[Edaily Reporter KYUNG GYEYOUNG ] If you search for sofas a few times on an online shopping site, sofas will inevitably continue to appear in your recommended products even after you’ve purchased one. This was a limitation of existing artificial intelligence (AI), which only recommended “what customers liked.” Now, hyper-personalized AI technology has been developed that takes into account the context—such as whether a consumer has already purchased a sofa or might ignore further recommendations—when making suggestions. #The result of over a year of collaboration between Shinsegae Co.,Ltd and a research team led by Professor Oh Min-hwan of Seoul National University’s Graduate School of Data Science has been accepted by the International Conference on Machine Learning (ICML), the world’s most prestigious AI conference.

In a recent interview with E-Daily, Professor Oh Min-hwan explained why ICML took notice of this research, stating, “It’s because we went beyond simply using past history to make recommendations; we mathematically modeled post-recommendation reactions—such as consumers feeling fatigued by recent recommendations or their interest growing as they see them more frequently.”

To address the impact of recommendations on consumers’ current preferences and purchasing behavior, the research team combined the “memory effect” with an existing recommendation algorithm (Contextual Bandit) designed to find the optimal suggestion for each context. When behavioral change data is added, the algorithm’s computations become exponentially more complex. Professor Oh explained, “This study not only efficiently solved these complex computations but also demonstrated mathematical optimal convergence for long-term cumulative performance under a defined model.”

In particular, this study incorporated 200 million purchase records held by Shinsegae Co.,Ltd., opening the door to practical application in the retail sector. According to preliminary simulations conducted by Shinsegae Co.,Ltd. using AI-based personalization technology, average transaction value increased by up to 46%.

Professor Oh noted, “By representing customer purchase histories and the department store’s brand portfolio as features (expressed as vectors), we can calculate which brand promotions are most advantageous for specific customers.” He added, “Compared to offering uniform promotions to all customers, implementing personalized promotions is likely to increase both response rates and average transaction value.” He added, “If the recommendation AI functions effectively, not only will retailers maximize profits, but customers will also be able to shop more efficiently and experience higher satisfaction, creating a win-win situation for both parties.”

The quality of Shinsegae Co.,Ltd.’s data also played a key role in deriving these research findings. Professor Oh emphasized, “There were no stereotypes based on specific age groups or genders, and each customer had distinct preferences, so there was ample room for the AI to make better recommendations through learning,” adding, “Even for new customers, their preferences can be identified after a few interactions based on what the system has already learned.”

Shinsegae Co.,Ltd plans to incorporate the findings of this research into its “AI Sales Agent,” scheduled for launch early next year. The AI Sales Agent is a business analytics tool designed to assist buyers and store managers in product management and the development of promotional strategies based on data from the sales floor.

Professor Oh emphasized, “This collaboration was carried out through organic communication between the company and the research team, allowing us to incorporate knowledge and know-how from the retail sector into the AI model,” adding, “Through the integration of online and offline data and digital transformation (DX), we will be able to achieve a seamless, organic connection where offline experiences lead to online ones and online experiences lead to offline ones.”