SKT가 지난 7월 미국 워싱턴 DC에서 개최된 정보 검색 분야 세계적 권위 학회 ‘SIGIR 2024’에서 자체 개발 추천 모델 알고리즘 연구가 우수 논문상(Best Paper Honorable Mention)을 수상했다.
Awarded at SIGIR 2024, the world's leading information retrieval conference
Self-developed 'One Model Version 2.0' algorithm
SKT won the Best Paper Honorable Mention award for its self-developed recommendation model algorithm research at 'SIGIR 2024', a world-renowned conference in the field of information retrieval held in Washington D.C., USA in July.
The winning paper at SIGIR (International ACM SIGIR Conference on Research and Development in Information Retrieval) 2024 is a study on SKT's 'One Model Version 2.0'.
SKT said, “We proposed an algorithm that improves recommendation prediction performance by synergizing data from various service domains,” and added, “This paper was selected for the Excellent Paper Award, which is given only to the top 0.6% of received papers, as it was highly evaluated for the novelty of the algorithm, the empiricality of commercial deployment, and the reliability of the results through extensive experiments.”
SKT commercially distributed version 1.0 of its self-developed recommendation model, 'One Model', last year. Research on the algorithm of this model has been accepted at 'CIKM 2023 (ACM International Conference on Information and Knowledge Management),' one of the best academic conferences in the field of information retrieval.
This One Model version 2.0 is said to have improved recommendation performance and learning efficiency compared to version 1.0.
Meanwhile, SKT is integrating or refining various types of individual behavior logs in chronological order, predicting the customer's next behavior through the 'One Model algorithm', which is the result of this research, and performing personalized recommendations that take into account the customer's multidimensional characteristics.
For example, it was introduced in the Cross-Domain Sequential Recommendation method. It is a method that comprehensively analyzes the customer's behavioral data in various service domains, such as rate plan subscription history, T-deal shopping history, and membership usage history, and recommends service benefits or products that match the customer's needs and interests at the most recent point in time.
One Model actually learns more than 10 different data domains simultaneously and provides recommendations from various channels within SKT as a single model. According to SKT, it has the effect of improving customer response rates by up to 3 times compared to existing recommendation methods.
In particular, to effectively utilize data from multiple service domains, a single-domain learning model (Pacer) and a multi-domain learning model (Runner) are configured in one architecture. It has attracted attention as a way to suggest mutual synergy effects through a complementary learning method.
SKT said, “Currently, the model is being applied to the recommendation system of SKT’s AI personal assistant service, Adot, T membership, and rate plan recommendations, and it is planned to be expanded to various product recommendations, such as the subscription product T Universe and AI curation commerce T Deal, within the year.”
Jeong Do-hee, AI data manager at SKT’s AI service business division, said, “SK Telecom has proven its AI capabilities once again by winning an excellent paper award from a world-renowned academic society this year, following last year’s award,” and “Going forward, we will apply advanced personalized technologies to all of our services to further increase customer satisfaction and accelerate our evolution into a global AI company.”