Cloud storage services provide both individual and business users with valuable functionalities like backup, data replication across different locations, data sharing, and collaborative work. These services often feature tiered storage options with various pricing models based on the classification of stored objects. Our research explores an increasingly important aspect for users: accurately predicting whether objects will be frequently or infrequently accessed, and thereby assigning them to the most cost-effective storage tier. We introduce a machine learning framework designed to dynamically determine the suitable class for objects based on their access patterns, achieving an accuracy rate surpassing 80\%. This framework's effectiveness is assessed through trace-driven simulations utilizing Dropbox data. The findings indicate that our approach can reduce storage costs by up to 37\% compared to other data allocation strategies reported in recent research.