Balancing Relevance and Recency in the LinkedIn Home Feed

Feed Relevance vs. Timeliness: The Core Engineering Trade-off

The problem isn't just showing content; it's showing the *right* content at the *right* time. Users often complain about feeds that are either too stale (missing breaking industry news) or too noisy (filled with low-value updates from distant connections). The useful takeaway is that LinkedIn's feed ranking is not a simple recency sort; it’s a multi-signal machine learning model designed to optimize for predicted user engagement, balancing relevance against time decay.

Signals Driving Feed Weighting

The ranking algorithm incorporates several publicly observable signals. These include connection strength (e.g., 1st vs. 3rd-degree connections), content engagement history (likes, comments, shares), and temporal decay (how long ago the post was made). The model assigns weights to these signals rather than applying a fixed rule.

Scoring Example: Weighted Engagement

To illustrate, consider a simplified scoring mechanism where a post score ($S$) is a function of connection proximity ($C$), engagement rate ($E$), and time decay ($T$): $S = w_C C + w_E E - w_T T$.

For a 1st-degree connection post with high recent engagement, $C$ is high, $E$ is high, and $T$ is low, resulting in a higher $S$. For a 3rd-degree connection post with low engagement, $S$ will be significantly lower, even if it is recent.

Practical Limitation: API Visibility

A key limitation is that the exact weights ($w_C, w_E, w_T$) are proprietary and not exposed via public APIs. You can observe the *effect* (what shows up) but not the *mechanism* (how it's calculated).

Actionable Insight

If you are optimizing content for your network, focus less on posting frequency and more on driving high-quality initial engagement (comments over likes). High early engagement signals the model to boost visibility, overriding simple recency bias.