Source-linked AI summary
Personalized Transfer of User Preferences for Cross-domain Recommendation
Yongchun Zhu, Zhenwei Tang, Yudan Liu, Fuzhen Zhuang, Ruobing Xie, Xu Zhang, Leyu Lin, Qing He
TL;DR
Cold-start recommendation remains difficult, and CDR must transfer user preferences from an auxiliary source domain to a target domain despite user-specific relationships. PTUPCDR uses a meta network and task-oriented optimization to generate personalized bridges; experiments validate its effectiveness across cold- and warm-start settings.
Problem
CDR must transfer source-domain user preferences to the target domain, but shared bridges struggle to capture relationships that vary across users.
Method
PTUPCDR feeds users’ characteristic embeddings into a meta network that generates personalized bridge functions, trained with task-oriented optimization.
Results
Experiments on real-world datasets validate PTUPCDR’s effectiveness and robustness across cold-start and warm-start scenarios.
Takeaways & Limitations
Personalized bridges provide a framework for transferring source-domain preferences while accommodating user-specific relationships between domains.
Takeaways & Limitations
The task-oriented optimization setting is constrained by users who have only limited interactions and by the CDR setting’s overlapping-user formulation.
Abstract
from arXiv · showhide
Cold-start problem is still a very challenging problem in recommender systems. Fortunately, the interactions of the cold-start users in the auxiliary source domain can help cold-start recommendations in the target domain. How to transfer user's preferences from the source domain to the target domain, is the key issue in Cross-domain Recommendation (CDR) which is a promising solution to deal with the cold-start problem. Most existing methods model a common preference bridge to transfer preferences for all users. Intuitively, since preferences vary from user to user, the preference bridges of different users should be different. Along this line, we propose a novel framework named Personalized Transfer of User Preferences for Cross-domain Recommendation (PTUPCDR). Specifically, a meta network fed with users' characteristic embeddings is learned to generate personalized bridge functions to achieve personalized transfer of preferences for each user. To learn the meta network stably, we employ a task-oriented optimization procedure. With the meta-generated personalized bridge function, the user's preference embedding in the source domain can be transformed into the target domain, and the transformed user preference embedding can be utilized as the initial embedding for the cold-start user in the target domain. Using large real-world datasets, we conduct extensive experiments to evaluate the effectiveness of PTUPCDR on both cold-start and warm-start stages. The code has been available at https://github.com/easezyc/WSDM2022-PTUPCDR.
1 INTRODUCTION
Cross-domain recommendation transfers source-domain preferences to alleviate target-domain cold start, but shared preference bridges cannot capture user-specific relationships. PTUPCDR addresses this with meta-generated personalized bridges and task-oriented optimization, with experiments covering cold- and warm-start settings.
- CDR transfers knowledge from an informative source domain to the target domain to alleviate the cold-start problem.
- Existing CDR methods commonly learn one preference bridge shared by all users, despite relationships between domains varying across users.
- PTUPCDR uses users’ source-domain characteristic embeddings in a meta network to generate a personalized bridge function for each user.
- The generated bridge transforms source-domain user embeddings into target-domain initial embeddings for recommendation.
- Task-oriented optimization trains the meta network using the rating task directly instead of relying on target-domain user embeddings.
- Experiments on three cross-domain tasks evaluate PTUPCDR in both cold-start and warm-start scenarios, unlike existing methods focused only on cold start.
2 RELATED WORK
Related work addresses cross-domain recommendation through transfer learning, deep models, preference bridges, auxiliary information, and meta learning. PTUPCDR belongs to bridge-based and parameter-generating approaches while introducing personalized rather than shared bridges.
- Transfer learning leverages source-domain knowledge to improve target-domain learning or reduce the labeled examples required there.
- Deep CDR models transfer knowledge through neural cross-connections, cluster-level rating patterns, or latent orthogonal mappings across domains.
- Bridge-based CDR methods connect user preferences across domains, including source-embedding initialization and explicitly modeled preference bridges.
- Existing bridge-based works learn a shared bridge for all users, whereas PTUPCDR learns personalized bridges for each user.
- Cold-start methods may use contextual bandits or auxiliary information such as user attributes, item attributes, knowledge graphs, and auxiliary-domain samples.
