The battle of machine learning models
Boozt is constantly pushing to improve the shopping experience that we offer to customers. With more than 60.000 clothing items and 600+ different brands available on Boozt.com the options are endless. To make it easier for our customers to find the styles that they love we are exploring the best ways to use product recommendations on site.
With the product recommendations tool on Boozt.com, customers can see suggestions for other products that are most likely to be purchased given the product that is currently being viewed. At the time we introduced the machine learning based model, more than 3 years ago, we saw a significant uplift in product views, add-to-carts, and revenue generated from related items. Almost doubling the number of clicks and more than doubling the share of the revenue from products accessed via recommended items. Besides a few modifications, it is still the same model we use today for generating product recommendations on Boozt.com and Booztlet.com.
Exploring new technological solutions is key to elevate the user experience on our sites. We are always looking to see how we can improve. So when Google suggested we test their AI recommendation model we were very keen to see if it could be an option for us.
However, after more than a year of running the collaboration of their Recommendation AI the results were still inconclusive. The A/B tests we had been conducting, to compare Google’s Recommendation AI-based related product suggestions to our own method of product recommendations on site, had resulted in a draw.
The long journey of the back and forth battle of recommendation models that we went through is a good reflection of the complexity of building machine learning models. Although Google’s AutoML model showed promising results it was unable to outperform our original method of product recommendations on site. And even after several months of fixing bugs and initiating updates, we could still not figure out exactly why our simple original model was showing better results. Our educated guess is that the simplicity of our original model also makes it easier to train which is why it performs better when the optimisation parameters change.We are still exploring how we can continue to improve the shopping experience for our customers through more accurate product recommendations. Among other things we are exploring solutions within image recognition and visual tools as a way to suggest products that look similar to styles the customer has shopped for before.
