by Bernard Marr
In the “old days”, scheduling was seen as a precise science, and TV networks carefully selected what programs would run at what times to fit in with our lives and earn our loyalty.
For example, scheduling news bulletins in the early evening when we return from work, followed by light entertainment as we relax in the evenings and a late-night movie before bedtime.
With on-demand forms of entertainment, this is often no longer possible. Customers being able to watch what they want, when they want, caused a quandary. What if they consistently pick the wrong programs to watch, and end up not feeling they are getting sufficient entertainment bang for their buck?
How Is Artificial Intelligence Used In Practice?
Netflix uses artificial intelligence (AI) to predict which of its catalog of more than 10,000 movies and TV shows you are likely to want to watch next.
These are the recommendations that pop up immediately after a movie or show finishes, as well as the content that appears in the service's menus when Netflix loads up on your TV, laptop or tablet.
Netflix originally used IMDB ratings together with the user's past viewing habits, and indications they give about shows they are interested in when they first sign up, to come up with a “personalized schedule” of content that it thinks a viewer will be interested in.2
Since then, Netflix has built up a huge dataset of viewing habits – on 7 January 2018 alone, its viewers set a record by streaming 350 million hours of content in one day.3
This means it knows an awful lot about what shows and movies people with similar habits to yourself are likely to enjoy.
What Technology, Tools And Data Were Used?
The most significant data that goes into its personalized scheduling algorithms is individual customers’ past viewing habits.
Netflix developed (and open sourced) its own deep learning library called Vectorflow to process the data it collects on customer viewing.4
Essentially, this is a recommendation engine – a key use case for AI technology as used by Amazon for product recommendations and Facebook with its “people you may know” features.
Netflix breaks down its content (movies and shows) and tags them according to individual elements – action films, psychological thrillers, female protagonists – there are tens of thousands of different tags that can be assigned to individual pieces of content.
It then measures how content that fits these tags matches with individuals’ viewing preferences. When it finds that particular tags work well with viewers that match a particular profile (based on their viewing history), it will recommend that content to others who also match that profile.
It is also used for a number of other functions across the service, such as optimizing streaming quality to ensure users receive the best possible picture quality.
When Netflix data scientists develop new machine learning methods that potentially provide more accurate predictions about what their customers will want to watch, they test them by initially rolling them out to a subset of customers.
If they find that the overall metrics improve, then it will be rolled out across the network. These metrics include the number of hours of content watched by the customers, as well as the churn rate – the rate at which customers cancel their subscriptions because they can't find anything to watch.5
To optimize the streaming and picture quality, Netflix uses algorithms that analyze every frame in real time to work out how it can be compressed to the smallest size possible, while still retaining all of the data that makes up the image people will see.6
Elements of each frame such as lighting, complexity (how individual parts of the image differ from others) and how much of the image will be moving in the next frame are all considered by the algorithm.
What Were The Results?
Netflix is able to accurately recommend content to viewers based on their preferences, and the preferences of others who match their profile. This leads to customers who re-subscribe for longer, offering longer lifetime value to the company.
As a content producer itself, it is also able to create new movies and TV shows that more closely match what its viewers want to watch.
Netflix's AI compression algorithms for minimizing the size of files that have to be transmitted, and therefore improving streaming quality, managed to reduce data usage by a factor of 1,000.
An episode of Jessica Jones, which would otherwise require 750 megabits per second of bandwidth, was reduced to 750 kilobits.7
Key Challenges, Learning Points And Takeaways
Moving from a mail-order to a subscription model vastly increased the amount of data that Netflix was able to collect, not just about what customers watch, but how and when they watch it.
Providing customers with more accurate recommendations about what they might want to watch means that fewer customers will cancel their subscriptions due to not being able to find movies and shows they will enjoy.
AI enables these recommendations to become finely tuned, as they learn from an ever-growing dataset of customer habits.
Netflix was able to use the vast database of viewing habits it built up to begin producing its own output, guided by data on what its customers want.
Streaming high definition and ultra-high definition video uses huge bandwidth resources – resources that are limited and expensive. AI can reduce these overheads by learning to transmit only the important data.
