OPPO's Unified Modeling for App Distribution: Balancing Cost Reduction and User Value

In the current industry climate, cost reduction and efficiency are paramount, but preserving user value remains essential; OPPO's senior manager Lai Hongke discusses how the company addresses this tension in its app‑distribution platforms.

OPPO faces a unique recommendation problem: the app‑distribution scenario spans many categories (finance, logistics, travel, e‑commerce, games, social) and suffers from data sparsity because user interactions are tool‑oriented and infrequent, leading to long conversion chains.

To overcome these challenges, OPPO implements a "full‑scene unified modeling" approach that shares global features and samples across all distribution scenarios, dramatically expanding its feature store from 2 TB to 30 TB and increasing feature count from 100 million to over 10 billion.

The unified model uses a Multi‑Task Mixture‑of‑Experts (MMOE) architecture, allowing shallow and deep conversion targets to share embeddings, reducing the number of CVR models while improving conversion rates, especially for deep targets like game payments.

OPPO also introduced the oCPX capability, an intelligent cost‑per‑action bidding system that lets advertisers optimize for specific goals (views, downloads, registrations, payments) by estimating conversion rates and automatically adjusting bids.

Beyond algorithmic advances, OPPO’s engineering improvements—dynamic traffic throttling, user‑level flow control, and VIP‑focused resource allocation—have boosted system efficiency, raising traffic support and revenue by up to 20 % while maintaining or improving user experience.

Overall, the integration of unified modeling, advanced bidding, and scalable infrastructure demonstrates how OPPO balances cost reduction with user‑centric value creation in its recommendation ecosystem.