Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Measure Domain's Gap: A Similar Domain Selection Principle for Multi-Domain Recommendation

Yi Wen, Yue Liu, Derong Xu, Huishi Luo, Pengyue Jia, Yiqing Wu, Siwei Wang, Ke Liang, Maolin Wang, Yiqi Wang, Fuzhen Zhuang, Xiangyu Zhao

TL;DR

The paper tackles negative transfer in Multi-Domain Recommendation by introducing the Similar Domain Selection Principle (SDSP), a lightweight framework that explicitly measures domain gaps with a prototype-based distance and dynamically selects similar domains for each target domain. It combines flexible domain representation learning (MMOE-inspired) with a prototype encoder/decoder to learn domain prototypes, an asymmetric distance measure, and an epsilon-greedy, mask-based domain selection module to suppress irrelevant cross-domain transfer. The overall loss blends CTR optimization with a reconstruction term to stabilize prototype learning, achieving improved AUC across three public datasets when integrated with backbones like MMOE and PLE, while maintaining low latency. The work contributes a formal mechanism to quantify inter-domain gaps, a dynamic, model-aware selection policy, and empirical evidence that the approach reduces negative transfer and enhances MDR performance in real-world scenarios.

Abstract

Multi-Domain Recommendation (MDR) achieves the desirable recommendation performance by effectively utilizing the transfer information across different domains. Despite the great success, most existing MDR methods adopt a single structure to transfer complex domain-shared knowledge. However, the beneficial transferring information should vary across different domains. When there is knowledge conflict between domains or a domain is of poor quality, unselectively leveraging information from all domains will lead to a serious Negative Transfer Problem (NTP). Therefore, how to effectively model the complex transfer relationships between domains to avoid NTP is still a direction worth exploring. To address these issues, we propose a simple and dynamic Similar Domain Selection Principle (SDSP) for multi-domain recommendation in this paper. SDSP presents the initial exploration of selecting suitable domain knowledge for each domain to alleviate NTP. Specifically, we propose a novel prototype-based domain distance measure to effectively model the complexity relationship between domains. Thereafter, the proposed SDSP can dynamically find similar domains for each domain based on the supervised signals of the domain metrics and the unsupervised distance measure from the learned domain prototype. We emphasize that SDSP is a lightweight method that can be incorporated with existing MDR methods for better performance while not introducing excessive time overheads. To the best of our knowledge, it is the first solution that can explicitly measure domain-level gaps and dynamically select appropriate domains in the MDR field. Extensive experiments on three datasets demonstrate the effectiveness of our proposed method.

Measure Domain's Gap: A Similar Domain Selection Principle for Multi-Domain Recommendation

TL;DR

The paper tackles negative transfer in Multi-Domain Recommendation by introducing the Similar Domain Selection Principle (SDSP), a lightweight framework that explicitly measures domain gaps with a prototype-based distance and dynamically selects similar domains for each target domain. It combines flexible domain representation learning (MMOE-inspired) with a prototype encoder/decoder to learn domain prototypes, an asymmetric distance measure, and an epsilon-greedy, mask-based domain selection module to suppress irrelevant cross-domain transfer. The overall loss blends CTR optimization with a reconstruction term to stabilize prototype learning, achieving improved AUC across three public datasets when integrated with backbones like MMOE and PLE, while maintaining low latency. The work contributes a formal mechanism to quantify inter-domain gaps, a dynamic, model-aware selection policy, and empirical evidence that the approach reduces negative transfer and enhances MDR performance in real-world scenarios.

Abstract

Multi-Domain Recommendation (MDR) achieves the desirable recommendation performance by effectively utilizing the transfer information across different domains. Despite the great success, most existing MDR methods adopt a single structure to transfer complex domain-shared knowledge. However, the beneficial transferring information should vary across different domains. When there is knowledge conflict between domains or a domain is of poor quality, unselectively leveraging information from all domains will lead to a serious Negative Transfer Problem (NTP). Therefore, how to effectively model the complex transfer relationships between domains to avoid NTP is still a direction worth exploring. To address these issues, we propose a simple and dynamic Similar Domain Selection Principle (SDSP) for multi-domain recommendation in this paper. SDSP presents the initial exploration of selecting suitable domain knowledge for each domain to alleviate NTP. Specifically, we propose a novel prototype-based domain distance measure to effectively model the complexity relationship between domains. Thereafter, the proposed SDSP can dynamically find similar domains for each domain based on the supervised signals of the domain metrics and the unsupervised distance measure from the learned domain prototype. We emphasize that SDSP is a lightweight method that can be incorporated with existing MDR methods for better performance while not introducing excessive time overheads. To the best of our knowledge, it is the first solution that can explicitly measure domain-level gaps and dynamically select appropriate domains in the MDR field. Extensive experiments on three datasets demonstrate the effectiveness of our proposed method.

Paper Structure

This paper contains 36 sections, 2 theorems, 13 equations, 6 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem Statement
  3. Multi-Domain Recommendation
  4. Domain Selection
  5. Method
  6. Motivation
  7. Flexible Domain Representation Learning
  8. Prototype-based Domain Distance Measure
  9. Dynamic Similar Domain Selection
  10. Overall Loss Function
  11. Theoretical Analysis
  12. Experiments
  13. Experiment Setting
  14. Datasets.
  15. Baselines.
  16. ...and 21 more sections

Key Result

Theorem 1

Denote $\mathbf{Z}^i$ is the $i$-th domain's embedding learned from the original multi-domain data $\mathcal{X}$. If there exist a subset $\mathcal{S}^i \subseteq \mathcal{D}$ such that $\mathbf{Z}^i = g^j(\mathbf{X}^j)$, $\forall j \in \mathcal{S}^i$, where $g^j$ can be any function. With the optim where $\Delta_p^i = I(X_i; Y_i)-\min_{j \in \mathcal{S}^i} I(X^j; Y_i)$, $i \in \{1, \ldots, D\}$

Figures (6)

  • Figure 1: The performance of STAR and M2M with selected different domains on Movielens Dataset. "S" and "Y" are abbreviations for the SENIOR and YOUNG, respectively.
  • Figure 2: The framework of the proposed SDSP.
  • Figure 3: Hyperparameter Analysis of MMOE+SDSP and PLE+SDSP on Movielens dataset.
  • Figure 4: The ablation experiments of MMOE and PLE on the Movielens dataset. "Random" represents completely randomly selected, and "Average" denotes using the average representation to measure domain distance.
  • Figure 5: The comparison of the negative transfer of PLE and PLE+SDSP on Movielens dataset. "S" and "Y" are abbreviations for the SENIOR and YOUNG, respectively.
  • ...and 1 more figures

Theorems & Definitions (8)

  • definition 1
  • definition 2
  • Remark 1
  • Remark 2
  • Remark 3
  • Theorem 1
  • Remark 4
  • Lemma 2