How to implement personalisation using machine learning

Personalisation has long been cited as being the 'next big thing', but we're yet to see widespread adoption of sophisticated personalisation using machine learning. So why are marketers not yet embracing it?

The recent Econsultancy Conversion Optimisation Report confirmed that, while marketers are keen to get going on implementing personalisation, they just never seem to get around to it. Of those brands that actually are doing personalisation, it's typically very basic, focusing on emails, a homepage and maybe, a landing page or two.

There are a number of contributory factors to this, but invariably the stumbling block marketers face is the fact that their data is too siloed to make personalisation possible. The growing trend for machine learning - which relies on a solid, well-connected data network to be able to work efficiently - has highlighted this issue further.

By connecting and enriching your data - for example, linking up your returns and purchase data or ensuring your back-end data feeds into your analytics - you can really begin to personalise your customer experience in a meaningful way.

So what are the machine learning and personalisation techniques you can use to up the ante, and how do they work? Here are four key principles for employing machine learning in your personalisation efforts:

1. Use data mining

'Data mining' is when you use an algorithm to search for opportunities in the data. These algorithms are designed to 'mine' large data sets to identify patterns of behaviour that can help you uncover any identify issues with your current campaigns or opportunities for future growth.

For instance, the data may identify that PPC traffic converts well at weekend and not much during the week, so you might adapt your campaign spend based on that information. Or you may find a campaign isn't working well in a particular region and may adapt the messaging based on this insight.

Data mining is useful in identifying patterns, but it needs to be used in conjunction with experimentation and other conversion optimisation methods to be useful.

So why bother? It can save you time (and money): Data mining automates your initial insight work and can help you find opportunities that might be otherwise hard to spot. And if you have a tool that incorporates data mining as part of the service, it is even easier and quicker.

2. Use predictive algorithms

Predictive algorithms power recommendation engines such as 'what others also bought'.
Lots of commerce sites use these features, but Amazon has taken things to the next level.

Amazon uses its predictive algorithms to anticipate how motivated a user is to buy, user interest and the current sales volume of the individual products. It then provides product recommendations based on this behavioural and operational data (so, for example, you and I will get a different set of search results for 'green hat)'. This also explains why you sometimes get search results for a product with little or no reviews appearing above items with hundreds of reviews in the search results.

This personalisation tool accounts for 35% of Amazon's sales and is obviously very complex, requiring a huge amount of data maturity - but it shows what's possible when you put a strong data strategy in place.

3. Take advantage of best match algorithms

A term you might hear your Conversion Optimisation team using is 'best match algorithms' or 'multi-armed bandits', which is a method used in A/B testing to automatically favour a variation as the experiment progresses.

Say you're doing a standard multi-variate test where you send 50% of traffic to one version and 50% to the control. Dependent on traffic, you would usually allow the test to run for a set period of time, then allocate the winner and direct all traffic to that variation.

With a 'best match' algorithms, as soon as a winner becomes obvious, it will automatically redirect more traffic to that winning test. This means you'll capture value sooner.

The New York Times is currently employing this technique to test their headlines. Each story is given two different headlines which are split tested, and the winning variant is favoured as the experiment progresses until all the traffic is focused on the winning headline. This process usually takes less than an hour.

This type of algorithm is great for short-term promotions and campaigns but can also be used in long-term tests. Say you have two versions of a homepage, and over time, one proves much more popular, getting 95% of all traffic. Then a seasonal change occurs - for example, Black Friday - the page that initially delivered much lower performance might suddenly start performing better. The 'best match' the algorithm would detect this and start directing more traffic towards the second page, and then automatically switch things back once the seasonal change alters traffic figures again.

4. Connect your data

As these examples have shown, joining up your data sources is vital if you want to undertake more advanced personalisation. We know that is easier said than done, but if you can enrich your data to allow patterns to be identified, you'll be in a much better place than the majority of your competitors to create a relevant and engaging customer experience and improve your site conversion.

10 questions you should be looking to answer with your data