Course Project: First-Order Methods in Online Convex Optimization
Online convex optimization (OCO) deals with making sequential decisions in an uncertain environment where data arrives in a streaming fashion. Hence, it is necessary, and beneficial, to take a robust approach, by applying an optimization method that learns as more aspects of the problem are observed. This view of optimization as a process has led to successes in developing efficient algorithms for prominent online learning problems. In this project, we aim to understand the key concepts in online convex optimization, several classic first-order online convex optimization algorithms that are robust to adversarial attacks, and showcase the applicability of the framework to the online email classification task.
We demonstrate the potential of FTRL and OGD with the email categorization task by utilizing
a soft-margin linear SVM with hinge loss. Specifically, we have trained a online learning model
on the Spambase (spam_data.mat) dataset adapted from the UCI Machine Learning Repository
containing 4601 emails, each represented by 57 features. We then verified the theoretical regret bound