- PTUPCDR is a parameter-generating meta-learning method that uses a meta learner to predict network parameters.
3 MODEL
PTUPCDR transfers source-domain preferences to the target domain through user-specific bridge functions generated from transferable characteristics. It uses task-oriented optimization and a staged procedure to initialize target-domain embeddings for cold-start users and support warm-start users.
- 3.1 Problem Setting: CDR transfers knowledge from a source domain to a target domain using overlapping users, with source and target item sets disjoint.The framework represents users and items with domain-specific embeddings and uses overlapping users to support transfer.
- 3.2 Characteristic Encoder: An attention network forms each user’s transferable characteristic embedding by weighting interacted source-item embeddings according to their importance for predicting the personalized bridge.The attention scores are learned from item embeddings, and the weighted sum captures characteristics useful for transfer.
- 3.5 Overall Procedure: The personalized bridge transforms a source-domain user embedding into a target-domain embedding that initializes the user representation for target-domain recommendation.The transformed embedding is used directly for extreme cold-start prediction and can be fine-tuned with target interactions for warm-start users.
- 3.3 Meta Network: A meta network maps each user’s characteristic embedding to bridge parameters, producing a personalized bridge function whose parameters vary across users.The bridge can use different structures; PTUPCDR uses a linear layer and reshapes the generated parameter vector into a matrix.
- 3.4 Task-oriented Optimization: Task-oriented optimization trains the characteristic encoder and meta network against the ultimate rating task rather than directly matching transformed and target user embeddings.This procedure uses rating data as ground truth, alleviates effects from unreasonable intermediate embeddings, and provides more training samples to reduce overfitting.
- 3.5 Overall Procedure: PTUPCDR uses pre-training, meta, initialization, and test stages, and is designed for both cold-start and warm-start recommendation.Pre-training learns latent spaces separately for the source and target domains before the meta network generates personalized transformations.
4 EXPERIMENTS
Experiments on three Amazon cross-domain tasks evaluate PTUPCDR in cold-start, warm-start, and more complex model settings. Results support its effectiveness across these scenarios and attribute improvements to personalized transfer and attention-based characteristic encoding.
- Experimental Settings: Experiments use Amazon-5cores data across three tasks: Movie →Music, Book →Movie, and Book →Music, evaluated with MAE and RMSE.Each user or item has at least five ratings, and the source domain contains substantially more ratings than the target domain.
- Experimental Settings: The evaluation compares PTUPCDR with target-only, shared-embedding, and bridge-based CDR baselines under varying cold-start proportions.Test-user proportions are β=80%, 50%, and 20%; cold-start results are reported as means over five runs.
- Cold-start Experiments: Cross-domain methods outperform the target-only model in cold-start experiments, indicating that auxiliary-domain data helps alleviate target-domain sparsity.The cold-start section reports this finding across the evaluated cross-domain tasks and baselines.
- More Practical Scenarios: PTUPCDR remains effective with MF, GMF, and YouTube DNN, achieving the best results across the tested base models.The generalization experiments apply EMCDR and PTUPCDR to one non-neural and two neural base models, with β=0.2.
- Warm-start Experiments: In warm-start settings, PTUPCDR achieves the best results across test-user proportions, while bridge-based methods outperform target-only initialization.Warm-start models use pre-trained embeddings for initialization, and all models perform better than in the cold-start stage.
- Explanation of the Improvement: t-SNE visualizations show PTUPCDR-transformed embeddings better fit the target-embedding distribution than EMCDR embeddings.EMCDR transformed embeddings are described as concentrated, whereas PTUPCDR embeddings are scattered across and better fit the target feature space.
5 CONCLUSION
PTUPCDR addresses cross-domain recommendation by personalizing preference transfer across users. Experiments on real-world datasets validate its effectiveness in both cold-start and warm-start stages.
- PTUPCDR uses a meta network to generate personalized bridge functions from users’ characteristic embeddings.
- The framework targets the limitation that a single shared bridge cannot capture varied source–target preference relationships.
- Experiments on real-world datasets validate PTUPCDR’s effectiveness in both cold-start and warm-start stages.