[Paper Review] User-specific Adaptive Fine-tuning for Cross-domain Recommendations

Lei Chen, Fajie Yuan+4 more|arXiv (Cornell University)|Jun 15, 2021

Recommender Systems and Techniques40 references4 citations

TL;DR

This paper proposes User-specific Adaptive Fine-tuning (UAF), a novel method for cross-domain recommendation that dynamically selects which layers of a pre-trained model to fine-tune on a per-user basis using a policy network. UAF significantly improves cold-start user recommendation accuracy and robustness, especially under data-scarce conditions, outperforming global fine-tuning and existing transfer learning baselines across multiple datasets and tasks.

ABSTRACT

Making accurate recommendations for cold-start users has been a longstanding and critical challenge for recommender systems (RS). Cross-domain recommendations (CDR) offer a solution to tackle such a cold-start problem when there is no sufficient data for the users who have rarely used the system. An effective approach in CDR is to leverage the knowledge (e.g., user representations) learned from a related but different domain and transfer it to the target domain. Fine-tuning works as an effective transfer learning technique for this objective, which adapts the parameters of a pre-trained model from the source domain to the target domain. However, current methods are mainly based on the global fine-tuning strategy: the decision of which layers of the pre-trained model to freeze or fine-tune is taken for all users in the target domain. In this paper, we argue that users in RS are personalized and should have their own fine-tuning policies for better preference transfer learning. As such, we propose a novel User-specific Adaptive Fine-tuning method (UAF), selecting which layers of the pre-trained network to fine-tune, on a per-user basis. Specifically, we devise a policy network with three alternative strategies to automatically decide which layers to be fine-tuned and which layers to have their parameters frozen for each user. Extensive experiments show that the proposed UAF exhibits significantly better and more robust performance for user cold-start recommendation.

Motivation & Objective

• Address the cold-start problem in recommender systems where users have insufficient interaction data in the target domain.
• Overcome the limitations of global fine-tuning strategies that treat all users uniformly, despite their varying similarity to the source domain.
• Develop a personalized fine-tuning mechanism that adapts transfer learning policies per user to maximize preference transfer effectiveness.
• Improve robustness against overfitting in low-data regimes by selectively updating only relevant layers per user.
• Demonstrate the versatility of the proposed method beyond recommendation, including in user profile prediction tasks.

Proposed method

• Train a deep sequential recommendation model (NextItNet) on a large source domain dataset as a pre-trained backbone model.
• Introduce a policy network that predicts, for each individual user, which layers of the pre-trained model should be fine-tuned and which should remain frozen.
• Implement three variants of the policy network: hard-gating (UAF-Hard), soft-gating (UAF-Soft), and reinforcement learning-based (UAF-RL), to determine layer-wise fine-tuning decisions.
• Use the policy network’s output to apply dynamic layer-wise masking during fine-tuning, enabling user-specific adaptation of the pre-trained model.
• Train the policy network end-to-end with the recommendation model using backpropagation, allowing it to learn optimal fine-tuning policies per user.
• Replace the original NextItNet backbone in the policy network with a GRU to evaluate flexibility and performance, showing comparable results.

Experimental results

Research questions

• RQ1 Can user-specific fine-tuning policies improve cross-domain recommendation performance compared to global fine-tuning strategies?
• RQ2 How does UAF perform under low-data regimes typical of cold-start users?
• RQ3 Does UAF prevent overfitting better than standard fine-tuning when training data is scarce?
• RQ4 Can the proposed method generalize beyond recommendation to other downstream tasks like user profile prediction?
• RQ5 How robust and flexible is the policy network design across different architectures and tasks?

Key findings

• UAF-Hard, UAF-Soft, and UAF-RL achieved significantly better MRR@5 and HR@5 performance than Finetune-All and PeterRec on ColdRec-1 and ColdRec-2, with up to a 6.1% improvement in MRR@5 over PeterRec on ColdRec-1.
• UAF methods prevented overfitting more effectively than standard fine-tuning, with learning curves showing sustained improvement until 9 epochs, while Finetune-All began to degrade after 5 epochs.
• The policy network with GRU architecture (UAF-GRU variants) achieved comparable performance to the original NextItNet-based policy network, demonstrating the method’s architectural flexibility.
• UAF outperformed baselines in user profile prediction tasks (gender, age, life status), achieving higher accuracy and F1 scores across all three datasets, confirming its versatility.
• The ablation study confirmed that user-specific fine-tuning policies lead to better performance than fixed global strategies, validating the core contribution of personalization in transfer learning.