Qianchuan Unified Recommendation Framework: Architecture, Challenges, and Algorithmic Solutions

Qianchuan is a unified recommendation platform designed to serve numerous long‑tail scenarios (channels, venues, post‑purchase flows, brand walls, etc.) that have small individual traffic and diverse characteristics. It aims to aggregate these fragmented scenes into a single system for continuous optimization and revenue growth.

The platform faces five core difficulties: varying product sets across scenes, inconsistent scene goals, large strategy differences, diverse recommendation types, and limited resource capacity. Solving these requires flexible product selection, multi‑goal support, differentiated strategy configuration, multi‑type recommendation capability, and low‑cost, high‑efficiency extensibility.

Qianchuan’s architecture is organized into five layers: (1) APP Service Layer – connects various tail scenes; (2) Qianchuan Access Layer – provides two integration methods and assigns scene‑specific Qianchuan IDs; (3) DPP Layer – offers multiple recommendation modules (product, multi‑type, floor, brand) with a common framework; (4) Algorithm Layer – handles the full recommendation pipeline (recall, coarse ranking, fine ranking, strategy); (5) Infrastructure Layer – leverages machine‑learning platforms, indexing, feature services, and traffic‑control systems.

In the algorithm layer, recall designs five categories (I2I, U2I, etc.) to address behavior bias; coarse ranking employs dual‑tower and ESMM models to differentiate click‑only and click‑plus‑conversion goals; fine ranking improves user interest modeling and multi‑scene differentiation; the strategy stage provides interventions such as flow control, diversity re‑ranking, and multi‑type distribution.

Key iterative developments include: 4.1 Recall and coarse‑ranking evolution – handling behavior bias with I2I and vectorization, modeling interest bias via DSSM and MIND embeddings, and addressing goal differences with dual‑tower and ESMM models; 4.2 Ranking model evolution – adding transformer‑based long‑short term behavior modeling, explicit feature crossing, MMoE and POSO structures to capture scene and user group differences, and enhancing feature engineering for brand, category, and activity dimensions.

Future outlook focuses on (1) business support – expanding scene coverage, refining common requirements, and strengthening system stability; (2) algorithmic advances – further bias reduction, richer feature usage, whole‑screen re‑ranking, continuous learning, and real‑time ODL/LTR techniques; (3) generality – moving toward componentization, algorithm libraries, flexible expansion, and reproducibility.