Popularity Metric

What This Is

A unified popularity signal that aggregates across news volume, social media engagement, and celebrity coverage modelling the information cascade that turns an event into a cultural moment. The goal is a real-time metric that predicts sustained attention, not just momentary spike.

Popularity is not uniform. A story can trend on Twitter and die in 24 hours, or generate slow-building news coverage that sustains for weeks. These are different phenomena with different signals. We want to distinguish them.

The Signal Architecture

News frequency and sentiment We scrape headlines and compute: raw frequency, sentiment trajectory over time (rising negative? stabilising positive?), and source diversity (is this covered by one outlet or many?). A single outlet picking up a story is noise. Cross-source convergence is signal.

Social media engagement velocity Not raw engagement count, but the rate of change is attention accelerating or decelerating? A story with 100k mentions and falling velocity is less interesting than one with 10k mentions and rising. We model this as a differential equation on engagement: dE/dt as the primary feature.

Semantic coherence over time As a story evolves, the vocabulary used to describe it changes. We embed daily coverage and track semantic drift: is the narrative expanding (high drift) or consolidating (low drift)? Consolidating narratives typically peak before the next news cycle.

The Cross-Modal Fusion

Each modality gives a separate score. The fusion model learns which modality is most predictive for which entity type: celebrity coverage is dominated by social, policy stories by news, sports by both with a short lag. We use entity-type-conditioned attention for the fusion.

LLM Integration

We use an LLM to handle the parts that structured models can't: disambiguating entity names, inferring why a story is trending without explicit context, and generating human-readable summaries of the popularity dynamics.

What's Next

• Real-time data pipeline with streaming ingestion
• Predictive model: given current signals, will this story be relevant in 48 hours?
• Interactive dashboard with entity search and historical trend comparison