A customer avatar model based on Kolmogorov-Arnold networks

The increasing pace of development of e-commerce continues to present new challenges in terms of personalizing product search and recommendations. Monolithic search and recommendation systems have become cumbersome and are unable to effectively address the need for a deeper understanding of users on electronic trading platforms (ETPs) despite having access to comprehensive information about their interests and purchase histories. Collaborative filtering mechanisms which are widely used suffer from a lack of diversity in offerings and a reduced capacity to surprise users. Additionally, the low frequency of recommendation updates and the replacement of "personalized" with "similar to others" concepts contribute to these issues. We have approached the resolution of these issues by developing a shopping assistant named "Ellochka" that is individual for each user of ETP. The digital avatar model of the user continually searches for relevant products based on their history of interaction with ETP. We were guided by the principle of independence-avatar models do not share information with each other. When a new user joins, they are assigned a unique avatar model that evolves independently. Each avatar has its own language to generate search queries. The level of complexity of each avatar can vary depending on the intensity of its interaction with ETP. Continued interaction with the avatar allows for tracking of optimal purchase conditions, reminding users of expiration dates and the need for repurchasing frequently purchased items. Isolating the avatar allows it to be retrained after each event, without significantly impacting the overall search and recommendation system. The use of neural network architecture-based and Kolmogorov-Arnold networks in the avatar-model has led to improvements in the main indicators of search and recommendation effectiveness, namely, novelty and diversity.