Why do recommendation systems matter outside big tech?
Recommendation systems matter for any business with more products than a customer can browse, because what the system shows first largely decides what gets seen, bought and restocked.
Take a fictional online craft supplies shop, Willowmere Crafts, with about 4,000 products and a team of 15. The e-commerce lead, Hana, has a complaint from the buying team and a report from the agency that set up the 'You might also like' panel. Her notes say: 'same 40 products recommended on almost every page. new spring yarn range (220 items) has had almost no recommendation views in 3 weeks. returning customers get good suggestions, first-time visitors get random-looking ones. independent maker brands say they never appear.' She needs to explain to the founder what is going on before asking the agency for changes.
How do recommendation systems decide what to show?
Most recommendation systems use content-based filtering, collaborative filtering or a mix of the two. The difference is whether they compare items with items or people with people.
Content-based filtering
Recommends items similar to what a person already likes, using item features. A customer who bought chunky merino wool in teal is shown other chunky wools and teal yarns. It needs no data about other customers, but it only works as well as the item features someone has defined, and it rarely surprises people with something outside their usual taste.
Collaborative filtering
Recommends what similar people liked. If customers who bought a beginner knitting kit often went on to buy stitch markers and a pattern book, a new kit buyer is shown those. It can find unexpected but useful suggestions and needs no item descriptions, but it depends on a history of interactions.
Hybrid systems
Combine both, often with simple rules such as removing out-of-stock items, so each approach covers the other's weak spots.
What is the cold start problem?
The cold start problem is that a recommendation system struggles with new items and new users because it has no interaction history for them. Collaborative filtering in particular cannot place an item nobody has bought or clicked yet, which is exactly what happened to Willowmere's spring yarn range, and it has little to go on for a first-time visitor.
Common remedies are to fall back on content-based similarity for new items, to estimate a new item from others in the same category or from the same maker, to show a deliberately varied set of popular items to new visitors, and to ask one or two quick preference questions where that suits the shop.
How do feedback loops and fairness problems arise?
A feedback loop arises when the system learns from clicks on items it chose to show, so items shown often get more clicks, which makes them shown even more. Over time a small set of popular products crowds out everything else, as with Willowmere's 40 products, while items never shown never get the chance to prove themselves.
Fairness questions follow directly. If independent makers' products start with less history, the loop can keep them invisible regardless of quality, and customers lose choice. Ways to counter it include reserving a few recommendation slots for newer or less-exposed items, measuring exposure by maker and category rather than only total clicks, and refreshing the model's training data so it is not only learning from its own past choices.
What does a weak and an improved request to the agency look like?
A weak request says: 'The recommendations are not working, please make them better.' The agency may simply retrain on more of the same click data, which strengthens the loop.
An improved request says: 'Our panel appears to rely on purchase history only. Please (1) use product attributes, such as fibre, weight, colour and craft, to recommend new items until they have enough history; (2) show first-time visitors a varied set across categories rather than one bestseller list; (3) reserve two of eight slots for items with low exposure in the last 30 days; (4) report each month the share of recommendation views by maker and by product age; and (5) test changes against the current panel with a split of visitors, measuring orders, not only clicks.' Hana can now judge whether the work was done.
How do you check whether recommendations are working?
Check recommendations by looking beyond click-through to orders, returns and how widely exposure is spread across products and suppliers. A panel that lifts clicks on a few bestsellers while burying new ranges may be hurting the business. Compare changes with a control group of visitors who see the old version, because recommendations shown during a sale or a holiday will look better whatever the model.
What are the common mistakes with recommendation systems?
The most common mistake is optimising only for clicks, which rewards eye-catching items and the feedback loop. Others: launching a new range without content features so it suffers from cold start, judging success by comparing before and after without a control group, ignoring which suppliers or categories never appear, and using personal data for recommendations without telling customers in the privacy notice.
Where does this lesson sit in Tosheo.ai?
This lesson is part of the Machine Learning category in the machine learning course, after the lessons on how models learn and are evaluated. In Tosheo.ai the hands-on task asks you to write a brief for improving a fictional shop's recommendations, and the rubric feedback checks for cold start, feedback loops, fairness and a fair test. It is a professional foundation for briefing and judging this work; building a recommender in code is part of the optional Code track.