Two‑Level Store Recommendation and Experience Optimization in Taobao’s Daily Good Store

Daily Good Store is a classic Taobao product that centralizes high‑quality shop discovery for millions of sellers and buyers. It provides three tabs—Headline, Featured Street, and Shop Rankings—to help users find popular, niche, and scenario‑specific shops.

The technical challenge is a novel “two‑level recommendation” problem: recommending a set of shops while simultaneously recommending items within each shop’s window. Traditional two‑stage pipelines (shop recommendation → item recommendation) limit optimization space, so a dual‑link recommendation system was built to intertwine shop and item suggestions.

To evaluate and improve the system, a “scatter score” (NormedScatterScore, NSS) was introduced, measuring the average diversity of items (or shops) across positions in a feed. The metric normalizes the dispersion of similar content and serves as a key indicator alongside efficiency metrics.

Personalized scattering further aligns the distribution of recommended categories with each user’s long‑term preference distribution. An Earth Mover’s Distance (EMD) term penalizes mismatches between the session‑level category distribution and the user’s preference profile.

Beam search replaces the greedy search of the original MLR algorithm, enabling a more global optimum for the two‑level ranking problem. By caching repeated sub‑sequence scores, the approach remains computationally efficient while optimizing relevance, diversity, and personalization jointly.

Additional UI upgrades—such as richer shop window visuals, dynamic tags, and improved card layout—enhance the front‑end presentation, boosting both efficiency (PPS) and user experience.

Offline experiments and online A/B tests show modest gains in efficiency metrics and significant improvements in ecosystem health indicators after deployment.