Progressive search personalization and privacy protection using federated learning

Abstract

AbstractPersonalized search comes under recommendation systems which provide more user‐focused query results. Applications such as movie recommendations, document recommendations, and so forth are using machine learning algorithms to suggest user‐tailored content. These algorithms train on historical user query data and predict documents based on user preferences. The existing methodologies on personalization have shown two main limitations: privacy concerns over the usage of personal data and a lack of efficient personalizing strategies for better accuracy and overall standardization. In this paper, we propose a novel federated learning (FL) algorithm for personalized search results using temporal characteristics of the user query data. Individual user query data is used for developing specific client models while the aggregate of such models is developed for general search suggestions. The introduction of time‐series interpretation of queries provides for larger training data as well as a better understanding of a user's current needs and intent. The proposed algorithm is validated using the AOL4PS dataset and is evaluated on the efficiency of personalization, and the amount of data and time required to achieve the results. Its performance is compared with existing state‐of‐the‐art personalization algorithms that utilize deep learning and FL. Mean reciprocal rank (MRR) is the primary metric for measuring the algorithm's performance. After training for 35 federated rounds, the server model yielded an MRR score of 0.8638 while the client models were able to yield an average MRR score of 0.9308.