The Ulsan National Institute of Science and Technology (UNIST) announced on the 6th that a research team led by Professor Lee Yeon-chang of its Graduate School of Artificial Intelligence has developed a cross-domain technology called "Multi-TAP" that analyzes consumers' purchasing tendencies by detailed product category to recommend tailored products.
Cross-domain recommendation is a technology that uses information from data-rich fields to recommend products in other fields where data is scarce. Existing technologies had limitations in that they grouped entire product categories together for analysis, failing to reflect differences in consumers' detailed preferences, such as favoring high-priced products for computers while preferring low-priced popular products for home audio.
The research team solved this problem using a large language model (LLM). After dividing consumer tendencies into five criteria—price sensitivity, quality and popularity preference, usage frequency, and product category diversity—the team used GPT-4o to generate a "multi-persona" in sentence form reflecting detailed preferences.
This analytical information is converted into numbers and undergoes a process of calculating the relevance between the recommendation subject (target) and existing information (source). By applying a "target-adaptive doppelganger transfer" method that weights information with high relevance, the team prevented the phenomenon of reduced recommendation accuracy caused by unnecessary information interference.
In verification using purchase records from five of Amazon's product categories, Multi-TAP recorded the best performance in five of six cross-recommendation tasks. The rate of including actually purchased products in the top five recommendation list (HR@5) rose by up to 36.3% compared to existing technologies, while recommendation ranking accuracy (NDCG@5) increased by 42.6%.
"Even in fields with few purchase records, we can recommend products close to actual preferences, which will help advance the personalized services of online shopping malls," Professor Lee explained.
The research, in which researcher Kang Dae-hee participated as first author, was adopted by "ACM KDD," an international academic conference in the field of data mining. The conference will be held for five days starting on the 9th of this month in Jeju Island.
