SVM at Edge: Low Cost Caching Prediction for Connected Edge Intelligence in Federated Machine Learning

This research work targets a problem in edge intelligent devices where caching becomes challenging task. We propose a Federated Learning (FL) technique that uses local model as Support Vector Machines (SVM) and global model as a Neural Network. Each participant in one FL run would train a SVM on its local data and use it to make predictions. Further, local aggregated predictions and learned SVM boundary used to train a global NN model. SVM is a simple and fast algorithm that can work well on small, local datasets, while Neural Networks are capable of modeling more complex relationships in data and can achieve higher accuracy on large, global datasets. The combination of these two models balances the trade-off between local accuracy and global accuracy, while preserving privacy by only sharing predictions. Results of this approach demonstrate its potential to provide a more accurate representation of data across multiple participants in a federated learning setup to predict cache and increase cache hit.