On tuning a mean-field model for semi-supervised classification

Abstract Semi-supervised learning (SSL) has become an interesting research area due to its capacity for in scenarios where both labeled and unlabeled data are available. In this work, we focus on the task of transduction—when objective is label all presented learner—with a mean-field approximation Potts model. Aiming at particular study how classification results depend β find that optimal phase depends highly amount same study, also observe more stable classifications regarding small fluctuations related configurations high probability propose tuning approach based such observation. This method relies novel parameter γ then evaluate two different values said quantity comparison with classical methods field. evaluation conducted by changing available number nearest neighbors similarity graph. Empirical show effective allows NMF outperform other approaches datasets fewer classes. addition, one chosen leads resilient changes neighbors, which might be interest practitioners field SSL.