LLM-Enhanced Topical Trend Detection at Snapchat

This paper presents the first production-scale, end-to-end system for detecting topical trends on Snapchat, which integrates multimodal extraction, time-series burst detection, and LLM-based enrichment to significantly improve content freshness and user experience through global deployment.

The Gist

Imagine Snapchat as a massive, bustling digital city where millions of people are constantly posting short videos. In this city, "trends" are like sudden, massive street festivals or viral dance crazes that pop up out of nowhere. The challenge for Snapchat is that this city is too big and moves too fast for any human team to watch every corner and spot these festivals as they begin.

This paper describes a new, automated "City Watch" system built by Snapchat engineers to spot these trends the moment they start, using a special kind of artificial intelligence called a Large Language Model (LLM).

Here is how the system works, broken down into four simple steps:

1. The "Eyes and Ears" (Topic Extraction)

First, the system needs to understand what is happening in the videos. Since videos are a mix of pictures, sounds, and text, the system uses a team of AI "detectives."

• The Visual Detective: Looks at the video frames to see what objects or scenes are there (like a dog, a beach, or a concert).
• The Audio Detective: Listens to what people are saying (speech-to-text).
• The Text Detective: Reads any words written on the screen or in the caption.
• The Summarizer: Once these clues are gathered, a powerful AI (the LLM) acts like a skilled journalist. It takes all that messy information and writes a short, clear headline for the video, like "People dancing to a new song" instead of just a list of random words.

2. The "Surge Detector" (Burst Detection)

Just because a video is about "dogs" doesn't mean it's a trend; people post about dogs every day. The system needs to know when something is suddenly popular.

• Imagine a quiet street where usually 5 people walk by an hour. Suddenly, 500 people show up in the next hour. That's a "burst."
• The system tracks how many unique people are posting about a specific topic. It ignores how many people are watching (to avoid bias) and focuses on how many are creating.
• It uses a mathematical formula to compare today's numbers against the recent past. If the number of creators spikes significantly, the system flags it as a potential trend.

3. The "Quality Control" (Post-Processing)

Not every spike is a good trend. Sometimes a spike is just a glitch, spam, or something too vague like "funny videos."

• The Filter: The system uses AI rules to throw out bad topics. It removes anything that is too broad (e.g., "life") or anything that violates safety rules (sensitive or unsafe content).
• The Merger: Sometimes the system spots "World Cup 2026," "World Cup," and "World Cup qualifiers" as three separate things. The AI realizes these are actually the same event and merges them into one single, clean trend called "World Cup 2026." This keeps the list tidy and easy to understand.

4. The "Storyteller" (Trend Enrichment)

Once a trend is confirmed, the system doesn't just give it a name; it builds a profile for it.

• It picks a few representative videos from the trend and asks a super-smart AI to write a summary, assign a category (like "Sports" or "News"), and list which countries are talking about it most.
• Think of this as turning a raw data point into a polished news card that the rest of the app can use.

Why Does This Matter? (The Results)

The paper reports that this system has been tested and is now running globally on Snapchat.

• Accuracy: When humans checked the system's work over six months, it was correct about 92.8% of the time.
• Real-World Impact: The system is now used to help decide what videos appear on your screen (ranking) and what suggestions appear when you type in the search bar.
• The Outcome: Because the system spots trends faster, users see fresher, more relevant content. The tests showed that users liked the content more (higher "like" rates) and spent more time watching stories that were part of these new trends.

In short, this paper describes a smart, automated way for Snapchat to listen to the "buzz" of its entire user base, filter out the noise, and instantly tell the app, "Hey, everyone is talking about this right now—let's show it to people!"

Technical Summary

1. Problem Statement

The paper addresses the challenge of automatic topical trend detection at scale on short-video social media platforms, specifically Snapchat. Unlike traditional text-based networks, short-video platforms present unique difficulties:

• Multimodal Complexity: Content consists of video, audio, and text, requiring scalable systems to process heterogeneous data.
• Subtle Signal Detection: Trends often emerge through weak signals distributed across communities, making early detection difficult.
• Rapid Lifecycle: Trends evolve quickly, requiring near real-time adaptation while maintaining robustness against noise, spam, and ambiguity.
• Scale: The system must handle the massive volume of daily content creation on a global platform.

The goal is to identify emerging topics characterized by noticeable increases in content creation to improve content surfacing, personalization, and creator participation.

2. Methodology

The authors propose a scalable, end-to-end system powered by Large Language Models (LLMs) and Vision-Language Models (VLMs). The pipeline operates in four distinct stages:

A. System Overview

The system runs on a regular cadence to ensure freshness. It processes multimodal signals from "Snaps" (user videos) and publisher content.

