WeChat Hotspot Mining Platform: Architecture, Detection, and Presentation

With the rapid growth of self‑media, WeChat publishes millions of articles daily, creating information overload; quickly and accurately mining news hotspots is essential for helping users understand emerging events.

The platform integrates WeChat public articles with Tencent News, Tencent Video, and external media, leveraging both content and user behavior to construct hotspot resources, focusing on hotspot identification, display, and active learning.

A quality model extends traditional low‑quality indicators (spam, porn, clickbait) with news‑specific metrics such as newsworthiness, tone, and universality, and adds account‑level quality, authority, and regional tiers to better assess news content.

Hotspot discovery builds on Topic Detection and Tracking (TDT) techniques, using both offline (K‑means, HAC, DBSCAN, AP) and online (Single‑Pass) clustering, similarity measures (cosine, incremental TF‑IDF/TF‑ICF), and a stacking‑XGBoost ensemble for classification; multi‑modal features (text, images, user behavior) are extracted with CNN, Gated‑CNN, LSTM, BERT and fused via XGBoost for hotspot prediction.

Topic aggregation relies on online clustering, hierarchical agglomerative clustering, and a topic refinement mechanism to maintain a pure topic pool.

Hotspot prediction models combine multi‑modal features with DNN and traditional machine‑learning models, using feature extractors for text, images, sequential user behavior (LSTM), and account information, followed by an XGBoost fusion to forecast article popularity early after publication.

For hotspot presentation, popular topics combine topic aggregation, tracking, short description generation, summarization, and columnar layouts to strengthen user perception and interaction.

Topic tracking and event timeline involve story detection and event detection, using multi‑granular matching; an event element graph (EEG) represents articles as graphs of event elements, and the GIM (Graph‑based Interactive Matching) model applies multi‑layer GCN with interaction‑based feature crossing and attention masks to compute similarity at story and event levels.

Short description generation uses a multi‑source pointer network with attention to fuse article titles and content, incorporates an event word bank and named‑entity information, and applies length and semantic penalties; pre‑training on user query logs further improves quality.

Active learning addresses the high cost of long‑text labeling by generating informative and diverse samples, combining uncertainty with diversity, and using sparse reconstruction to present concise keyword summaries for annotators, achieving strong results on multiple public datasets.

The summary recaps the framework, highlights challenges such as fine‑grained hotspot capture, event core‑element extraction, and super‑topic graph construction, and points to future research directions like event graph building and future event prediction.