Your 'Recommended For You' Feed Is About to Get Faster and Smarter

When Netflix, Spotify, or Amazon suggest something, they're really answering one question: is item A similar to item B? A popular way to answer that is called the "Swing score." Instead of just counting how many people liked both items, it looks at which people liked both — and quietly discounts the serial shoppers who like absolutely everything. Think of a friend who recommends every restaurant in town: their opinion counts for less than a picky eater's. That makes Swing recommendations unusually good, which is why big companies use it.

The problem is arithmetic. Working out a Swing score gets dramatically slower as more users touch an item — roughly squaring the work each time. Scale that to billions of interactions, as real apps have, and it becomes impractical. Companies have worked around this by cutting corners, which makes the suggestions noticeably worse.

Two researchers, Runhao Jiang and Renchi Yang, wrote a new set of algorithms, ASC and K-ASC. Instead of doing all that heavy math, they use a clever sampling trick: for popular items they sample smartly, for rare items they compute directly, and they can prove mathematically how close their answer is to the perfect one. In tests across eight real datasets, their approach was orders of magnitude faster while matching the quality of much slower methods. It even handled graphs with a billion connections.

For you, the payoff is subtle but real: recommendations that feel less random, less repetitive, and refresh quicker, while the app spends far less computing power — savings that could show up as cheaper services. The catch is that these are approximations, not exact answers, and this is a research paper. No app has announced it yet, so don't expect your feed to change tomorrow.