Evolution, Algorithms, and Architecture of Qingteng FM's Recommendation System

Recommendation systems are now ubiquitous in the Internet, powering experiences on platforms such as TikTok, Taobao, and news feeds. Qingteng FM, the first domestic internet audio media platform, shares its ten‑year journey of building and refining its recommendation system for audio content.

Homepage Recommendation Evolution

1. Manual Recommendation Phase – Early on, the homepage was composed of modules manually maintained by operators. Only the personalized module used algorithmic recommendations; other modules required frequent manual updates, resulting in low efficiency.

To improve operational efficiency, a strategy‑based recommendation layer was introduced.

2. Strategy Recommendation Phase – Operators shifted from daily content updates to binding content libraries and selecting appropriate sorting strategies for each module. This reduced manual effort but introduced new challenges in module ordering and strategy selection.

3. Personalized Recommendation Phase – Data showed that personalized modules outperformed strategy modules. The number of items in the personalized module was increased from 3 to 6, and multiple modules were merged into a single personalized feed. AB tests confirmed higher metrics for the feed, and a double‑row layout was chosen for optimal performance.

Even with personalization, operators still support certain albums through a dedicated promotion system that generates multiple creatives (titles, covers) and promotes the best‑performing ones.

Algorithm Overview

1. Recommendation Pipeline – Content that meets recommendation standards (hundreds of thousands) enters the recall layer, which selects a few thousand candidates. The coarse‑ranking layer reduces this to a few hundred, and the fine‑ranking layer selects dozens for the re‑ranking layer, which finally orders the results while ensuring diversity.

2. Multi‑Channel Recall – Three recall types are used: content‑based (hot, attribute, new‑item strategies), collaborative filtering (User‑Based, Item‑Based), and embedding‑based (Word2Vec, BERT). Effective recall balances coverage, complexity, and business goals such as cold‑start and traffic support.

3. Coarse Ranking – Initially simple fusion strategies were used; later a dual‑tower model was adopted to efficiently combine multi‑channel recall results while reducing online computation.

4. Fine Ranking – Progressed from linear models (Logistic Regression, FM) to tree models (XGBoost) and finally deep models (DeepFM). DeepFM became the main model, delivering a 9.3% improvement in listening metrics.

5. Re‑ranking – Focuses on diversity using algorithms such as Maximal Marginal Relevance (MMR) and Determinantal Point Processes (DPP). MMR, with a tunable λ parameter, achieved better diversity‑accuracy trade‑offs, increasing average album exposure by 8.84% and category exposure by 7.06%.

System Architecture

The recommendation system follows a three‑layer architecture: offline (data processing, model training, reporting), near‑line (real‑time feature processing, recall, coarse ranking), and online (user request handling, fine ranking, re‑ranking, and delivery).

Model deployment uses a Scala Play framework for feature extraction shared between Spark (offline) and Play (online), ensuring consistency. The prediction service supports multiple models, versions, and automatic updates, improving deployment efficiency.

Future Outlook

Future work includes improving new content exposure, integrating more business scenarios (albums, live, playlists, programs, radio), addressing cold‑start for new and dormant users, and advancing real‑time model training and multi‑objective ranking.

Recruitment Notice

Qingteng FM’s Intelligent Operations team is hiring recommendation algorithm engineers and Golang developers. Interested candidates can email [email protected] to discuss opportunities directly with the technical lead.