Recommendation systems on websites are very similar no matter where they appear. The key differences between these systems lie behind the basic functions and structures of the sites that use them. Here we will discuss a few special issues that may arise when dealing with recommendation engines on classifieds sites.
The cold start problem
Let’s start with a little overview of what happens when you visit a classifieds site. From the very first moment you step into a site, the recommendation engine faces the so-called “cold start problem”. The fuel for the recommendation engine is the data it collects from users: what users view and buy. But when there is no user history to rely on, the engine simply doesn’t know what to recommend. At this point recommendations are based on metadata.
Then you start browsing, which is when the recommendation system starts doing what it’s best at: analyzing your actions on the site and then showing you items you might be interested in, but didn’t search for. Everything you do on the site helps the engine create more and more accurate recommendations for you.
Ads just come and go
It’s no surprise that one of the most distinctive features of classifieds sites is a continuously, fast-changing list of items for sale. There are plenty of items in an incredible variety, with masses of new products added every single day. Unlike auction sites, classifieds sites do not show when an item is sold.
The huge amount of items alone requires a serious performance from the systems to create accurate recommendations in real-time. But when you take account of the items already sold but which are still in the sites’ inventory, the task of providing users with valuable recommendations gets even more difficult.
A very popular feature for most classifieds site seems to be the ability to organize the items according to regions. Users usually benefit from this, since they don’t have to spend a lot of money on travel or delivery costs and they might be able to see and handle the items before actually buying them.
However, segmenting their selections also brings one of the biggest disadvantages classifieds sites could face, as narrowing searches means fewer items offered to a user, and as a result, users will have less chance to find the item they want. If they don’t find anything that matches what they are looking for, they will simply leave the site unsatisfied, possibly never to return again.
A way to get past this issue is offering products from areas next to (or close to) where the user lives. However, this might be problematic as well. If the items that match the visitors’ taste are only available in far away areas, they might just leave the site.
The smartphone factor
It comes as no wonder that now, with smartphones hiding in the pockets or bags of almost everyone, the number of people visiting classifieds sites from phones or tablets is everything but negligible. 15% of bomnegocio.com’s users only ever access the site from smartphones.
Most classifieds sites already have a mobile-friendly layout, or have at least started developing one. Compare the size of a desktop screen to a smartphone screen. For a classifieds site, acknowledging this difference is vital. Since you can display more items on a desktop device, users will distinguish relevant items at ease.
But when you have a user visiting your site from their smartphone, you’re limited to showing only a handful of items at once. This is when you need recommendations to be as accurate as possible. If a user is not impressed by the results, they will simply leave your site. Having a reliable recommendation engine is key to leaping over these hurdles.
Hide the bad, show the good!
This brings us to a serious task classifieds sites and their recommendation systems face: deciding what is worthy of being shown (and recommended) and what should be hidden. What recommendation engines need to know for this is the quality and reliability of both the ads and the users who posted them. However, this is something that most recommendation systems are simply not capable of doing.
Our data miners and developers at Gravity R&D have managed to create an algorithm which can predict how trustworthy items and their advertisers are, based on metadata and user actions. This is not only useful for determining which ads should be shown to which users, but also to fend off even the slightest possibility of fraud. As with all motors, a recommendation engine should be primed to fire up quickly from cold, be smart enough to use its fuel efficiently and have the horsepower to take the driver where they want to go.
This article was originally featured on November 13, 2013, on the blog of Gravity Research and Development, the company behind Yusp. Gravity R&D is a technology expert serving omnichannel recommendations for major clients on 5 continents.