The effects of Machine Learning on your site's search bar
Published on May 2019 by Arjan Franzen
Personally, I don’t like the so-called “Silicon Valley Inspirational trips”. Don’t get me wrong: they sound like fun but going there just for inspiration seems a bit forced. Then: what do you call it when you happen to be in San Francisco, you are in the software business and you get ‘inspired’? 😉
This is what happened to me when I was in San Francisco in April 2019 to give a talk about Maxeda, Cloud Transformation and all the cool things we did there. “Since we’re already in SF” we decided to meet up with the cool team from Constructor.io. (see photo)
Constructor.io. is a Search ML startup based in downtown San Francisco. Their products enable e-commerce platforms to provide a better search experience for their users. Constructor.io. uses Machine Learning to train the site on what certain products ‘mean’ when they search for keywords. Sounds promising, let’s see what they can tell us.
We had a great time and left knowing a lot more about why Machine Learning is crucial to e-commerce platforms. “be more like Google when you’re doing search” was inspiring to us: we faced challenges with a lot of manual work for the ±100k products that need to be online. Automating ranking based on customer behaviour is crucial because ‘human brain resources’ are always limited!
Update: Nov 2020. After a smooth implementation project, the new constructor.io search engine is powering this large regional DIY retailer’s online platform.
Picture Left (before situation): no Machine Learning (ML) is used in the search. Instead: Literal text search is used by the product search engine of the (now old) site. no good support for compounded words. Dutch (like German & Turkish; compound words. ‘glues them together’). In this example, we are looking for a Garden Shed.
This is 1 word in Dutch but is glued together out of 2 words: Garden and House. if you search for the 2 individual words you get very bad results. (Left). No sheds, just wall paper from a brand that has the 2 individual words as brand name, that fact ranks the wallpaper products higher.
When using Machine Learning, ‘the Machine’ (the site) learns to compensate for this human error. By observing the behaviour of users searching for products and then clicking on a certain product and buying certain products, the machine learns what words mean.
The result, the picture on the right (the after situation) is that you get the 100% match for the Garden Sheds.
A second scenario describes the ‘before’ and ‘after’ situations when searching for a hammer. On the left, the ‘before’ situation shows products where 'hammer’ is in the title of the product. safety hammer, wooden hammer and a drill. Obviously, these are not the products we need when we are looking for a hammer.
When implementing a Machine Learning search, the product in the picture on the right is good results for hammers.
Implementing Machine Learning product search does miracles for conversion, and user experience and saves a lot of time of Product Managers improving search queries by hand. In the old scenario finding better converting synonyms was manual work, highly error-prone and simply too much work with around 100k products online. Machine Learning Search is at the core of the world-beating success of Google (Since the early 2000s). It is soon the de facto standard in all search engines.
Supercharge your Software Delivery!
Implement DevOps with Agile Analytics
Implement Site Reliability with Agile Analytics
Implement Service Level Objectives with Agile Analytics
Implement DORA Metrics with Agile Analytics
