OPPO Advertising Recall Algorithm: Architecture, Model Selection, Evaluation, and Optimization

This article shares OPPO's practical experiences and explorations on advertising recall algorithms.

The main contents include: background introduction, main recall model selection, offline evaluation metric construction, sample optimization practice, and model optimization exploration.

Background : The old recall architecture performed directional filtering, truncation, and then personalized recall, which suffered from performance and latency issues. The new architecture introduces ANN (approximate nearest neighbor) to enable full‑advertisement personalized recall and adopts a "single main path + multiple auxiliary paths" multi‑recall mechanism. The main path uses a learning‑to‑rank (LTR) approach, while auxiliary paths such as ECPM and cold‑start branches improve diversity and fairness, resulting in an accumulated 15% ARPU increase.

Main Recall Model Selection : The goals are consistency, generalization, and diversity. Three possible model choices are precise value estimation, ranking learning, and classification learning. Based on online experiments, OPPO selected the "set selection" (collection selection) approach, which aligns the recall stage with downstream ranking objectives.

LTR Prototype Model : A dual‑tower architecture is used, where each pairwise sample consists of multiple instances. Positive samples are the top‑K ads from the ranking stage, and negative samples are randomly sampled exposure ads. The model is trained with a ranking loss.

Offline Evaluation Construction : Three stages are described. The first stage uses a simple train/test split, but the AUC is overly high (≈0.98) and not useful for iteration. The second stage employs full‑library Faiss retrieval with GAUC and Recall metrics, where Recall measures the overlap between top‑K ranking ads and the model's top‑N results. The third stage introduces segmented sampling evaluation, splitting negatives into Easy, Medium, and Hard groups to enable finer analysis.

Sample Optimization Practices : A price‑sensitive model (bid_part) was added to improve bid sensitivity from 5% to 90%, yielding a ~1% ARPU lift. Hard‑negative feedback was incorporated, and medium negatives were mined via manual rules. Model‑driven medium negatives were explored using in‑batch negative sampling and pointwise loss, which proved more effective than pure hard‑negative addition.

Large‑Scale Multi‑Class Formulation : Two solutions are discussed—negative sampling (NCE) and sample softmax. Sample softmax with a temperature coefficient consistently outperformed negative sampling in offline experiments.

Joint Training Across Scenarios : Three strategies are compared: fully independent models per media, fully unified training with shared negatives, and a hybrid approach where each media has its own ad tower while sharing a user tower. The hybrid method balances commonality and media‑specific characteristics and has shown performance gains.

Model Optimization Exploration : Improvements to dual‑tower interaction include SENet (feature‑level attention), DAT (early interaction with AMM and CAL modules), and implicit feature sharing via shared semantic‑tag embeddings. Generalization enhancements involve the CDN architecture (memory vs. general experts with gating) for cold‑start ads, and PPNet (personalized gating network) for multi‑scenario optimization.

Outlook : Future work will focus on strengthening the ECPM branch, adapting to ad productization and creative intelligence trends, and further refining the recall‑ranking division while maintaining the core goal of surfacing high‑value ads.

Q&A : The article concludes with a short Q&A covering algorithm complexity considerations, the impact of learning ranking in recall, offline Recall metric parameters (N and K), sample difficulty ratios, and the relationship between offline AUC and online performance.