Source-linked AI summary
Transfer-Meta Framework for Cross-domain Recommendation to Cold-Start Users
Yongchun Zhu, Kaikai Ge, Fuzhen Zhuang, Ruobing Xie, Dongbo Xi, Xu Zhang, Leyu Lin, Qing He
TL;DR
Cold-start recommendation is difficult because EMCDR mapping functions rely on limited overlapping users and may generalize poorly. TMCDR addresses this with transfer pre-training and a task-oriented meta network, and experiments across six cross-domain tasks on Amazon and Douban support its effectiveness and compatibility.
Problem
EMCDR learns mappings from limited overlapping users, which can bias generalization for cold-start users in the target domain.
Method
TMCDR uses transfer-stage pre-trained source and target models and a meta-stage task-oriented network to transform source user embeddings into the target feature space.
Results
Across six cross-domain tasks, TMCDR outperforms most compared methods and improves most EMCDR-based methods across varied base models.
Takeaways & Limitations
TMCDR provides a general framework compatible with EMCDR-based methods and multiple base models, including MF, BPR, ListRank-MF, and CML.
Abstract
from arXiv · showhide
Cold-start problems are enormous challenges in practical recommender systems. One promising solution for this problem is cross-domain recommendation (CDR) which leverages rich information from an auxiliary (source) domain to improve the performance of recommender system in the target domain. In these CDR approaches, the family of Embedding and Mapping methods for CDR (EMCDR) is very effective, which explicitly learn a mapping function from source embeddings to target embeddings with overlapping users. However, these approaches suffer from one serious problem: the mapping function is only learned on limited overlapping users, and the function would be biased to the limited overlapping users, which leads to unsatisfying generalization ability and degrades the performance on cold-start users in the target domain. With the advantage of meta learning which has good generalization ability to novel tasks, we propose a transfer-meta framework for CDR (TMCDR) which has a transfer stage and a meta stage. In the transfer (pre-training) stage, a source model and a target model are trained on source and target domains, respectively. In the meta stage, a task-oriented meta network is learned to implicitly transform the user embedding in the source domain to the target feature space. In addition, the TMCDR is a general framework that can be applied upon various base models, e.g., MF, BPR, CML. By utilizing data from Amazon and Douban, we conduct extensive experiments on 6 cross-domain tasks to demonstrate the superior performance and compatibility of TMCDR.
1 INTRODUCTION
Cold-start recommendation is addressed through CDR, but EMCDR mapping functions trained on few overlapping users may generalize poorly. TMCDR combines transfer pre-training with a task-oriented meta network and supports multiple recommendation base models.
- 1 INTRODUCTION: Cross-domain recommendation leverages information from an auxiliary domain to improve recommendation in a target domain and address cold-start problems.The motivation includes both new-user and new-item cold-start settings.
- 1 INTRODUCTION: Limited overlapping users bias EMCDR mapping functions and degrade performance for cold-start users in the target domain.The mapping function is trained using one sample per overlapping user, with experiments containing 894 to 37,388 overlapping users.
- 1 INTRODUCTION: TMCDR combines a transfer stage with a meta stage to improve cross-domain cold-start recommendation.The transfer stage uses pre-trained source and target models, while the meta stage learns a task-oriented transformation into the target feature space.
- 1 INTRODUCTION: TMCDR is designed as a general framework applicable to EMCDR-based methods with MF, BPR, and CML base models.The framework is presented as a replacement for the training procedure of most EMCDR-based methods.
- 1 INTRODUCTION: The meta stage trains a task-oriented network through simulated learning and cold-start phases using MAML-style optimization.Training tasks separate users who provide learning signals from users who simulate cold-start cases, and optimize rating or ranking objectives rather than embedding distance.
2 MODEL
TMCDR combines pretrained source and target models with a task-oriented meta network for cross-domain cold-start recommendation. Its meta-learning procedure simulates cold-start adaptation using disjoint user groups and supports multiple recommendation base models.
- 2.1 Transfer Stage: TMCDR uses source and target models pretrained on all domain data before learning the task-oriented meta network.Using all data allows the pretrained models to capture information from items beyond those interacted with by overlapping users.
- 2 MODEL: The framework supports MF, BPR, ListRank-MF, and CML-style base models by replacing EMCDR mapping functions with the task-oriented meta network.The supplied evaluation caption identifies recommendation performance across six CDR tasks, with significance testing for TMCDR_MF versus the best baselines.
- 2.2 Meta Stage: The meta network implicitly transforms source embeddings of overlapping users into the target feature space.The source and target pretrained embeddings remain fixed while the task-oriented network is learned.
- 2.2 Meta Stage: Training tasks split disjoint groups of overlapping users into learning and cold-start phases to simulate adaptation and evaluation.The group strategy addresses instability caused by tasks containing only two users with potentially sparse interactions.
- 2.2 Meta Stage: The learning phase minimizes task-oriented loss on D_a, adapts parameters with gradient descent, and evaluates the adapted network on D_b.The meta-optimization updates the original parameters using loss computed with the adapted parameters.
3 EXPERIMENTS
Experiments evaluate TMCDR on six Amazon and Douban cross-domain tasks using cold-start users, ranking metrics, and single- and cross-domain baselines. TMCDR generally outperforms compared methods and improves multiple EMCDR-based models across tasks.
- Experimental Settings: The evaluation uses Amazon and Douban data across four Amazon scenarios and two Douban scenarios, with about 20% of overlapping users held out as cold-start users.AUC and NDCG@K measure ranking performance, while the baseline set includes CMF, BPR, ListRank-MF, CML, CST, EMCDR, and SSCDR.
- Recommendation Performance: TMCDR outperforms most compared methods across six CDR scenarios evaluated with AUC and NDCG@10.The comparison includes single-domain and cross-domain baselines, with results reported in Table 1.
- Recommendation Performance: EMCDR_MF largely outperforms EMCDR_ori_MF on all scenarios because training with all samples captures items absent from overlapping-user interactions.The experiments therefore favor using all available samples rather than only overlapping-user samples.
- Recommendation Performance: CML outperforms all methods on NDCG@10 in Scenario 3, while different base models perform differently across scenarios.This result motivates comparing multiple recommendation models rather than assuming one base model is uniformly best.
- Recommendation Performance: CST’s use of a source model as a pretrained model with parameter regularization also produces remarkable results.The authors use this observation to support the effectiveness of fine-tuning recommendation models from pretrained models.
- Generalization: TMCDR improves most EMCDR-based methods across tasks, supporting its generalization across MF, BPR, ListRank-MF, and CML variants.Figure 1 compares EMCDR-based baselines with corresponding TMCDR modifications; red lines denote baselines and blue lines denote modified methods.
4 CONCLUSION
The paper addresses cold-start cross-domain recommendation by replacing limited-overlap mapping with a transfer-meta framework. Experiments on Amazon and Douban validate TMCDR’s effectiveness and compatibility with varied base models.
- 4 CONCLUSION: TMCDR uses a task-oriented meta network to address overfitting from limited overlapping users in EMCDR.The framework is motivated by meta learning and is designed for cold-start users.
- 4 CONCLUSION: TMCDR is a general framework applicable to EMCDR-based methods with base models including MF, BPR, and CML.The conclusion also identifies compatibility with varied recommendation models as a central contribution.
- 4 CONCLUSION: Experiments on real-world Amazon and Douban datasets validate TMCDR’s effectiveness and compatibility.The evaluation covers six cross-domain tasks.