HiNet: A Hierarchical Information Extraction Network for Multi-Scenario Multi-Task Learning in Recommendation Systems

Recommendation systems in large platforms such as Meituan’s to‑restaurant service involve many heterogeneous scenarios (e.g., homepage feed, sub‑channels) and multiple business tasks (CTR, CTCVR). Modeling each scenario independently leads to duplicated development, data sparsity in long‑tail scenarios, and high maintenance cost, while naïvely merging all data ignores scenario‑specific information.

To address these challenges, the authors propose HiNet (Hierarchical Information Extraction Network), a two‑layer architecture that jointly extracts information across scenarios and tasks. The model consists of a Scenario Extraction Layer and a Task Extraction Layer, and introduces a Scenario‑aware Attentive Network (SAN) to weight the contribution of other scenarios, as well as a Customized Gate Control (CGC) module inspired by PLE to mitigate negative transfer among tasks.

In the Scenario Extraction Layer, three components are employed: (1) a shared‑expert network that captures common patterns across scenarios, (2) scenario‑specific expert networks for unique characteristics, and (3) the SAN module that computes attention weights for cross‑scenario information based on scenario embeddings. The outputs of these components are combined to form a rich scenario representation.

The Task Extraction Layer builds on the scenario representation and uses a CGC module that contains task‑shared and task‑specific experts. The CGC gate learns a weighted combination of expert outputs, allowing each task (e.g., CTR, CTCVR) to benefit from both shared knowledge and task‑specific signals while reducing parameter interference.

The overall training objective is a weighted sum of losses for all tasks across all scenarios, with hyper‑parameters controlling the relative importance of each task’s loss.

Extensive offline experiments were conducted on six real‑world scenarios from Meituan’s to‑restaurant platform. The authors compared HiNet against strong baselines such as Shared Bottom, MMoE, PLE, HMoE, and STAR. Results (Table 2) show that HiNet consistently achieves higher AUC for both CTR and CTCVR across all scenarios.

Ablation studies (Table 3) demonstrate that removing the hierarchical structure or the SAN module leads to significant performance drops, confirming the necessity of both components.

Online A/B tests in two production scenarios further validate HiNet’s effectiveness: the model yields notable improvements in CTR, CTCVR, and order revenue compared with baseline models (Table 4).

In conclusion, HiNet provides a unified framework for multi‑scenario multi‑task recommendation, achieving superior offline and online performance. Future work will explore integrating graph neural networks to further enhance cross‑scenario information propagation.