being manufactured, all without leaving their own
office.
44. Results, Key Challenges, Learning Points & Takeaways
With Einstein, Salesforce has effectively positioned itself as the
first provider of AI- as-a-service for CRM.
Delivering AI as –a-service has the potential to drive strong
economic growth by empowering businesses of any size to take
advantage of these powerful tools & technologies.
CR can effectively by managed in an automated way via ML by
algorithms that can learn the most effective approaches to
marketing & managing relationships with individuals according
to the profile they fit.
Salesforce makes data ownership a unique selling point of their
service , meaning that their customers do not have to let their
valuable customer-data out of their hands to take advantage of
its cloud-based services.
46. Uber
Business Model around disruptive data – pairing public hire drivers with
passengers by correlating location data from both parties smartphones.
Connecting waiting passengers far more quickly than conventional taxi
operators.
Uber has invested heavily in AI- AI First company – end to end of the
business.
AI helping to serve
How to serve customers with minimum expense through driver wages &
mileage.
Speedy response for the ride request- competitive edge .
Dealing with passengers who are drunk & abusive – Night rides.
How AI used in Practice
AI is used as a core business – Connecting & dispatching drivers to
passengers for pick up – Most efficient route.
Power company’s “ Surge Pricing” model – Based on supply & demand –
encourage drivers to clock in- reducing customer wait time.
47. Uber
How AI used in Practice
Hotel, Airlines & public transport have used this technique
balancing supply & demand- peak pricing .
Predictive technology to adjust pricing in real time.
Using AI analyzing the pick up point & destination-Business or
personal and suggest which account to be used if there are two
accounts.
Uses Machine learning algorithms to segments customers for
marketing programs , promotions , track frequency of opening
the app etc.
Recent patent application reveals that Uber has developed a
technology to predict whether a customer may be drunk.
Uses Machine learning within its Uber Eats food delivery
platform.
Can predict how long it will take a customer food to arrive.
48. Uber
Technology
Uber uses GPS data from passengers & drivers smart phone &
map data to plan routes between two points.
Data gathered from the journeys made is fed back into learning
algorithms with the aim of giving customers more accurate ETA
for their rides- shortening the waiting time.
Suggests the passengers a shorter waiting time for pick up
depending on the traffic condition- nearby location- so that
customer can move to that location instead of waiting.
Machine learning platform – Michelangelo – Data lake where it
logs all of its transactional and customer behavior data.
Uber’s current AI research division – Uber AI labs was formed
and acquired Geometric Intelligence in 2016.
Conducts research on deep learning & neural networks- going
beyond Uber business cases.
50. Results, Key Challenges, Learning Points & Takeaways
Shorter waiting time for rides & efficiently routed journeys
leading to customer satisfaction.
Customer retention – high lifetime value to business.
Success with Machine learning & predictive models- Scaling up
& global reach out.
Machine learning has been applied to all areas of business –
operational efficiencies & customer service.
Uber has disrupted the traditional taxi hire business globally.
51. Mail your comments to
ramaddster@gmail.com
End of Part -4
Will continue the summary in Part -5