Recommender systems (also called recommendation systems) are a powerful application of machine learning in ecommerce. They enable retailers to curate a personalised selection of products for each customer, based on each individual’s preferences. This is only possible through advances in machine learning, which can deliver accuracy and relevance at scale.
The benefits of getting this right are clear: Customer research from ecommerce company Barilliance found that “product recommendations account for up to 31% of eCommerce site revenues.”
A Salesforce report on personalised shopping finds that “shoppers that clicked on recommendations are 4.5x more likely to add items to cart, and 4.5x more likely to complete their purchase.”
Although technology is central to ecommerce recommendations, this process is really about people. Consumers appreciate receiving recommendations on a website, much as they appreciate personal recommendations from friends or family.
Resolving the paradox of choice
There are psychological explanations for this. A famous study in 2000 asked 50% of participants to choose from a selection of 24 jams, with the other 50% picking from a selection of just six jams.
Just 3% of shoppers bought from the larger selection. 30% made a purchase from the selection of six jams. This is known as ‘the paradox of choice’ or ‘choice paralysis’, and it should be familiar to ecommerce retailers.
The lesson: choice is good, but we can have too much of a good thing. Retailers must use customer and trends data to narrow the field for their audience, laying out a small, curated selection from which the consumer can make a simpler decision.
And to do that at scale, we need some assistance from machine learning recommender systems.
What are recommender systems?
Recommender systems guide users towards the products or content that they are most likely to prefer.
They achieve this by evaluating the relationships between items in the retailer’s inventory, between users and products, and between different users.
That is to say, a recommender system can detect similarities between different items in an inventory. For example, Amazon achieved huge success in the early 2000s when it implemented such a system.
In fact, the journal IEEE Internet Computing names Amazon’s 2003 paper,“Amazon.com Recommendations: Item-to-Item Collaborative Filtering”, as the piece of research that has “best stood the test of time”.
Amazon can then layer on its user data. This matches people with items they otherwise would not find in its endless digital shelf.
Even if we do not always see their inner workings, plenty of other big technology companies use recommender systems.
For example:
Google uses machine learning to recommend new apps to users and new videos to watch on YouTube.
40% of app installs on Google Play come from recommendations.
60% of watch time on YouTube comes from recommendations.
Spotify’s recommender system identifies similarities between the musical tastes of different users. It then uses this insight to suggest tracks to these individuals that they have not yet heard, but that similar users have heard and liked.
Many recommender systems perform the following functions:
Candidate generation: The system filters a large corpus of potentially millions of items down to a smaller subset that are worthy of further investigation. There may be multiple candidate generators, each using a slightly different set of criteria to whittle down the potential list of recommendations.
Scoring: The next stage takes the candidates from step one and narrows the range down to a set of items that the end user will see. This stage takes into account the user’s past interactions on the site, their declared preferences (through zero-party data), and their similarity to other users, among many other factors.
Re-ranking: This is an ongoing process, during which the system takes into account the success of the existing recommendations. For example, the user may click to dislike the recommendations or they may simply ignore them. This provides a signal that the recommendations can improve, so the system re-ranks the order and removes items the user does not wish to see. This should encourage longer-term engagement; the more the user interacts, the better their recommendations will be.
Recommender systems are also at their most effective when they are prominent in the onsite experience. Research finds that recommendation widgets placed above the fold on an ecommerce page are 1.7 times as effective as those below the fold.
How Cadeera helps retailers develop recommender systems
Cadeera builds bespoke recommender systems for retailers, covering the full pipeline from pre-processing to evaluation. Our ontology for the fashion and home decor industries also ensures that we embed domain expertise into our client’s recommendations.
Cadeera also puts the retailer in control, with dials to move different factors up or down within the recommendation logic. This means the customer receives thoughtful recommendations, not the blanket suggestions that off-the-shelf solutions rely on.
