Social Feed Ranking: Model Architecture

With a rich feature vector assembled from batch and streaming pipelines, spanning privacy-safe user embeddings, real-time engagement counters, and cross features capturing viewer-poster affinity, the next design decision is the most consequential one. Given hundreds of heterogeneous features per viewer-poster-content triple, how do you design a single model that simultaneously predicts click-through rate, long-press dwell, reshare probability, comment likelihood, and creator equity signals? This is the exact question a Staff+ candidate faces in a MAANG ML system design interview, and answering it well requires more than listing model names.

This lesson covers three architectural pillars that together form a production-grade feed ranking model. First, multi-task learning with shared and task-specific towers enables joint optimization across competing objectives. Second, mixture-of-experts layers handle the radical differences between content types like video, text, and images. Third, loss function design directly combats metric cannibalization, where short-term click optimization silently erodes long-term platform health. The design goal is a unified architecture that balances immediate engagement with sustainable creator equity and user retention.

Multi-task learning for joint optimization

Social feed ranking is inherently multi-objective. The platform needs to predict $P(\text{click})$, $P(\text{reshare})$, $P(\text{meaningful comment})$, $P(\text{long dwell})$, and a creator equity score for every candidate post. A naive approach trains separate models per task, but this wastes compute and misses the opportunity for shared representation learning. Users who reshare content also tend to click on it first, so the latent representations useful for one task carry signal for others.

Shared-bottom and MMoE architectures

The simplest multi-task setup is the shared-bottom architecture, where a single shared embedding and feature interaction layer ...