Notes
1Netflix, Shareholder's letter, 16 July 16 2018: https://s22.q4cdn.com/ 959853165/files/doc_financials/quarterly_reports/2018/q2/FINAL-Q2- 18-Shareholder-Letter.pdf
2It's Foss, Netflix Open Source AI: https://itsfoss.com/netflix-open-source-ai/
3Variety, Netflix Subscribers Streamed Record-Breaking 350 Million Hours of Video on Jan. 7: https://variety.com/2018/digital/news/netflix- 350-million-hours-1202721679
4Netflix, Introducing Vectorflow: https://medium.com/@NetflixTechBlog/ introducing-vectorflow-fe10d7f126b8
5Nvidia, How Netflix Uses AI: https://blogs.nvidia.com/blog/2018/06/01/ how-netflix-uses-ai/
6The Motley Fool, Netflix Streaming gets an AI Upgrade: https://www.fool. com/investing/2018/03/15/netflix-streaming-gets-an-ai-upgrade.aspx
7The Motley Fool, Netflix Streaming gets an AI Upgrade: https://www.fool. com/investing/2018/03/15/netflix-streaming-gets-an-ai-upgrade.aspx
25
Press Association: Using Artificial Intelligence To Cover Local News Stories
The Press Association (commonly referred to as PA) is a UK-based news agency that provides text and video news stories, photography, copywriting, TV listings and sports coverage to local, regional and national newspapers, magazines and TV stations across the country.
In 2017, it announced a partnership with Urbs Media to roll out news stories written by artificial intelligence (AI) “journalists” to local newspapers.
What Problems Is Artificial Intelligence Helping To Solve?
The local news industry in the United Kingdom has been in decline since the arrival of the internet, with more people turning to social media and online sites to keep up with local events. This has led to newspapers closing and journalists being made redundant.1
This has left a gap that has been described as “dangerous for democracy”. Local newspapers in effect operated as the “eyes and ears” of the general public into local political matters and regional administration, as well as matters involving healthcare and criminal justice.
Without reporters covering these beats, the public can't hold local authorities to account, and demand answers when they are needed.
How Is Artificial Intelligence Used In Practice?
PA partnered with data-driven journalism specialists Urbs Media to build an AI system capable of mass-producing localized news stories based on data that is fed into it.
While still employing human journalists to pinpoint the stories that are in need of coverage, the system uses AI algorithms to tell the story, and localizes it for newspapers and web
sites around the country.2
The service – known as Reporters and Data and Robots (RADAR) – was funded through a grant from Google's Digital News Initiative Fund, which aims to help journalism thrive in the digital age by harnessing new advances in technology.3
RADAR says that the project is not supposed to replace human journalists, but will make it easier for them to cover stories in a way that is relevant to local audiences by spotting trends in datasets such as government open data.
It then creates localized news reports explaining the impact of these trends at a local level.
As well as established local newspapers and outlets, the stories are available to the new breed of “hyperlocal” news sites that have sprung up in many communities to fill the gap left by the declining local newspaper industry.4
What Technology, Tools And Data Were Used?
The key AI technology used in the RADAR project is natural language processing and generation.5
This means it can “read” tables of statistics and information and parse them into news stories in human, natural language (English in this case).
For example, if it is given a list of average wait times that occur when members of the public call for an ambulance, it will be able to localize these by inferring which regions have good response times, average response times or poor response times.
Most of the data used is taken from open datasets published by government agencies, covering areas such as healthcare, education, law and order or demographic data.
What Are The Results?
News stories created with the RADAR system are now available to more than 1,000 local news outlets through the PA news feed.6
Mass-scale localization of news stories means there is more chance that important issues will be brought to the public's attention via local news outlets whose budgets are extremely tight.
When the AI surfaces issues that demand more in-depth and rigorous investigation, then human journalists can be assigned to stories to look for issues such as underlying causes that are not obvious from the data.
The stories can also help combat the problem of “fake news” – as it has been shown that when journalists neglect to cover issues of local importance, there is usually someone ready to provide their own take on the matter, and often this will be driven by personal or anecdotal experience rather than hard facts and data.
Overall this should lead to a more informed public, which has the information it needs at its fingertips to make decisions through local democracy.
Key Challenges, Learning Points And Takeaways
Local news outlets are severely pressured by budget constraints, which has left a potentially dangerous hole in their ability to cover issues of local importance.
AI can quickly and accurately compile news reports in easy-to-understand natural human language, simply using public datasets.
This will leave less room in the news ecosphere for peddlers of misinformation and “fake news”.
Human journalists will have more time available to carry out in-depth investigations into the background issues that may not be picked up by the AI working on the data alone.
