Feed generation at ShareChat
summary
ShareChat serves 85M monthly active users and ingests about 1M new posts daily, so feed generation combines a popularity bandit, autoencoder long‑term embeddings, LSTM+attention short‑term models and graph nets for creator/location affinity. A Bayesian mixer weights outputs (they speed convergence by exposing 100 rankings to 200 randomized user clusters). Feature stores span BigTable/BigQuery/Redis, serving millions qps; deployments run on Kubernetes TF-Serve, Dataproc/EMR and Dataflow. Cold starts, sparse signals and millisecond latency remain practical limits.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:46ShareChat - Who we are?
- 1:25Key Performance Statistics
- 2:03Internet Pyramid
- 2:33Understanding next billion users
- 6:17to address the core need of next billion users
- 7:14Content on ShareChat
- 7:55Who creates the content?
- 9:24Feed Generation is Important!
- 10:02Feed Generation: Multiple Objectives
- 14:27Heterogeneous Information Network
- 16:26Multiple Response Types
- 17:37What is Important?
- 18:38Architecture
- 20:38Algorithms in Play
- 27:02Data Challenges
- 27:56Tech Stack
- 28:57Other problems at ShareChat
related talks
22mGenerating a Billion Personal News Feeds
03mServing Billions of Personalized News Feeds with AI
42mBuilding a near real time personalized feed using data streaming and Aerospike
42mScaling to Billions of Real-time Recommendations Daily
01h 22mWho needs ChatGPT? Rock solid AI pipelines with Hugging Face and Kedro
09mHigh Order Bit: The In
19mRushin Shah, Engineering Manager, Facebook
28mEmbeddings