TLDR: GMC is a novel generative AI framework for Multi-Target Cross-Domain Recommendation (MTCDR). It tackles the challenge of simultaneously improving recommendations across multiple domains, especially when shared users or items are absent. GMC achieves this by learning domain-shared discrete semantic identifiers for items and training a unified generative model that integrates knowledge across all domains. It further refines performance with domain-specific fine-tuning, demonstrating superior results on real-world datasets without relying on extensive auxiliary pre-training data.
In today’s digital world, recommender systems are essential for helping us navigate the vast amount of information and products available online. From suggesting movies to recommending products on e-commerce sites, these systems aim to deliver personalized content to users. However, a significant challenge arises when trying to improve recommendations across multiple, distinct platforms or ‘domains’ simultaneously. This is known as Multi-Target Cross-Domain Recommendation (MTCDR).
Traditional MTCDR methods often hit a wall because they rely on shared elements, like the same users or items existing across different domains. In many real-world scenarios, such overlaps are rare or non-existent. Other approaches try to model user preferences and item features as universal semantic representations, but these typically demand massive amounts of auxiliary data for pre-training, which isn’t always available. This leaves a clear need for more effective solutions.
Inspired by the recent breakthroughs in generative artificial intelligence, a new research paper introduces a novel approach called GMC, which stands for Generative paradigm-based approach for Multi-target Cross-domain recommendation. The core idea behind GMC is to use a shared set of ‘semantic identifiers’ for items across different domains. Think of these identifiers as a universal language that allows the system to understand and integrate knowledge from various platforms within a single, unified generative model.
Here’s how GMC works in three main steps:
Semantic Identifier Learning
First, GMC employs an ‘item tokenizer’ to create these domain-shared semantic identifiers for each item. It uses large language models (LLMs) to understand the textual descriptions of items (like titles, brands, and categories) and then converts these into discrete, hierarchical codes using a technique called Residual Quantization Variational Autoencoder (RQ-VAE). To make these identifiers even more effective, GMC introduces a ‘domain-aware contrastive loss’. This clever addition ensures that items from the same domain have more similar identifiers, while still allowing for cross-domain sharing at a deeper, more abstract level. This means the system can capture both domain-specific characteristics and universal semantic information.
Unified Recommender Training
Once items have their semantic identifiers, the recommendation task is reframed as a ‘next-token generation’ problem. All the training data from every domain is combined to train a single, unified sequence-to-sequence model. This model learns to predict the next item’s semantic identifier based on a user’s historical interactions, effectively fusing and transferring recommendation knowledge across all domains simultaneously.
Also Read:
- Unlocking New Discoveries: A Multi-View Approach to Identifying Unseen Data Categories
- PgM: A New Framework for Enhanced Multimodal Learning
Domain-Specific Fine-tuning
Even with a unified model, each domain has its unique characteristics. To account for this without undermining the shared knowledge, GMC uses a lightweight tuning strategy. It incorporates ‘LoRA modules’ (Low-Rank Adaptation) into the unified recommender. These modules are small, trainable components that are updated using domain-specific data, while the main part of the unified model remains frozen. This allows the system to adapt to individual domain nuances efficiently, without requiring a separate, full-sized model for each domain.
Extensive experiments conducted on five public datasets demonstrate that GMC significantly outperforms a range of existing recommendation methods. Its ability to effectively integrate and transfer multi-domain knowledge through these shared semantic codes, combined with the power of the generative paradigm, gives it a strong advantage in recommending items across diverse platforms.
This innovative approach addresses critical limitations of previous methods, such as the reliance on overlapping users or the need for vast auxiliary pre-training data. By mapping item information from multiple domains into a shared discrete semantic space, GMC facilitates structured knowledge transfer and mitigates the fragmentation often seen in domain-isolated recommenders. For more technical details, you can refer to the full research paper: Generative Multi-Target Cross-Domain Recommendation.
