Machine Learning – Web Analytics Wednesday May 2016

We recently wrote a blog post about the future of artificial intelligence and the likelihood of our jobs being automated by a robot. This got us thinking, and we decided to hold the latest #WAWlondon on the theme of machine learning.
There is a lot of information widely available online and rather than simply define what it is, at Lynchpin we are known for our practical approach. Therefore we were delighted to have three different speakers all using machine learning techniques now so our audience could get a real sense of what they can explore for their own businesses.
Harvinder Atwal – Machine Learning: What They Don’t teach you at Coursera
Our first speaker was Harvinder Atwal, Head of Customer Insight and Marketing Optimisation at MoneySuperMarket. MoneySuperMarket is one of the most visited websites in the UK and Google’s biggest PPC customer. Around 1.4 billion people saved money last year by using their website.
From Harvinder’s talk, clearly machine learning is “soo money supermarket” and if you have been delightfully irritated by their off the wall ad campaigns they have done some brilliant t-werk lately.
Harvinder explained various ways MoneySuperMarket are using machine learning, but also the practicalities of deploying the algorithms, on-boarding the wider business and getting projects shipped on time – the stuff they don’t teach in the textbooks.
MoneySuperMarket have used machine learning successfully in their comparison tools. Using all the data they already hold on their customers they were able to create a supervised regression model that could quickly tell users how much they should be paying for energy bills, insurance and even broadband based simply on postcode.
This helps customers determine whether it is worth switching provider without the laborious task of repetitively filling out reams of forms. The outcome here is not necessarily just around short term profit, but delivering a better customer experience and setting MoneySuperMarket apart from the competition.
Reality Check – Don’t expect everyone in the business to understand
A machine learning regression model is regular programming flipped on its head where you use outcomes present in historical data to train the algorithm. Harvinder made it look easy as he talked about customer personas and unsupervised clustering, however, he admitted it’s not as easy as it looks and there are many practical considerations.
One that we see time and again at Lynchpin is getting key stakeholders to buy into something they perhaps don’t fully understand. One way to do this is to run and test models first and then scale from there. This is something we look at doing with our clients and was also mentioned later by fellow speaker Richard Fergie from e-Analytica.
Richard’s talk was about using machine learning to predict uplift in PPC campaigns. It’s better to relay results to the marketing department in terms of risk and reward and pounds and pence rather than trying to explain and proving a mathematical hypothesis.
 
Machine Models incur costs and require technical maintenance
Harvinder quoted another Lynchpin mantra: when implementing a new model onto the website or indeed anything from scratch make sure you build measurement in from the start.
Building and maintaining a machine learning model costs money and a good chunk of that cost will be development costs. MoneySuperMarket hands over to developers to turn models into high throughput services, which is a vital but potentially time consuming process.
Models also require monitoring when in production, and potentially retraining (or even terminating) as consumer and/or market behaviour evolves.
Beware feedback loops and document like crazy
A benefit of machine learning is that new data will quickly feedback into the model, but this can risk a key variable or customer segment being missed and a negative result being reached: ultimately customers don’t all behave the same. As well as some brilliant practical demos, Nick Walker highlighted some of the interesting results that could arise from unsupervised models being allowed to run the world without suitable checks and balances! And delighted the audience with facial recognition that will change which advert you are served , the ultimate real time and physical world personalisation.
Harvinder highlighted that models will almost certainly need to be revisited in the future, and documentation with a proper code log and discipline in removing dead code to make it easier to debug and add new functionality in the future is a critical success factor.
Quote of the evening: “Documentation is like sex: any is better than none”
Web Analytics Wednesday happens every two months. Sign up to the meetup group here.