Exploring and Practicing Personalized Recommendations at Kuakan Manga

Kuakan Manga, founded in 2014, operates a popular comic app with over 200 million users, primarily young, highly engaged readers. The platform combines long‑form comics and short‑form content (posts, videos) and aims to deliver personalized recommendations across its homepage, discovery, world pages, and related‑content sections.

Technical Challenges include handling varied content formats, capturing continuity and multiple interest points in long comics, and dealing with fragmented user attention for short content. Additionally, understanding comic images and community‑specific language poses difficulties for both image and text analysis.

Content‑Based Recommendation relies on a rich tag system covering work attributes, user distribution, and creation themes. Early models used these tags to build item profiles and user interests, achieving a 35% increase in daily active user (DAU) reading frequency.

Collaborative Filtering was introduced later, employing item‑based, user‑based, and model‑based approaches. The team evaluated nearest‑neighbor libraries (Nmslib vs. Faiss) and selected Faiss for its GPU support and HNSW indexing. Real‑time user‑based CF was implemented using Faiss IndexIVFFlat and IndexIVFPQ pipelines.

Ranking Models combine multiple recall results. CTR prediction models evaluated include Logistic Regression (LR), Factorization Machines (FM/FFM), Gradient Boosted Decision Trees (XGBoost), and Deep Neural Networks (DNN). XGBoost was chosen for its balance of feature engineering effort and performance.

System Architecture follows a three‑layer design: an offline layer for batch feature engineering and model training, a near‑line layer for real‑time user profiling via Kafka and Flink, and an online layer for serving recall, ranking, and recommendations. Supporting tools handle tag weighting, result tracking, and monitoring.

A/B Testing Platform provides device, user, and traffic randomization, supports orthogonal and mutually exclusive experiments, and offers configurable metrics with significance analysis.

Results and Future Work show that after deploying recall and ranking models, DAU reading counts rose by over 30%. Ongoing challenges include improving feature extraction, addressing cold‑start issues, and exploring deep‑learning models for better content understanding.