Demographic Filtering
Before a system knows anything about what you have clicked, it can still know roughly who you are — age band, country, language, device. Demographic filtering recommends from that alone: find the people you resemble on attributes, and show what they liked.
How it works
The mechanics are deliberately simple. Users are described by a demographic profile, grouped into segments either by hand-drawn rules or by clustering, and each segment carries item preferences aggregated from its members' behaviour. A new user is assigned to a segment from their registration data and served that segment's list. Nothing about the individual's own interaction history is required, which is the entire point — and also the limit, since every member of a segment receives the same recommendations.
Why it survives
As a standalone recommender it is weak; as a cold-start fallback it earns its place. A brand-new user has no rows in the user-item matrix, so collaborative filtering has nothing to compute from, and the alternative to a demographic guess is the global popularity list. In a hybrid design it is usually the first stage of a cascade or a feature source folded into a ranking model, where country and device are simply additional inputs alongside behavioural features.
The caution
Recommending on demographic attributes means recommending on protected characteristics, directly or by proxy. Age and gender inferences drive what people are shown and can harden into stereotype, which is an algorithmic-fairness problem rather than an accuracy one. Production systems therefore keep demographics as one signal among many, never as the whole policy.
Demographic filtering recommends items by matching users on attributes such as age, gender, location, language or device, then serving what similar-profile users preferred. It needs no interaction history, which makes it a practical cold-start fallback and a useful signal inside a hybrid recommender. Accuracy is coarse, since everyone in a segment sees the same list, and the reliance on protected attributes raises fairness concerns.
Picture a bookshop assistant who has never met you and knows only three things: your age, the city you live in, and the language you read in. They can still make a decent guess, because they know what other people like you have enjoyed. That guess is demographic filtering.
It exists for the awkward first moment. On the day you join a service, it has no idea what you like, so it leans on who you resemble instead. The weakness is obvious: everyone your age in your city gets the same suggestions, and the method can reduce you to a stereotype. It works best as a temporary stand-in until your own choices give the system something better to go on.