Recommendation engines employ plenties of algorithmic methods (math-based instructions) with the purpose of generating product recommendations. Predictive algorithms constantly consider relationships, see the most important ones here, with a few examples in the brackets:
Item Hierarchy (Horse equipment store: You have just purchased a saddle therefore you will need spare straps for the saddle)
Content-Based Filtering with Item-Item Similarity (Netflix: movie “Falling Down” is similar to “Taxi Driver” therefore the user will likely watch it)
Attribute Based (Gamer e-store: If you like games set in medieval scenario with ambitious female characters, this 2 attributes will appear in the content and product recommendations the store is displaying to you)
Collaborative Filtering harnesses the similarity between User and User (Example from a cake shop - confectionary: : “who bought cake fireworks also bought numbers made from fondues”)
A simple example of collaborative filtering
The Collaborative filtering
Collaborative filtering is a method used widely in recommendation systems. It is predicting the taste of a user based on other users’ similar interests. These connections are built up from browsing different items in an eCommerce store, clicking on several products during a session, and eventually adding to items and purchasing together with other products. Collaborative filtering assumes if User 1 viewed/bought the same item as User 2, User 1 is likely to view/buy other items that User 2 did as well.
Let’s have a simple example of collaborative filtering with a store with 3-4 buyers only.
Alec has bought a sun umbrella and a beach mat, Brya has bought a sun protection cream. Cedric is currently browsing the store and put a sun umbrella in his cart. What to recommend to Cedric during his checkout? Cedric will be recommended by a beach mat, because he has indicated a similar interest as Alec.
Imagine these scenarios with not only 5 but 100 or 1000 products having relevant meaning to not 4 but 1000 customers and we can understand how many aspects a reco engine has to take into account. In addition, these user profiles and product attributes must be assessed and grouped in real time otherwise the customer can be lost in the cloud of irrelevant products very quickly.
Collaborative filtering is a well-spread algorithmic method running beneath the hood, and it can help you to understand your customers better from their behaviour patterns, but this is not the only way to deliver recommendations: content based filtering also plays a pivotal role.
The Content-based filtering
In the Content-based filtering method (also called product clustering) a set of data or a set of specific characteristics gets related (manually or by the AI) to a product (title, category, tag, price etc.) in order to identify the product and sort them to similarity groups.
An example: products having the “female” tag, will be more likely to be recommended with other products categorized as female:
Bikinis tagged as pink will be more likely to appear together with other pink clothes.
At the first hearing it doesn’t seem complex, but with scaling up content-based filtering also has to make numerous calculations.