Online Multiclass Learning with "Bandit" Feedback under a Confidence-Weighted Approach

Data volume has been increasing explosively in recent years and learning methods are vitally important to extract key information in such mass data. Traditional offline learning requires multiple traversals to the dataset, thus frequently suffering from lack of computational resources. Online learning can benefit in shrinking total time consumed by training model and lowering computational capacity. However they often converge slowly due to memory loss. Considering partial feedback, we uniquely propose online Confidence- Weighted learning in Bandit setting (CWB) for lower cumulative error and higher convergence rate. Specifically, historical information is preserved to adjust the weights of features for speeding up the convergence rate. Moreover, we novelly integrate the random sampling into the confidence-weighted learning, which can balance the exploitation and exploration in bandit setting. Extensive experiments demonstrate efficiency and effectiveness of our proposed scheme.