Evolution of OPPO Commercial Advertising Targeting: From Differentiated to Intelligent to Untargeted Practices

Introduction The article shares OPPO's commercial advertising targeting evolution—from differentiated targeting to intelligent modeling and finally to untargeted crowd recall—covering user profiling, non‑standard recommendation systems, and recall technology.

01. OPPO Commercial Background and Targeting Logic

OPPO commercial ads are delivered in four main scenarios: browser feed ads, app‑store search ads, app‑store recommendation ads, and alliance ads. Monthly active users exceed 450 million, with over 16.5 billion ad deliveries and more than 5 billion alliance ad requests per day.

The ad playback chain retrieves ads from a candidate pool built by the marketing platform, then passes them through retrieval, coarse ranking, fine ranking, and re‑ranking before returning the winning ad to the user. The focus of this article is on label‑based targeting in the retrieval stage.

02. Advertising Targeting System and Modeling Practice

The targeting architecture consists of four layers: data sources, data mining techniques, user targeting, and targeting recommendation/untargeted technology.

Data sources include advertiser‑provided data (conversion, creative), OPPO ad data (creative, behavior), and third‑party data (browser, app‑store, search behavior).

Mining techniques involve text classification, NLP, XGBoost, statistical modeling, recall, and data processing.

User targeting is divided into basic targeting (demographics, device, app behavior), interest targeting (behavior interest and industry interest), and DMP‑customized audience packs.

Interest‑based tags are recommended and combined with budget‑driven intelligent scaling. The article later details intelligent modeling and untargeted crowd recall.

Industry interest tags are built via statistical modeling of advertiser‑provided conversion signals, scoring users with a decayed time factor α and weighting of user, behavior, item, and tag attributes.

Industry‑specific tags are created by collecting positive (click/convert) and negative (impression without click) samples and training classification models. Four modeling stages are described:

Stage 1: Single‑machine XGB binary classification.

Stage 2: XGB + Spark for large‑scale prediction.

Stage 3: DNN binary classification with model fusion (wide & deep).

Stage 4: Multi‑objective modeling sharing layers to handle sparse industry features.

03. Automated Intelligent Targeting Modeling Practice

Background: progression from no targeting → behavior‑tag targeting (20‑45% cost reduction) → industry‑custom tags → tag recommendation.

Explorations include DNN binary classification, FM for tag embeddings, and DSSM twin‑tower models for ad‑tag vector interactions. Challenges remain: tag‑to‑tag dependence on user‑selected tags, data sparsity of ad‑tag pairs, and coarse granularity of tag audiences.

The OPPO model adapts the paper "Learning to Build User‑tag Profile in Recommendation System" and introduces customizations such as weighted IMEI‑tag scores, tag2vec embeddings, feature crossing, and a wide‑deep architecture that jointly optimizes ad‑tag and user‑tag scores.

Tag recommendation evolves through three stages:

Stage 1: Tag combination recommendation to scale audience.

Stage 2: CTR‑based scoring within a single tag for fine‑grained scaling.

Stage 3: Untargeted recall directly fetching high‑conversion users.

04. From Intelligent to Untargeted Crowd Recall

Smart expansion network uses a DSSM twin‑tower to compute ad and user vectors, applies OCPX cost control and budget capping, and adds an online user‑tower for inference, enabling crowd expansion.

Recall sample strategy differentiates between ad‑centric crowd recall (finding users for a given ad) and user‑centric ad recall, requiring careful negative sample construction.

DMP LookAlike iteration leverages twin‑tower embeddings with Spark LR for crowd expansion, simplifying complex features.

05. Future Thoughts and Summary

Commercial ad delivery trends move from CPC to oCPX smart bidding, demanding broader audience exploration; thus intelligent targeting and untargeted crowd recall become essential.

Targeting evolution follows regular → intelligent → untargeted, providing stable boundaries for audience expansion at each stage.

Algorithmic practice emphasizes problem‑driven solutions, flexible migration of industry solutions, and model reuse across scenarios (e.g., wide‑deep structures reusing industry‑specific multi‑objective outputs).

Today’s sharing concludes with thanks to the audience.