Ranking the AI Feed: A Practical Playbook for “Home” and “For You” - #2
A concrete recipe to rank Home and For You feeds for AI platforms—signals, justification strings, and safeguards against popularity spirals.
Feeds are where attention compounds—or evaporates. For AI platforms, a good feed is half relevance, half serendipity. Here’s a practical blueprint I’ve used to make Home and For You feel alive from day one.
The “Home” Feed: Relevant First, Serendipitous Second
Blend signals with explicit justification text:
Relevance & Authority
- Creator affinity: “You like /TheBloke” → their models/datasets rank higher.
- Topic affinity: “Trending in LLMs” (match to followed tasks/topics).
- Creator authority: prioritize uploads from trusted creators (likes, downloads).
Network & Trends
- Out-of-network, high-authority: “Suggested: /teknium uploaded a model”
- Community activity: “Actively being discussed in the Community”
- Geography: “Trending in New York/India”
Operational cues
- Deployment popularity: “Popular in Text2Text Inference Endpoints”
- Search momentum: “Trending in Search”
Each item gets a short chip explaining why it’s here (“Suggested · Based on your likes”). This improves dwell time and reduces suspicion of “black-box” feeds.
The “For You” Feed: Discovery by Design
For You is for exploration. On day zero, you can personalize it by enhancing onboarding (more below), but the ranking leans into:
- Collaborative filtering across users × creators
- Topic-level breadth (ensure coverage; don’t cluster only on LLMs if the user also picked CV)
- Diversity constraints (cap consecutive items per creator; rotate long-tail)
Guardrails Against Popularity Spirals
- Freshness windows (e.g., 48–72 hours) so new uploads get a fair shot.
- Diversity quotas for emerging creators.
- Decay functions on authority so “once-viral” isn’t “always-viral.”
What to Ship First
- Justification chips (low effort, high trust).
- Creator + topic affinity ranking.
- Out-of-network authority panes injected every N items.
- Diversity constraints in the ranker.
The result: a feed that feels personal, transparent, and lively—and that turns consumers into contributors.
The document below is a more detailed discussion of how Hugging Face can implement a feed that can draw on many different signals and become the home feed of choice amongst their large and growing user base.
At Koo, my team and I developed For You and Home feeds of content from diverse creators from across geographies and languages. We studied the feed algorithms of top social network platforms like Meta and Twitter and developed key insights from our own user base. The insights shared in these posts reflect many of the learnings my team and I had over the years building personalisation features at Koo.
