by Bernard Marr
29
Verizon: Using Machine Learning To Assess Service Quality
Verizon started life as one of the “baby Bells”, coming into existence initially as Bell Atlantic when the US Justice Department forced the breaking up of the Bell telephone conglomerate in 1984.
Today, as Verizon Communications, it is one of the largest communication technology companies in the world. It is the No. 1 provider of wireless subscription services in the United States1 and it offers high-speed fibre optic broadband services to millions of US subscribers through its Fios service.2
Until recently, Verizon's main source of data on how well the network was running and the quality of service experienced by its users came from customer feedback.
It now monitors traffic and data across its network and uses machine learning to understand how service quality is affected by usage spikes, as well as external factors such as the weather and changing customer habits.
Verizon brought additional machine learning expertise into the business through its acquisition of Yahoo! in 2017.
What Problem Is Artificial Intelligence Helping To Solve?
Monitoring a network of the scale of Verizon's to understand where faults and outages occur takes a monumental amount of effort.
Traditionally, this has been done through customer feedback –essentially waiting for something to go wrong and the complaints about poor service to start flooding in.
It was only possible to react to problems after they occur – meaning that even if Verizon was able to find the cause and fix it, customers had already experienced a drop in their quality of service.
Being able to predict where problems arise before they affect customers would be preferable, but until machine learning was sufficiently advanced, the analytics tools needed to make such accurate predictions were available.
How Is Artificial Intelligence Used In Practice?
Verizon's machine learning algorithms crunch data gathered from all of its network elements and use the insights to understand how and when outages and faults occur.
This means that it is able to recognize when situations are occurring that are similar to those that have caused network problems in the past – for example, spikes in customer data use, or extreme weather conditions that could lead to equipment failure.
It does this by analyzing all of the factors available and establishing “normal” levels of operation. It then looks for indicators of outliers – events that fall outside of normal patterns of behavior – and attempts to establish their cause.
Verizon's director of network performance and analytics, Matt Tegerdine, told me: “The beauty of this is that we don't just look at one singular data source like interface statistics – we're also going out and collecting things like environmental statistics, CPU usage on routers – we use machine learning to learn what ‘normal’ is.”
The aim is to listen to as many network elements as possible and then use predictive modelling to ensure that customers receive an interruption-free service wherever possible.
In fact, customer satisfaction is the metric that underpins the entire strategy, with reducing customer “churn” – the number of customers who do not renew their subscription – the priority.
Verizon has also introduced a chatbot that operates through Facebook Messenger. Customers are able to use the familiar conversational interface they use to chat with their friends to ask natural language questions about what's on television, get technical advice such as information on how to reset their router and ask for up-to-date billing information.3
What Technology, Tools And Data Were Used?
Tegerdine told me: “It's a very complex ecosystem of different data sources, and it's that combination that drives a lot of insights and is where the value of the analytics increases.”
In fact, Verizon's predictive analytics algorithms monitor 3 GB of data every second, streamed from millions of network interfaces, customer routers, sensors that gather temperature and weather data, and operational data, which includes customer billing records.
Verizon's chatbot uses natural language processing and neural net technology to answer customer questions through Facebook's Messenger platform.
What Were The Results?
In 2017, Verizon predicted 200 “customer impacting” events before they happened, thanks to its machine learning-driven predictive technology.4
Because of this, these issues were remedied before they caused a problem to its customers.
The telecoms giant was also able to use insights generated through its network monitoring platform to drive business and marketing decisions. Through the extensive monitoring and testing of the service, engineers were surprised to learn that their 750 MB/s service was actually consistently providing speeds of 1 GB/s into customers’ homes. This meant they were able to rebrand their service as a 1 GB service, leading to a noticeable increase in sales.5
Key Challenges, Learning Points And Takeaways
Predicting customer-impacting events means that they can be fixed before they cause issues, leading to higher customer satisfaction and reduced churn rates.
The size and reach of Verizon's network mean that there is a wealth of data available that can be used to predict these events.
The fact that much of Verizon's data is internal data and not available to other businesses gives it a competitive advantage.
