Deep Tree Matching (TDM): Evolution and Practice in Large-Scale Retrieval at Alibaba

The match stage of modern recommendation systems must retrieve a small set of relevant items from billions of candidates under strict latency and compute constraints, typically using a two‑stage funnel of recall (match) followed by ranking.

Traditional two‑stage recall, such as item‑based collaborative filtering, is simple but cannot incorporate advanced deep models. To overcome this, Alibaba introduced an inner‑product (dual‑tower) vector retrieval pipeline that transforms match into a single‑point scoring problem using user and item embeddings.

Building on this, the Deep Tree Matching (TDM) framework aggregates all items into a hierarchical interest tree, enabling a one‑stage retrieval that reduces the number of computation steps from billions to a logarithmic depth (e.g., 30 levels for a billion items). The tree structure is designed to support arbitrary complex models, and beam‑search is used to efficiently explore the tree while guaranteeing global optimality through a max‑heap property.

TDM evolved through three versions: TDM 1.0 introduced the max‑heap tree and beam‑search; TDM 2.0 added joint model‑index learning to align offline training with online retrieval; TDM 3.0 further integrated model, index, and retrieval optimization, addressing the mismatch between training objectives (multi‑class probability fitting) and retrieval goals (maximizing recall).

Key innovations include: (1) constructing training samples directly from beam‑search results to align data distribution; (2) redefining node labels to satisfy the max‑heap constraint, ensuring that locally optimal beam‑search decisions are globally optimal; (3) iterative training where each minibatch triggers a beam‑search to generate hierarchical samples, followed by a binary‑classification loss per node.

Extensive offline experiments on public datasets and online A/B tests in Alibaba's targeted advertising scenario show double‑digit improvements in recall over baselines such as ItemCF and YouTube‑DNN. The system also adapts to dynamic ad inventories by combining a static tree with real‑time inverted/forward indexes.

Future work aims to extend the tree structure with graph‑based approximations, incorporate multi‑objective metrics like diversity and explainability, and eventually merge match and rank into a unified pipeline for ultra‑large‑scale retrieval.