Knowledge Enhanced Multi-Domain Recommendations in an AI Assistant Application

TL;DR

This work tackles cold-start and cross-domain recommendation in a conversational AI assistant by unifying knowledge graph enhancement with multi-domain learning. It introduces a three-component model (user encoder, item encoder, scoring function) with variants that incorporate KG embeddings via a TransE-based objective and a fusion block, and leverages domain-specific plus shared user representations through GNNs. Empirical results on a large, real-world dataset across music, video, and books demonstrate additive benefits from KG enhancement and multi-domain modeling, with the strongest performance in zero-shot scenarios when combining all components. The findings support the practical deployment of KG-enhanced multi-domain recommendations in personalized AI assistants, improving both overall and new-user experiences.

Abstract

This work explores unifying knowledge enhanced recommendation with multi-domain recommendation systems in a conversational AI assistant application. Multi-domain recommendation leverages users' interactions in previous domains to improve recommendations in a new one. Knowledge graph enhancement seeks to use external knowledge graphs to improve recommendations within a single domain. Both research threads incorporate related information to improve the recommendation task. We propose to unify these approaches: using information from interactions in other domains as well as external knowledge graphs to make predictions in a new domain that would not be possible with either information source alone. We develop a new model and demonstrate the additive benefit of these approaches on a dataset derived from millions of users' queries for content across three domains (videos, music, and books) in a live virtual assistant application. We demonstrate significant improvement on overall recommendations as well as on recommendations for new users of a domain.

Figures (1)

• Figure 1: The combined MultiDomain KGE+GNN model provides recommendations for user $u$ based on their interaction graph and recommends items $v$ which are enhanced with KG embeddings using KG edges of the form $(h,r,t)$. KG entity and relation embeddings are trained using KG embedding task, while the item embeddings and interaction module are trained over the recommendation task. The user model contains one shared Graph Neural Network and multiple unique ones for each domain. Note that the domain specific GNNs still uses all prior interactions from multiple domains, but the weights and output is specific one domain.