How Precise Recommendation Systems Work: Architecture and User Behavior Mining

Precise Recommendation Project Background

In social networks each user is a node; in Weibo users follow others, creating directed edges. Information overload makes it hard for users to find relevant posts, so recommendation systems are needed to match user interests with content.

Technical Architecture

The recommendation system consists of four layers: data generation, storage, candidate filtering, and ranking. Data generation extracts user interests offline and stores them in user and post databases. The storage layer uses NoSQL databases such as Redis. The candidate filtering layer matches user interests with candidate posts. The ranking layer orders candidates using CTR and machine‑learning models such as Logistic Regression, FM, and GBDT.

Offline Data Mining

Four targeting methods are used: demographic, behavioral, geo, and look‑alike. The article focuses on behavioral targeting and look‑alike targeting.

Behavior‑Based Mining

User interactions (repost, comment, like, favorite, click) indicate interest levels. Text of interacted posts is tokenized (e.g., jieba), stop words removed, and TF‑IDF scores computed to extract core words, which become user tags stored in a database.

Interaction‑Based Mining with Authors

If a user has interacted with an author, the user is likely to engage with that author’s future posts. Historical interaction data are processed and stored for candidate selection.

Similar‑User Mining

Similar users are identified via @‑mentions, community detection, and collaborative filtering.

@‑Based Similar User Mining

Posts containing @ mentions are used to extract the mentioned users and their followers, forming a candidate set.

Community‑Based Similar User Mining

The social graph is treated as a directed graph; inbound and outbound edges are merged to create an undirected graph. Label Propagation with weighted edges (α=0.6, β=0.4) discovers communities, after which low‑entropy (inactive) communities are filtered out.

Collaborative Filtering

User‑based collaborative filtering computes similarity between users based on overlap of follower sets, then recommends items liked by similar users.

When a post needs to be delivered, it is targeted to selected active communities or similar users based on matching post topics with community tags or the author’s community.