Content Mining and Recall Model Practices in the Quanmin K Song Recommendation System

Recommendation systems are typically divided into a recall stage and a ranking stage. In the UGC platform Quanmin K Song, content mining is essential to select high‑quality user‑generated works as candidates for the downstream ranking process.

1. Content Mining – Importance The platform has two types of users: content creators and content consumers. Creators want their works to be seen, while consumers seek enjoyable content. Solely focusing on short‑term personalization can concentrate traffic on a few creators, leading to creator churn. Therefore, a content‑centric mining mechanism is needed to maintain a healthy ecosystem.

2. Content Mining – Goals The goal is to continuously discover high‑quality content and build reliable traffic channels. This is achieved through two modules: a content admission mechanism that filters uploaded works, and a distribution channel that ensures selected works reach users.

3. Content Mining – Iterations Initially, manual scoring was used, which was slow and inefficient. Later, handcrafted features with models were tried but had limitations. The current solution uses deep neural network models that evaluate audio/video quality, combined with lightweight human review, to process millions of daily uploads.

4. Content Mining – Solution The pipeline consists of three parts: (i) content admission via a deep neural network trained on manually labeled good and bad samples; (ii) tiered distribution that gives initial exposure based on user level and scales up based on feedback; (iii) recall using a dual‑tower model that encodes only the item’s source information.

Recall Model Overview The recall system includes three categories: attribute‑based recall, model‑based recall (dual‑tower embeddings), and other recall methods such as real‑time hot items, location‑based services, and social‑graph recall.

Attribute‑Based Recall User preferences are computed in real time from platform-wide behaviors (search, play, like, share, etc.). High‑preference dimensions (e.g., tags, singers) are used to retrieve matching items, providing strong perceived relevance.

Attribute Vectorization To overcome the rigidity of pure attribute matching, attribute embeddings are learned so that similar attributes (e.g., similar accompaniments or stylistically related singers) can be expanded during recall.

Recall Model Practice – Model‑Based Recall Three typical architectures are used: embedding‑only matrix factorization, single‑tower with softmax (YouTube‑DNN), and dual‑tower with cosine similarity (DSSM). Quanmin K Song adopts the dual‑tower approach.

Challenges include a much larger search space than ranking, the need for efficient online retrieval (preventing the use of cross‑features), and the necessity of debias mechanisms to avoid feedback loops.

Negative Sampling In recall, random negative sampling works better than using only unclicked exposures, because the candidate set is the entire item pool.

Enhancing Single‑Tower Expressiveness Since dual‑tower models cannot use user‑item interaction features, stable user behavior sequences and basic profile information are leveraged to build richer user representations, which are then used to select dynamic signals (play, click) for more relevant recall.

Final Model and Optimizations After applying hot‑negative sampling, hard‑negative mining, and sequence‑attention mechanisms, the model achieved noticeable online gains.

Debiasing A bias loop was identified where high‑exposure items are over‑estimated. A global sample augmentation combined with propensity‑score weighting (IPS) was deployed to reduce the weight of high‑exposure positives and increase that of low‑exposure positives, resulting in measurable improvement.

Summary and Outlook Future work will focus on finer‑grained traffic control for faster cold‑start of quality content, replacing statistical real‑time profiling with model‑based profiling, and incorporating multi‑objective considerations into the recall stage.