How human intervention, as well as utilising accurate, high-quality content can help overcome the challenges of Sitecore machine learning
Since the launch of Sitecore 9, Sitecore machine learning has become even more of a popular topic in the digital marketing and tech world. After all, the technology behind Sitecore machine learning will play an increasingly huge part in achieving the personalised experiences the platform is renowned for.
Machine learning has great potential for the future, nobody is doubting that. What isn’t often covered is the specific challenges the technology could face, not through any fault in its technology, but from the content it has to deal with and serve up to users on a daily basis.
The high amount of content online means that there’s essentially good content, bad content and irrelevant content. Through human assistance, such as working with machine learning and creating great content, we can help to create a scenario in which we can utilise the technology to provide outstanding personalised experiences.
For now, here are the challenges machine learning faces, particularly when faced with bad content.
Before my development career, I worked closely with various people who had been in the armed forces. Of all the tales I’d heard about IT issues in the forces, the story that sticks in my mind the most on the topic of machine learning is about spies.
Decoding surveillance photography to track enemy movements is a skilled job. Spotting images that show decoy tactics almost seems like an ideal job for automation – doesn’t it?
Initially, a company attempted to use machine learning to filter out any images depicting enemy tanks. The idea then, was that human expertise could focus solely on images that were of interest in the analysis process. Next, the developers of the system collated a large number of existing images (some showing tanks and some without) and went on to train their computer. When the machine learning model was completed, the developers were able to reliably separate the resulting test data into two groups.
Thrilled with the results, they went on to classify new photos as they were being taken. This did not go so well. When presented with the “live” data, the system’s accuracy dropped to less than 50%. The project team were puzzled – why did their system work so well with the training data, but fail so badly with real data?
After further investigation, the developers realised their mistake: there was a slight correlation in their training data, between “images that contained tanks” and “images taken when the sun was high in the sky”. Their system had got its metaphorical wires crossed and took their direction to spot “images taken around lunchtime” based on what the shadows looked like.
As far as the future of Sitecore machine learning goes, it’s important for us to consider the consequences before we use the technology to make website personalisation decisions. Machine learning systems are complex, and as far as our traditional Sitecore approach to rules-based personalisation to choose relevant content goes, the outcomes of a system like that are obvious.
You can see the decisions the computer might make on screen, and the logic the system will apply to it is readable for humans. However, that is often not the case with a machine learning system – the choices are concealed in “tensor algorithms” which, without advanced mathematics knowledge, aren’t easy to understand.
Not understanding how the computer will make specific decisions, is a challenge that is difficult to guard against once you’ve committed to trusting your computer to make important choices for you.
Fundamentally there are some things which computers, and robots if you like, find easier to do better than humans, just as there are some things which they will struggle with.
Crunching big data to win at complicated games is something modern computers can do reliably, yet differentiating between images of dogs and muffins isn’t something they’ve quite mastered yet.
Human brains are still naturally advanced in making the complex decisions that we’re yet to see machines fully grasp, such as instinctively arriving at the right answers for visual recognition. This level of decision making can be applied to choosing content for web pages, including selecting appropriate imagery.
I deliberately use the term ‘appropriate’. Another area that AI just hasn’t been able to learn is if the content it’s choosing could be interpreted as something offensive (or hilarious depending on your audience’s sense of humour).
So whilst we can laugh at the computer fails, inappropriate content and poor advert positioning, we do need to recognise that they do have the potential to do harm to brands and their reputations.
First and foremost, not cutting humans out of the picture is probably the simplest and most obvious answer – at least for the time being. If we’re wanting to rely on computers to make decisions for us, when it comes to the production of website content – the best option is likely to be those where automated choices come from a pool of data that humans have pre-examined.
Success is likely where the computers are used to present options to content editors, who can then decide whether they are appropriate or not. “Flags of countries that are red” is the sort of request a computer can answer quickly, whilst the editor can spot which flags might be relevant for their specific web page.
A huge benefit of artificial intelligence is how it can save editors time, in this case by collating relevant data and leave the really important decision of “which image will be best for my users” to the human brain.
The alternative is that computers will become more advanced than humans and there will be a ‘rise of the machines’ and robots will take over the world – that opens a whole new debate which I’ll leave for another day.
Working and living in an increasingly digital landscape it’s important to be aware of the risks and challenges AI can pose and make sure you think them through and test with live data before deploying new systems publicly. Artificial intelligence will soon become a standard in every digital marketer’s toolbox, and for businesses, and more specifically, digital marketers and content creators, it’s important to work towards addressing the challenges and making the most of the opportunities it presents sooner than later.