Building an Intelligent News Recommendation System Inside Sohu News App

What this post added

This post details how Sohu News integrated Milvus into their news recommendation system. It explains the use of semantic vector search with BERT-as-service for encoding news articles and user profiles, the workflow for generating user semantic vectors daily, and the insertion of these vectors into Milvus. A key technical contribution is the optimization strategy using MinHash and Jaccard distance to reduce semantic vector extraction time from days to hours. Additionally, it describes a method for resolving short text classification errors by querying Milvus with short text vectors to find similar long text articles and using a voting mechanism for reclassification. The post highlights Milvus's role in improving real-time performance and the efficiency of text classification.