Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Your decision path does matter in pre-training industrial recommenders with multi-source behaviors

Chunjing Gan, Binbin Hu, Bo Huang, Ziqi Liu, Jian Ma, Zhiqiang Zhang, Wenliang Zhong, Jun Zhou

TL;DR

This work tackles cross-domain pre-training for industrial recommender systems by recognizing that user decision paths strongly influence multi-source behaviors. It introduces HIER, a framework that fuses a knowledge-graph–based GNN encoder with hierarchical decision-path modeling, augmented by two contrastive-learning modules: exemplar-level CL for items and information bottleneck CL for users, plus an $L_0$ sparsity term for path selection. The learning objective combines $\mathcal{L}_{IB}$, $\mathcal{L}_{ECL}$ and $\lambda_3 \|\mathbf{Z}\|_0$, enabling robust, transferable representations from source domains and lightweight fine-tuning on target domains using universal user encodings and mean-pooled item representations. Experiments on four real-world target domains show strong gains in both few-shot and zero-shot settings and significant online CTR improvements, validating HIER’s practical potential for industry-scale, domain-agnostic recommender systems. Overall, the approach demonstrates how explicit decision-path modeling and knowledge-graph–driven pre-training can enhance cross-domain transfer in complex, multi-source environments.

Abstract

Online service platforms offering a wide range of services through miniapps have become crucial for users who visit these platforms with clear intentions to find services they are interested in. Aiming at effective content delivery, cross-domain recommendation are introduced to learn high-quality representations by transferring behaviors from data-rich scenarios. However, these methods overlook the impact of the decision path that users take when conduct behaviors, that is, users ultimately exhibit different behaviors based on various intents. To this end, we propose HIER, a novel Hierarchical decIsion path Enhanced Representation learning for cross-domain recommendation. With the help of graph neural networks for high-order topological information of the knowledge graph between multi-source behaviors, we further adaptively learn decision paths through well-designed exemplar-level and information bottleneck based contrastive learning. Extensive experiments in online and offline environments show the superiority of HIER.

Your decision path does matter in pre-training industrial recommenders with multi-source behaviors

TL;DR

This work tackles cross-domain pre-training for industrial recommender systems by recognizing that user decision paths strongly influence multi-source behaviors. It introduces HIER, a framework that fuses a knowledge-graph–based GNN encoder with hierarchical decision-path modeling, augmented by two contrastive-learning modules: exemplar-level CL for items and information bottleneck CL for users, plus an sparsity term for path selection. The learning objective combines , and , enabling robust, transferable representations from source domains and lightweight fine-tuning on target domains using universal user encodings and mean-pooled item representations. Experiments on four real-world target domains show strong gains in both few-shot and zero-shot settings and significant online CTR improvements, validating HIER’s practical potential for industry-scale, domain-agnostic recommender systems. Overall, the approach demonstrates how explicit decision-path modeling and knowledge-graph–driven pre-training can enhance cross-domain transfer in complex, multi-source environments.

Abstract

Online service platforms offering a wide range of services through miniapps have become crucial for users who visit these platforms with clear intentions to find services they are interested in. Aiming at effective content delivery, cross-domain recommendation are introduced to learn high-quality representations by transferring behaviors from data-rich scenarios. However, these methods overlook the impact of the decision path that users take when conduct behaviors, that is, users ultimately exhibit different behaviors based on various intents. To this end, we propose HIER, a novel Hierarchical decIsion path Enhanced Representation learning for cross-domain recommendation. With the help of graph neural networks for high-order topological information of the knowledge graph between multi-source behaviors, we further adaptively learn decision paths through well-designed exemplar-level and information bottleneck based contrastive learning. Extensive experiments in online and offline environments show the superiority of HIER.
Paper Structure (16 sections, 15 equations, 6 figures, 1 table)

This paper contains 16 sections, 15 equations, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. The proposed approach
  3. Overview
  4. Hierarchical Decision Path Enhanced Representation Learning
  5. Improving Item Representation via Exemplar-level Contrastive Learning
  6. Improving User Representation via Information Bottleneck Contrastive Learning
  7. Improving Decision Path via Hierarchical Structure
  8. Learning of HIER
  9. Experiments
  10. Datasets
  11. Overall Performance
  12. Ablation study
  13. Analysis of Exemplar-level Contrastive Learning
  14. Analysis of Information Bottleneck Contrastive Learning
  15. Online Performance
  16. ...and 1 more sections

Figures (6)

  • Figure 1: A toy example of possible user decision path when interacts with a sharing-bike.
  • Figure 2: Overall architecture of HIER.
  • Figure 3: Ablation study
  • Figure 4: The influence of exemplar number w.r.t. NDCG.
  • Figure 5: The influence of intent number w.r.t. NDCG.
  • ...and 1 more figures