Insights from 58 Group Technical Salon: Recommendation Systems, Image Creative Optimization, and Deep Learning Online Prediction Service

Background

On December 15, 2018, the 58 Group Technical Engineering Platform and HR Magic Academy co‑hosted the fifth 58 Group Technical Salon at the Beijing headquarters, featuring four speakers who shared experiences on recommendation system architecture, search ad image creative optimization, creative platform design, and deep‑learning online prediction services.

1. Video Recommendation System Architecture

Rapid growth of video content creates challenges for recommendation systems. The team responsible for the video recommendation platform described a micro‑service based architecture that isolates numerous modules and employs distributed inverted indexes for different video categories, while maintaining in‑memory forward indexes for high‑frequency queries, thereby improving retrieval efficiency. An asynchronous multi‑path recall mechanism is used to overcome single‑machine bottlenecks and ensure comprehensive recall results.

2. Image Creative Optimization in Search Advertising

Ad image creatives significantly affect CTR. The presented solution consists of offline and online stages: offline selection builds a candidate pool using historical CTR and manual labels with pointwise or pairwise ranking models; online, a CNN extracts image features which are combined with user and ad features for final CTR‑based ranking. For creative understanding, user query intent is classified, and offline image classification aligns images with the same category, enabling selection of the highest pCTR creative per keyword.

3. 58 Commercial Creative Platform Architecture

The platform addresses massive creative content generation by programmatically creating ads, improving efficiency and quality. It comprises a publishing side, a Creative‑Builder, dynamic creatives, and an effect data module. The Creative‑Builder extracts tags from user text, applies rule‑based templates, and generates images via Java AWT or complex tables via WebDriver + Chrome.

4. Deep Learning Online Prediction Service Architecture

The service provides a unified online inference capability for TensorFlow and Caffe models, built on Kubernetes and Docker to manage GPU/CPU resources. It consists of a web management layer (handling JAR uploads, hot‑loading, resource monitoring), an algorithm layer (supporting custom model images and gRPC communication), and a cluster management layer (load balancing, high‑availability scheduling, and comprehensive resource monitoring via ETCD, Heapster, etc.).

5. Summary

Recommendation systems and computational advertising technologies have broad applications within 58 Group and the industry, while the deep‑learning online prediction service underpins these use cases, enabling stable deployment of various deep‑learning models. Participants gained valuable insights across recommendation architecture, image creative optimization, creative platform design, and online prediction services, and look forward to further practical achievements.