Notes
1Recode, A merged T-Mobile and Sprint will still be smaller than AT&T or Verizon: https://www.recode.net/2018/4/30/17300652/tmobile-sprint-att -verizon-merger-wireless-subscriber-chart
2Verizon Fios: https://www.verizon.com/home/fios/
3Knowledge@Wharton, Tapping AI: The Future of Customer Experience at Verizon Fios: http://knowledge.wharton.upenn.edu/article/ competing-with-the-disruptors-a-view-of-future-customer-experience- at-verizon-fios/
4Verizon, How Verizon is using artificial intelligence and machine learning to help maintain network superiority: https://www.verizon.com/about/ our-company/fourth-industrial-revolution/how-verizon-using-artificial- intelligence-and-machine-learning-help-maintain-network
5Forbes, The Amazing Ways Verizon Uses AI And Machine Learning To Improve Performance: https://www.forbes.com/sites/bernardmarr/2018/ 06/22/the-amazing-ways-verizon-uses-ai-and-machine-learning-to- improve-performance/#2b859af07638
30
Viacom: Using Artificial Intelligence To Stream Videos Faster And Improve Customer Experience
Viacom is a vast media network, which as well as taking in household name brands such as Nickelodeon, Comedy Central and MTV, spans 400 YouTube channels, 60 Instagram pages, 430 Facebook pages and 100 Twitter handles.
It has invested heavily in real-time analytics across its networks, using artificial intelligence (AI) platforms to draw out insights that it can use to improve the customer experience.
What Problem Is Artificial Intelligence Helping To Solve?
With so many channels – taking in both TV brands and social media – data is plentiful regarding customer viewing habits, preferences and convenience. Making sense of all of that data can be tricky. A business like Viacom needs to understand how every variable, from Facebook posts “likes” to the time it takes to launch a streaming video, impacts on the amount of time viewers will spend with it.
It also has to ensure there is sufficient bandwidth availability to pump its content out to its customers around the world. Getting this wrong can lead to buffering and stutter during video playback – a factor that is very likely to get customers looking elsewhere for their entertainment.
Viacom's senior director of product analytics, Dan Morris, told me: “Delivery of video is at the core of everything we do, and our goal is to be exceptional at that.
“But there are a lot of variables at play – we have internal systems talking to external systems, we have content delivery, we have ad servers, and on the user side there's a whole bunch of environmental factors like wi-fi connectivity which we really have no control
over.”
How Is Artificial Intelligence Used In Practice?
Viacom used network data and social media signals to understand as much as possible about how its audience consumed its services. From this it was able to determine its “North Star” metric – the most important target to achieve to hit its goals.
Facebook in its early days deduced through running analytics on its users’ behavior that if it could get them connected to seven friends within 10 days, they were likely to become long-term users of the platform.1
Viacom's North Star metric turned out to be that customers who were hooked on two or more shows had a 350% more likely chance of becoming loyal, long-term viewers of Viacom output.
By persuading them to watch four shows regularly, that probability leapt up to 700%.
This insight means the company was able to dedicate more resources to persuading customers who are already enjoying one show to also watch a second, third or fourth.
Long-term, loyal viewers who look to Viacom's properties first when looking to spend some time viewing content are what attract advertisers to the network, and generate the company's revenue.
Viacom also uses AI algorithms to monitor the flow of data and available bandwidth across its online streaming video platforms.
The quality of its video feeds is constantly monitored to understand where customers are receiving poor service.
Here, it was able to ascertain that two variables had the highest impact on whether customers would continue to watch. These were “time to first frame” – how long the video took to start playing – and the rebuffering rate – how frequently the video stutters during playback to load more data.
By using AI-driven analytics to reroute bandwidth availability to keep these two metrics at optimum levels, Viacom enhances its customers’ experience.
What Technology, Tools And Data Were Used?
Viacom put together a seven-person data science team to orchestrate automated data capture and analytics across its hundreds of social media channels.2
It now has a tool that pulls information from social networks as frequently as every five minutes, monitoring the performance of content marketing posts promoting its brands, driving traffic to its websites or suggesting customers “tune in” to popular shows.
“Social war rooms” are established around all of the network's “tentpole” shows to understand how its viewers’ experience is affected by variables that can be influenced through social media – such as post engagement rates, timing and choice of channel.3
To monitor and analyze network signals that allow it to observe the behavior of its streaming video playback, Viacom built another platform utilizing machine learning, using Apache Spark and Databricks, running on Amazon Web Services.4
The system proactively monitors video feed quality and automatically allocates resources when user experience drops below optimal levels.
What Were The Results?
Viacom's social media analytics platform helps it to measure the impact posting different content at different times of the day and on different networks has on customer viewing habits.
It helps Viacom to distribute the resources at its disposal for promoting content – for example, social influencers – where the analytics suggests they will positively affect metrics – such as the number of shows a customer enjoys.
Through its Databricks system, Viacom was able to drive an overall reduction of 33% in the time it takes streaming videos to start when played over its web services. This improvement in viewer experience was shown to increase customer retention and drive brand loyalty.
