The efficiency of elastic optical networks (EONs) can be improved if services are provisioned based on traffic patterns and dynamic reconfiguration of network resources, reducing network expenses by cutting down on wasteful over-provisioning of network resources. Recently, there have been advancements in using machine learning (ML) in EONs to enable automated operations. This paper proposes an ML-assisted dynamic spectrum partitioning-based allocation scheme, named ML-DSP, for EONs to suppress blocking performance. The ML model predicts the request pattern of each source-destination pair in the network for each time interval. An integer linear programming (ILP) is formulated to optimally partition the spectrum into different partitions for each demand type using the ML prediction to enhance the resource utilization. When ILP is not tractable due to scalability, we introduce a heuristic for partitioning the spectrum. The first-last fit (FLF) spectrum allocation policy is used to allocate the connection requests. The simulation results indicate that ML-DSP outperforms the conventional static partitioning-based scheme that uses the first fit allocation policy in terms of blocking ratio, fairness, and spectrum time utilization ratio.