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

The article introduces OPPO's advertising recall algorithm, starting with a background overview that highlights limitations of the old recall pipeline—namely performance constraints and suboptimal placement of personalized recall after truncation.

A new recall architecture is described, featuring the integration of Approximate Nearest Neighbor (ANN) search to enable full‑inventory personalized recall and a multi‑path recall mechanism consisting of a primary "consistent" LTR‑based path and several auxiliary paths such as ECPM, cold‑start, and other specialized branches.

The primary recall model's objectives are broken down into three dimensions: consistency with downstream scoring, generalization across unseen data (both common and individual patterns), and diversity to avoid recommendation echo chambers.

Model selection is examined through a review of YouTube's 2016 paper, leading to three recall paradigms—precise value estimation, collection selection, and classification learning. After weighing trade‑offs, OPPO adopts the collection‑selection approach for its main recall.

Offline evaluation metrics are constructed in three stages: an initial simple split yielding overly high AUC, a full‑library Faiss‑based evaluation using GAUC and Recall with carefully chosen K and N parameters, and a segmented sampling evaluation that categorizes negative samples into Easy, Medium, and Hard to enable finer analysis.

Sample optimization practices are detailed, including a bid‑price sensitivity model that dramatically improves price responsiveness, incorporation of hard negatives to boost diversity, and automated mining of medium negatives via in‑batch negative sampling and pointwise loss.

Large‑scale multi‑class classification techniques are explored, comparing Negative Sampling (NCE) and Sample Softmax methods; experiments show Sample Softmax with temperature scaling yields the best performance.

Further model refinements cover dual‑tower interaction enhancements such as SENet for richer feature weighting, DAT for earlier interaction, and implicit feature sharing for semantic tags.

Generalization improvements address mixed‑distribution challenges using expert‑based gating (CDN) for cold‑start ads and PPNet for multi‑scenario personalization, demonstrating measurable ARPU gains.

The article concludes with a forward‑looking outlook on extending ECPM support, adapting to evolving ad productization, and maintaining the core role of recall in surfacing high‑value ads.