It was our first Umbraco Spark conference. Steve Temple, our Technical Director, needed a talk. He had become interested in the Machine Learning tools offered by Azure. After relatively little time Steve was able to put together a Machine Learning prototype using Azure's tools and created a demo for two of our clients.
One tool was an image recognition tool. He applied it to matchday images from Bristol City Football Club. With minimal training it was quickly able to learn to correctly identify players from different teams and tag those images appropriately with each team name with surprising levels of accuracy.
Another he created was a recommended products tool. He applied this to products from Beyond Retail’s website, Tap Warehouse. Again, with minimal training, the tool started to recommend relevant products. It wasn’t perfect but it was clearly doing something intelligent. We already had a system in place for recommendations based on a series of logic – but you would get occasionally get obscure recommendations back, that didn't really fit with the product. After a considerable period of data collection, the machine learning system was giving us back far more relevant products. A black shower recommended other black showers, and similar style items from similar ranges, similar price, similar styles etc.
Beyond Retail loved it.
Turning a prototype into a full production-ready feature took some more effort. It required tweaking of the system to train the system to make smart recommendations; as on occasion it was still capable of giving odd results. It also took a bit of convincing within Beyond Retail that algorithms could do a smarter job than the existing logic-based approach or real people manually curating recommendations. In fact to prove that machine learning would provide smarter recommendations than the old system we decided to A/B test it.
Machine Learning won.
Here are some stats:
- The Machine Learning Recommended Products click rate is up 104.26%
- The ML variant's conversion rate is up 0.83%,
- Average order value is up 2.25%,
- Items per order is up 0.61%
Some of these numbers may look small, but on a busy ecommerce website, the aggregated effect of these incremental improvements is significant. Machine Learning had more than paid for itself well before we even ended the A/B test. Machine Learning has been rolled out onto all three of Beyond Retail’s websites
Machine Learning is an exciting tool. It has many applications. Product recommendations are an obvious candidate as is image recognition. It is something that can be tweaked and improved with manual intervention, telling it to weigh some metrics more heavily than others and of course, it loves data – the more you give it both in types and quantity the smarter it gets. It has the power to automate away tedious tasks and to improve your conversion and AOV.
What could Machine Learning do for your website? Get in touch if you'd like some ideas