Traditional or default Software-Defined Networking (SDN) generally relies on reactive or static routing, where a new flow entry is added for a new arriving packet in a switch. This can create inefficiencies in processing dynamic network conditions. To solve this issue, a Machine Learning-based Proactive Rerouting Scheme (MLPRS) has been developed to dynamically balance load in real-time networks by rerouting flow and enhancing Quality of Service (QoS) compared to the default SDN. However, SDN can face challenges from conflicting flows occurring due to priority and action adjustments that may affect network throughput and bandwidth. Therefore, this article introduces an Online Active ML-based Conflict Flow Avoided Proactive Rerouting Algorithm (OAMLCFAPRA) to discriminate between normal and conflict flows in SDN based on the behavioural changes of flows over time. To achieve this, an online learning algorithm with a customized weighting scheme called Iterative Confidence-Weighted Learning (ICWL) is executed in the controller plane. Initially, it preprocesses the generated flows and trains the ICWL to identify them as normal and conflicting based on their behavioural features. Then, the priorities of each flow are assigned by the ICWL, and the normal flows are directed to OpenFlow. Additionally, the ICWL classifies the conflicting flows into several types based on their features such as priority, IP address, and action. Moreover, the betweenness centrality algorithm determines each link’s significance, and the load on each link is monitored. When the most significant link is overloaded, the flow is rerouted to prevent congestion in real-time. Finally, experimental results show that the ICWL achieves high accuracy in classifying conflict flow types and the OAMLCFAPRA increases network throughput and bandwidth compared to conventional algorithms.