I went to Selfridges this morning; I was in Oxford Street and had half an hour to kill, so I drifted in for a mooch. I walked in the main door, managed to dodge the perfume-muggers, and took the escalator. I was one of thousands of people who did this today, and what struck me was that anybody walking in through that door, be they old, young, regular visitor or tourist, gets hit by the same smell-jockeys and sees the same counters; we all have the same experience.
Such are the realities of the physical world, where moving all the counters around for the latest customer is impractical. But what if we could?
What if we could know what sort of things they're likely to buy, though they might not know themselves? What if we could reorganise the store for them?
The web can deliver this, but until now, the tools available meant a lot of resources were needed to make it happen. Today, leading-edge websites are using artificial intelligence to bridge that final gap to personalisation. Early on, webmasters used A/B testing to identify what worked. They'd try a red button for some consumers and a green button for others; if red generated more sales, it'd be rolled out, and testing would move on to another element.
Later, sites used multivariate testing (MVT). Recognising that it's the combination of different elements that leads to success, MVT mixes up the testing of many variables such as copy, images, colours and layout at the same time, and uses statistical analysis to pick the most-effective permutations. The technique generates an ideal average; there are more sophisticated applications that can further segment these ideal page configurations according to customer type, allowing different pages to be displayed to different types of customer.
To do this, a decision tree is created and a rule-set is overlaid. For example: 'If user has placed product x in basket but abandoned funnel in past five days, re-present product x', a simple binary statement, creating two potential outcomes. On any website, though, there are potentially hundreds of actions and other variables that have to be legislated for, meaning a decision tree quickly becomes unwieldy.
Though MVT transformed the effectiveness of websites, two problems remained. First, the segmentation is difficult to scale beyond fairly coarse granularity. Second, it changes over time, due to external influences like the weather. So MVT is an effective tool to design the framework, but it's not dynamic enough to decide what product to show.
What's needed is a system that can analyse thousands of data points and decide 'in the moment' what the customer is most likely to be interested in.
Machine learning, a branch of artificial intelligence, enables computers to develop their responses based on empirical data. Used in the stock market, it has also been applied to detecting credit-card fraud and, crucially, evolving this automatically in response to changing information.
Many US websites are applying this technology to their purchase funnel, and it's taking hold in the UK too. Purchase history, abandoned carts, search keyword, temperature and other variables are fed into a dynamic model. This enables an instant decision on which product or content to show, giving customers what they want.
Publishers use it to recommend stories, retailers to suggest which shoe or jacket you might like. If Selfridges could do this in Oxford Street, I might never see another perfume counter.
- Andrew Walmsley is a digital pluralist
30 SECONDS ON ... MACHINE LEARNING
- Machine learning is a branch of artificial intelligence concerned with allowing computers to evolve their behaviours and make intelligent decisions based on empirical data.
- Many machine learning systems aim to sideline the need for human intuition in analysing data; others take a more collaborative approach. Of course, intuition can't be bypassed completely, since any system is, after all, designed by a human.
- Machine learning is already applied to uses such as speechand handwriting-recognition, detecting credit-card fraud, medical diagnosis, robot locomotion and stock market analysis.
- Some experts, such as author Martin Ford, argue that such technology could result in rising unemployment, as machines progress from being able to fill routine jobs to even exceeding human performance in skilled roles that require educated employees.
- Other reputable sources have predicted machines will be able to perform low-invasive surgery by 2017. Office and hospital jobs are not the only ones under threat; the US department of defence predicts that its first robotic soldiers will be ready for the battlefield by about 2035.
- All of which brings us to a 2006 UK government study, which predicted that robots could one day demand the same legal rights as humans.