A Machine-Learning-Based Action Recommender for Network Operation Centers

Failure management and cost-aware traffic engineering are two important tasks done in Network Operation Centers (NOC). These are performed by expert technicians who must carefully analyze the network state and the flow of incoming alarms to decide how, where and when to take actions on the network. While based on implicit guiding principles, these network actions are very hard to automate with explicit rules due to the high complexity of the system; hence NOC action is essentially a manual process today. To automate part of that process, in this paper we introduce an Action Recommendation Engine (ARE) that can learn implicit NOC action rules with supervised machine learning from historical data. As a result, ARE can recommend suitable action(s) to remedy network faults and engineer the traffic to minimize costs, all while maximizing the users' Quality of Experience. To quantify the effectiveness of different NOC action scenarios, we introduce the QoE-OPEX metric which balances between users' quality of Experience and ISP's operational costs. After proper model training on 56,000 data points with 66 features, we demonstrate that ARE can effectively reproduce implicit action-taking logic of NOC technicians, thus moving us one step closer to reliable autonomous networks and fully-automated NOCs.