Key Challenges, Learning Points And Takeaways
The impact of social signals such as post engagement can be demystified with AI, and their impact on core business processes more deeply understood.
Social media offers unparalleled opportunity to get to know your customer, but you may need advanced tools such as AI to cut through all of the noise and find the insights that matter.
AI is now powerful enough to implement real-time monitoring and automate resource management across vast data networks, such as Viacom's streaming video output.
Identifying a key driver of success – a North Pole metric – is a primary use case for many AI analytics initiatives in business.
Notes
1Startup Marketing, How Chamath Palihapitiya put Facebook on the path to 1 billion users: https://ryangum.com/chamath-palihapitiya-how-we-put-facebook-on-the-path-to-1-billion-users
2Digiday, How Viacom uses artificial intelligence to predict the success of its social campaigns: https://digiday.com/media/viacom-uses-artificial-intelligence-predict-success-social-campaigns/
3Digiday, How Viacom uses artificial intelligence to predict the success of its social campaigns: https://digiday.com/media/viacom-uses-artificial-intelligence-predict-success-social-campaigns/
4Databricks, Customer Case Study, Viacom: https://databricks.com/wp-content/uploads/2018/04/viacom-case-study.pdf
Part 4
Services, Financial and Healthcare Companies
31
American Express: Using Artificial Intelligence To Detect Fraud And Improve Customer Experience
American Express handles more than 25% of US credit card spending, accounting for $1.1 trillion of transactions in 2017,1 and is the world's most valuable financial services brand according to Forbes.2
The company puts data and analytics, driven by machine learning, at the heart of everything it does; however, two of its key use cases are detecting fraud and improving customer experience.
What Problem Is Artificial Intelligence Helping To Solve?
Global credit card fraud causes losses of around $20 billion to businesses and customers every year.3 This often takes the form of card not present fraud when stolen or forged details are used to purchase goods or services over the internet or telephone.
Card payment processing systems must be built to handle large volumes of transactions every minute of every day to be useful to businesses and consumers. This means attempts at fraud must be detected quickly and there is only a small window of opportunity to do so. Incorrectly flagging up valid transactions as fraudulent is inconvenient for customers, and if it happens too often, they will look to other methods of payment to avoid the hassle.
As fraudsters are often technologically adept themselves, they have developed and deployed high-tech systems of their own to circumvent anti-fraud security systems. This can include “spoofing” location data to make it appear to security systems that transactions are originating from a different part of the world, or identity fraud to make it appear that a trusted customer is behind a particular transaction.
An “arms race” has been ongoing for decades as banks and fraudsters seek to outsmart each other. Artificial intelligence (AI) represents the latest evolution of this competition, and while fraud will almost certainly always exist, banks and credit card companies hope they can use technology to give their customers the confidence they need to conduct business through their networks.
How Is Artificial Intelligence Used In Practice?
Banks and other financial institutions such as card issuers and insurers have always used patterns presented in historical data to attempt to detect fraud. Examples include watching out for cardholders making uncharacteristically high-value transactions, or transactions that appear to originate from outside their home country.
Transactions, once found to be fraudulent, are logged, and characteristics are flagged as possible indicators that can be used to suggest future transactions may also be fraudulent. The personal details of the person making the transaction, the place where the transaction originates and the suppliers, goods and services involved are all potential indicators.
These allowed financial institutions to build models that can be used to predict the trustworthiness of future transactions, but they are cumbersome to build and often could only be updated with new information infrequently rather than in real time.4
American Express has built AI systems that are able to read in data from card transactions around the world in real time as they happen. This means fraudulent characteristics can be logged and fed back into detection algorithms almost in real time. Having access to far more data means more complex patterns of characteristics can be examined.
This means that even if fraudsters are able to spoof or fake certain characteristics of the transaction, there's a better chance that algorithms will detect anomalies in other characteristics quickly enough to warn that the transaction is suspect.
Aside from this, American Express also uses machine learning in several ways that are designed to improve the customer experience by giving added value to card users. One example is through its acquisition of AI-driven “personal travel assistant” app, Mezi.5
The idea here is that as well as helping you to securely spend money, AI assistants built into card apps can help you decide where to spend it by offering personalized recommendations based on your habits and previous purchase history – much like a recommendation engine of the type used by Amazon and other online retailers.
What Technology, Tools And Data Were Used?
Data primarily comes from historical transaction records as well as information gathered about individual customers when they sign up to become American Express cardholders.
Using machine learning to spot fraud among the millions of transactions occurring every day requires sophisticated storage solutions that are able to cope with ingesting and making available large volumes of data. To achieve this, American Express uses a Hadoop-based distributed storage infrastructure.6