EdgeRec: An Edge‑Computing Based Real‑Time Recommendation System

In the era of pervasive wireless usage, information‑feed recommendation faces severe latency problems because traditional systems rely on paginated client requests that trigger server‑side sorting, limiting the ability to react instantly to users' evolving interests.

EdgeRec addresses these issues by leveraging edge computing on mobile devices, enabling the client to perform real‑time user intent perception, on‑device re‑ranking, and instantaneous insertion of items, thus reducing dependence on cloud latency and allowing more responsive, personalized feeds.

The system architecture consists of two main modules: (1) Heterogeneous User Behavior Sequence Modeling, which captures both positive (clicks, purchases) and negative (exposures, dwell time) feedback across item exposure and item page‑view behaviors; and (2) Context‑aware Reranking with Behavior Attention Networks (BAN), which fuses item‑level context with real‑time user behavior using attention mechanisms to produce a locally optimized ranking.

Behavior modeling treats each user action as a <Item, Action> pair, encoding item sequences with GRU networks and action features with identity mappings, then concatenating them for a unified behavior embedding. The BAN reranker treats candidate items as queries and the encoded behavior items as keys/values, applying attention to prioritize items similar to those the user has recently interacted with.

Offline experiments and large‑scale online A/B tests (including a Double‑11 shopping event with 5 billion runs) demonstrate that EdgeRec’s click‑oriented reranking improves click‑through rate by 10 % and its conversion‑oriented reranking raises transaction value by 5 %, confirming the effectiveness of edge‑side inference and real‑time adaptation.

In conclusion, EdgeRec showcases the first successful deployment of edge‑computing for recommendation at scale, highlighting the potential for on‑device deep models to overcome cloud latency, enable personalized training per user, and open new avenues for intelligent edge applications.