How machine learning plays a key role in Amazon retail and Kindle services

Amazon director of machine learning Ralf Herbrich describes how machine learning improves services

Machine learning already plays a critical role in Amazon services, ranging from its online store to features within Kindle, all the way to servers helping to match people playing online video games.

That's what Ralf Herbrich, Amazon director of machine learning, told the audience at the Yandex Data Factory's 'Machine Learning and Big Data: Business Challenges' conference in Berlin, Germany.

He made the comments during in a panel discussion about using big data analytics in the real world. Amazon offers a machine learning service to enterprise customers, but Herbrich offered three examples of how machine learning is helping Amazon bring benefits to users of its more mainstream services.

First, Herbrich discussed how machine learning enables Amazon to provide the lowest prices possible to online shoppers by using algorithms to make comparisons with the prices of the same products being sold by other online retailers.

"One of the premises [of Amazon] is that you have very low prices. That sounds easy, but when you think of it, we have about 20 million products which we buy and sell, so in order to set the lowest price we have to find what the price for that item is," he explained.

However, Herbrich pointed out that Amazon could never do this manually, so machine learning and analytics are used in order to quickly make price comparisons for millions of products.

"With 20 million products you can't just separate it out, hire 20 people and have each of them manually check the prices for a million products; that's where machine learning is used," he said.

"A predictive model goes out and tries to find products - or products of a similar description - to set the prices," he continued, adding: "That's a machine learning power."

Herbrich also described how Amazon uses machine learning to improve the experience offered by its Kindle e-book reader, detailing its role in what he dubbed an ‘X-Ray button'.

If you click that button on the top bar, what you get is an index into the book that supplies people who are mentioned in that story," he said, then explained how this benefits the readers of complex texts, giving the example of 'Game of Thrones'.

"So if you have A Game of Thrones you have Drogo, and Targaryens and Lannisters. You see all the places they're mentioned in an index and quickly jump to there. And, you can see the reference to Wikipedia and the about page," said Herbrich.

He described how this "could probably be done by hand" for some books, but that isn't the case for the majority of texts.

"A lot of these works need a semi-autonomous process, where you use machine learning to predict which tokens, and which words in the book, are actually the beginning, the end or the middle of a place, organisation or a person," he said.

The machine learning algorithm does this for books in various different languages, providing the reader with features that can help them search for terms in the book without having to manually scan it themselves.

"It's a machine learning process where the machine scans data. It's a new, pre-processing feature, which produces a machine model," said Herbrich.

Machine learning is also harnessed in online video games, he continued, especially when it comes to matchmaking in online shooters on the Microsoft Xbox console for games such as Halo.

"Every time you play online matches in these video games, that outcome of the match - which person was playing on which team, and which team came in first or second - is fed into a predictive model, with updates based on skill in order to predict the probability of winning of every player that's playing online against every other player," he said.

The reason, he said, is to provide players with the most potential of an evenly matched game.

"So the matches get created by optimising towards a 50 per cent chance of winning or losing. That's a machine learning model at work, constantly matching millions of people every day," Herbrich concluded.

Amazon recently unveiled an AWS Internet of Things platform for connecting cloud devices.