Strategies for Reducing Cost and Improving Efficiency in Recommendation Systems with Alibaba Cloud PAI‑Rec

The presentation introduces strategies for cost reduction and efficiency improvement in recommendation systems, focusing on Alibaba Cloud's AI platform PAI‑Rec, which serves many enterprise customers facing similar challenges.

Key challenges include long development cycles due to complex algorithm pipelines, high training resource consumption, and costly online inference. EasyRec and TorchEasyRec provide componentized and configurable tools to shorten development time, lower technical barriers, and support GPU‑accelerated training.

Feature engineering consistency between offline and online environments is emphasized, with FeatureStore offering unified management of offline and real‑time features. The system supports up to 1,200 features, and feature processing operators (FG) are shared between SQL‑based offline jobs and C++‑based online services.

Model deployment follows a four‑week workflow: data processing, one‑click recall configuration, feature customization, engine training and deployment, and final online testing. Experienced users can launch new scenarios in just over two weeks.

EasyRec supports various data sources (OSS, MaxCompute, HDFS, Hive, Kafka) and multiple model types (Wide&Deep, DNN, transformer). Configuration is graph‑based, allowing flexible assembly of blocks such as transformer, DNN, and concat_blocks.

The PAI‑Rec engine handles core recommendation stages—recall, filtering, coarse ranking, and re‑ranking—through configurable components that can read from storage systems like Hologres. A/B testing and metric management are integrated for continuous optimization.

Performance benchmarks show that TorchEasyRec on two A10 GPUs reduces training time from ~1 hour (120 CPU cores) to about 10 minutes, and GPU inference cuts overall cost by roughly 50 % compared to CPU.

Q&A highlights include details on training speed gains, supported feature counts, inference latency (typically <200 ms), and the use of graph neural networks for social‑graph‑based recall, which can improve click‑through rates by over 50 %.

Reference links are provided for the EasyRec codebase, documentation, and the PAI‑Rec product page.