Bandwidth fragmentation is a critical problem for Elastic Optical Networks (EON), and spectrum defragmentation is the most important strategy to mitigate this phenomenon. In this work we propose a Machine Learning (ML) based method for estimating the Blocking Rate, which, when exceeding a threshold, triggers a defragmentation process. This is done in order to achieve better results in terms of the number of blocking demands and the number of re-routed connections. The performance of the proposed method was compared with two other known strategies: fixed-time (FT) defragmentation, and triggering based on one fragmentation metric (BFR). Simulation results were evaluated using two multi-objective metrics. Experimental results show that the proposed method is more efficient than the other two, being the best method in 85.7% of comparisons using the Pareto Coverage metric, and obtaining 47.4% of non-dominated solutions in the Pareto Front.