MetaForecast: Harnessing Model-Agnostic Meta-Learning Approach to Predict Key Metrics of Interconnected Network Topologies

Abstract

Meta-learning, an approach in machine learning that focuses on “learning how to learn prioritizes generalization over specialization, mirroring the human’s ability to derive generalizations from experiences and specialize when tasks are repeated. Training a meta-model requires the procurement of similar tasks or similar data distribution. In our study, we explored a model-agnostic meta-learning approach to predict telemetry data collected from a network of devices. We also proposed a custom architecture “MetaForecast” wherein a meta-learner learns the generalized intricacies of each site/device’s data, allowing us to fine-tune the base learner and create site/device-specific models. Based on our experiments, we have observed that by using MetaForecast in such complex telemetry system we can:1)Significantly reducing the training time of a forecasting model for newly added devices/sites: Our proposed approach enables fine-tuning to a site-specific model within a small number (less than 10) of epochs.2)Minimizing data gathering requirements: By requiring fewer epochs for model tuning, our approach greatly reduces the data gathering needs. Hence, a new site doesn’t necessitate extensive historical data beyond a few recent entries based on granularity.3)Enabling day 1 prediction: We assert that if a new site/device is added, the new model can be trained within a few epochs and doesn’t rely on a large amount of past data for training. To establish a baseline, we compared the performance of our MAML-inspired architecture to individual models per site and transfer learning. Our findings revealed approximately a ~ 10% reduction in mean squared error, a ~ 50% reduction in computing resources, and a ~ 65% reduction in data gathering requirements. Based on our comprehensive research, we assert that the integration of meta-learning techniques and our proposed architecture yields notable improvements in forecasting accuracy, accompanied by substantial reductions in training time, data requirements, and computing resources.