Notes
1BBC, Death of the Local Newspaper: https://www.bbc.co.uk/news/uk-43106436
2The Drum, How PA and Urbs Media will use robots to strengthen local news, rather than devalue it: https://www.thedrum.com/opinion/ 2017/08/10/how-pa-and-urbs-media-will-use-robots-strengthen-local- news-rather-devalue-it
3Google, Radar (Round 3): https://newsinitiative.withgoogle.com/dnifund /dni-projects/radar/
4Press Association, More than 1,000 UK regional news titles now have access to stories jointly written by journalists and AI as RADAR launches new website: https://www.pressassociation.com/2018/06/18/more-than- 1000-uk-regional-news-titles-now-have-access-to-stories-jointly- written-by-journalists-and-ai-as-radar-launches-new-website/
5Press Association, Trial of automated news service underway as RADAR makes its first editorial hires: https://www.pressassociation.com/2017 /12/12/trial-automated-news-service-underway-radar-makes-first- editorial-hires/
6Press Association, More than 1,000 UK regional news titles now have access to stories jointly written by journalists and AI as RADAR launches new website: https://www.pressassociation.com/2018/06/18/more-than- 1000-uk-regional-news-titles-now-have-access-to-stories-jointly- written-by-journalists-and-ai-as-radar-launches-new-website/
26
Spotify: Using Artificial Intelligence To Find New Music You Will Love
Spotify is a streaming music service that launched in 2008 and now has 180 million active users and 83 million subscribers.1
Like other online services that have risen to prominence in the last decade, such as Amazon and Netflix, offering its users a large catalog of content at prices that undercut traditional delivery methods is only part of the recipe for success.
The “icing on the cake” is Spotify's advanced predictive technology powered by machine learning. This makes it possible to present that content in a way that people can make sense of and enjoy.
One successful way it has managed this is through its Discover Weekly playlists, which give users an artificial intelligence (AI)-curated playlist of new music it thinks they will enjoy.
What Problems Is Artificial Intelligence Helping To Solve?
With millions of songs at their fingertips, users are never short of music to listen to. However, they may find it difficult to discover new bands and artists in the way that radio listeners did in the past.
While it may be simple enough for them to search their favorite band's or singer's name and hear their latest releases, unearthing new talent from the thousands of new tracks added to the service every day is a trickier proposition.
How Is Artificial Intelligence Used In Practice?
Spotify presents users with 30 new tracks every week that it thinks they will love via their own personal Discover Weekly playlist.
To those of us who grew up making mix tapes for our friends by copying songs onto cassette, it's like having a new best friend who happens to be an AI.
Another way to look at it would be to think of the AI as filling a role traditionally played by the radio DJ – reading their audience's taste and playing songs they think they will enjoy.
One breakthrough that led to the Discover Weekly podcast was the realization that it wasn't just what Spotify recommends that matters to users, but also how it recommendeds it.2
As users have become used to the “playlist” concept as a form of music curation since the early days of digital music, it made sense that Spotify uses this format for presenting its automated recommendations.
What Technology, Tools And Data Were Used?
As with Netflix's recommendation engines, data for Spotify's Discover Weekly playlists is gathered by monitoring its users’ listening habits.
This makes it possible to begin to build up recommendations through a process called collaborative filtering.3
As a simple example, consider Person A regularly listens to music by Artist X and Artist Y. Another user, Person B, regularly listens to Artist Y and Artist Z.
With this data, a collaborative filtering algorithm can deduce, with some certainty, that Person A might enjoy being introduced to Artist Z, and Person B might enjoy the output of Artist X.
Of course, with millions of users and millions of songs, the matrix that is constructed to enable these suggestions to surface is considerably more complex than in this example, which is why AI algorithms are necessary to deliver these sorts of insights at scale.
It also looks for negative signals – skip a song within the first 30 seconds of it playing, and Spotify's AI algorithms will take that as a sign that you don't like it, and will give less weight to others similar to it in its recommendations.4
Spotify's recommendation engine goes further than that though, also utilizing audio analysis and natural language processing to create recommendations.
Audio analysis breaks do
wn each individual track into its constituent parts – for example, tempo, beat, pitch of the notes, types of instruments and sounds used, and the prominence and pattern of lyrics.
This allows it to fine-tune its calculation of the probability that a certain user will like a particular track by comparing these elements to those of their favorite songs, as well as songs enjoyed by other users who match their listening preferences.
Natural language processing takes in external data – text found online relating to particular tracks. Spotify crawls the web to find news articles and blog posts that talk about tracks. It analyzes the sentiment of text describing each song – whether it is frequently described as “upbeat”, “funky”, “melancholy” or “heavy” – and uses this data to determine how receptive an individual user is likely to be to it.5
Spotify uses deep learning and neural nets to bring all of this information together and make recommendations it knows – to a high degree of probability – its users will love.6
How about if you let a friend borrow your login details? Well, it turns out that Spotify is well aware that a proportion of its members do this. So, its AI algorithms are smart enough to ignore drastic, but short-lived changes in listening habits.
Spotify does not have its own data centers – in 2018 it completed migration of its entire platform to Google Cloud. This allows it to scale more quickly without having to continuously upgrade its infrastructure to cope with new users coming on board.7