"Looks Just Like the Product I Chose?!" AI Developed for Multi-Persona Personalized Recommendations [IT Item of the Day]

[Financial News] An artificial intelligence (AI) technology has been developed that makes recommendations by reflecting consumers' preferences, which vary by product category. The rate of actual purchases was up to 36% higher than with existing technologies. According to Ulsan National Institute of Science and Technology (UNIST) on the 7th, a research team led by Professor Lee Yeon-chang at the UNIST Graduate School of Artificial Intelligence developed 'Multi-TAP,' a cross-domain recommendation technology that can accurately capture consumers' purchasing tendencies, which differ by detailed product category.
Multi-TAP is a cross-domain recommendation technology that distinguishes purchasing tendencies that vary even within the same product category, while reflecting information from other product categories only to the extent that it is relevant to the current recommendation. As a result, it can more accurately capture consumers' detailed preferences than existing technologies. Under previous methods, a user might choose an expensive, highly rated product when buying a computer, but opt for a cheaper, popular one when buying home audio equipment, and such differences were not properly reflected.
Multi-TAP analyzes purchase histories and product information to determine, in three levels, how much users consider price, ratings and the number of reviews when buying a product, as well as how often they use products and how many different subcategories they have purchased within the same product category. A large language model (LLM) then turns this into sentences that highlight differences by product category. These sentences are converted into numerical data and combined with purchase behavior information learned by LightGCN, an existing recommendation algorithm, before products are recommended.
Using an LLM allows the recommendation model to learn not just a list of three levels — high, medium and low — but also which product category shows which tendency. The system also calculates how relevant information from other categories is to the product category being recommended, and adjusts how much of that information is reflected.
In experiments using purchase records from Amazon's electronics, home and living goods, sports equipment, apparel and toy categories, the system delivered the best performance in five of six major cross-recommendation tasks. The share of actual purchased items included in the top five recommendations was up to 36.3% higher than the previous best technology.
Professor Lee Yeon-chang said, "Because consumers' preferences change depending on the situation and the specific product category, Multi-TAP divides purchasing tendencies into multiple 'multi-personas' and reflects those differences in recommendations." He added, "It could improve personalized recommendations for online shopping malls by suggesting products that are closer to users' actual preferences, even in product categories with limited purchase history."
The study was accepted to ACM KDD, the top international conference in the field of data mining. This year's conference will be held in Jeju Island for five days starting on the 9th.

[email protected] Yeon Ji-an Reporter