B. Step 1: Multimodal Topic Extraction

• Input Processing: Videos are processed by lightweight multimodal models to generate textual signals:
  • Visual: A VLM (e.g., BLIP2) analyzes sampled frames to produce visual tags from a fixed taxonomy.
  • Audio: Automatic Speech Recognition (ASR) generates transcripts.
  • Text: Optical Character Recognition (OCR) extracts on-screen text.
• Unification: These signals are combined with user-provided captions and hashtags into a single textual representation.
• Summarization: A text-only LLM summarizes this representation into free-form topical phrases or entities (rather than selecting from a fixed taxonomy), creating candidate topics for downstream analysis.

C. Step 2: Time-Series Burst Detection

• Metric: The system tracks the number of Unique Users (UUs) posting about a specific topic over time. Viewer-side signals are excluded to avoid bias from ranking algorithms.
• Algorithm: A multi-scale burst detection algorithm is applied:
  1. Pre-filtering: Topics with fewer than $M$ UUs are removed to reduce long-tail noise.
  2. Moving Maxima & Averages: A sliding window calculates the moving maximum of user counts, followed by a moving average of these maxima to establish a robust baseline.
  3. Lift Calculation: A "lift score" is computed as the ratio of current activity to the smoothed baseline.
  4. Trend Scoring: Lift values across multiple temporal windows are aggregated using a weighted harmonic mean. This emphasizes recent activity while penalizing weak signals, ensuring only sustained, multi-scale growth receives high scores.

D. Step 3: Post-Processing (LLM-Enhanced)

This stage refines candidates to ensure quality and compliance:

• Sensitive Filtering: LLMs detect and remove topics violating safety policies.
• Generality Filtering: LLMs remove overly broad topics (e.g., "funny videos") that lack specificity.
• Precision Control: Adaptive thresholds on trend scores and user levels filter results.
• Topic Consolidation: An LLM clusters semantically similar topics (e.g., "World Cup 2026" and "World Cup") and selects the most representative canonical form to eliminate redundancy.

E. Step 4: Trend Enrichment

To make trends actionable for downstream systems:

• MLLM Analysis: A Multimodal LLM (e.g., Gemini 2.0 Flash) analyzes a representative set of videos associated with the trend to generate concise descriptions.
• Synthesis: A text-only LLM aggregates these descriptions to produce a human-readable summary, extract key details, assign a canonical category (e.g., Sports, News), and generate structured metadata (e.g., top countries, detection time).

3. Key Contributions

1. First Production-Scale End-to-End System: This is the first published system for topical trend detection on short-video platforms at production scale.
2. Hybrid Architecture: A novel integration of multimodal topic extraction, time-series burst modeling, and LLM-based consolidation/enrichment.
3. Efficient LLM Utilization: A strategy that balances semantic accuracy with computational efficiency by using lightweight models for full-scale extraction and heavier MLLMs only for enrichment on a subset of data.
4. Comprehensive Evaluation: Demonstrates high precision through continuous offline human evaluation and measurable improvements via online A/B testing.

4. Results

Offline Evaluation

• Precision: Over a six-month period (July–December 2025), independent annotators reviewed 1,278 trends. The system achieved an overall precision of 92.8%.
• Sensitivity: A trend score threshold of 1.8 was selected for production, optimizing the trade-off between precision and coverage.

Online A/B Testing

The system was deployed globally and integrated into content ranking and search. Key improvements included:

• Content Ranking:
  • Spotlight Story Views (US): +0.86%
  • Content Like Rate: +11.5% (indicating strong user affinity for timely content).
  • Content Timeliness (<1 day): +1.89%
• Search:
  • Impression Timeliness (<3 days): +2.38%
  • Popular Account Open (US): +26%
  • PostType Abandonment Rate: -1.28% (reduced).

5. Significance

This work demonstrates the practical viability of combining Large Language Models with traditional information retrieval and temporal modeling in a high-scale, real-world environment.

• Platform Impact: It enables Snapchat to dynamically surface fresh, culturally relevant content, directly improving user experience and creator engagement.
• Research Contribution: It provides a blueprint for handling the heterogeneity and noise of short-video data, showing that LLMs can effectively bridge the gap between raw multimodal signals and structured, actionable trend intelligence.
• Scalability: The architecture proves that LLMs can be deployed cost-effectively at global scale by strategically partitioning tasks between lightweight extraction and heavy enrichment.