NetEase Cloud Music Advertising System: Algorithm Practice and Model Evolution

The presentation introduces NetEase Cloud Music's advertising system, highlighting its massive MAU (>100M), dispersed ad slots, and weak recommendation nature, which creates a core problem of balancing advertiser ROI with media revenue.

Core Challenges : Advertisers seek precise targeting and low cost, while media aim for higher click‑through rates and revenue through auction mechanisms.

Economic Models : Comparison of GFP (generalized first‑price) and GSP (generalized second‑price) auctions, emphasizing the need for machine‑learning‑driven CTR estimation to achieve optimal ROI.

CTR Prediction :

Problem: Predict click probability to guide bidding.

Data: Positive samples are clicked impressions; negatives are non‑clicked.

Model history: LR → FM/GBDT → DNN variants (Wide&Deep, DeepFM, DCN, DCN+Wide+Attention).

Key improvements: sample selection, handling imbalance, feature crossing, activation‑function change (ReLU → PReLU), learning‑rate decay, large batch size.

CVR Prediction :

Multi‑objective optimization (oCPC) balances conversion rate with cost.

Unified vs. separate modeling of CTR and CVR; calibration using bucketed PCVR.

User Vector Modeling :

Cold‑start for new ads addressed via Lookalike expansion of seed users.

Methods: manual tags, binary classification, clustering, collaborative filtering, and recent dual‑tower (DSSM) models.

Dual‑tower architecture learns separate user and ad embeddings; cosine similarity measures relevance.

Similarity scoring uses clustering of seed vectors, weighting by historical CTR, and max‑weighted cosine similarity.

Practical Recommendations :

Focus on business‑specific data characteristics when optimizing models.

Ensure data reflects true user intent.

Maintain communication with peers to broaden perspective.

Overall, the talk emphasizes that human behavior is highly predictable (≈93%) and that algorithmic improvements should respect both individual stability and collective patterns.