Responsive Calibrated Web Personalization System with Online Local Variational Inference for the Logistic Regression Mixture Model

Improved computing environments performing large-scale data processing and high-speed computational processing facilitate the delivery of new algorithms to businesses while considering cost efficiency for small-scale investments. Implementing the proposed method more as a criterion for feasibility and economic rationality in specific problem areas rather than as an approach to generic issues, we aim to develop technologies of practical use in the real world. Recently, it has become possible for customers to monitor their buying behavior through smart devices, and with the improvement of computing performance, it has become possible to improve the accuracy of prediction and recommendation cycles through active online learning. This study proposes a method for dynamically recommending products that are highly likely to be selected by the user by combining the user's reaction with reuse of knowledge and real-time online learning to cyclically repeat feedback that is more specific to the user.We propose a method to sense streaming data by utilizing a user's behavior, intervening a user's behavioral change through interactions, such as recommendations, and evaluating the user's buying intention and interest in each product. Using the evaluation results for recommendations helps achieve positive feedback and effectively support the selection of more exciting or different products. We propose a recommendation method specific to individual customers based on past transaction data, where changes can be monitored in real-time by reusing the knowledge acquired in advance through batch processing of knowledge discovery and data mining and processing the stream data in real-time online. We will present the implementation of our proposed method targeting the database system and machine learning algorithm.