New release of Softconfidenceweighted.jl I just implemented Exact Soft Confidence Weighted <http://arxiv.org/pdf/1206.4612v1.pdf>(SCW) in Julia, and published as a package Softconfidenceweighted.jl <https://github.com/IshitaTakeshi/SoftConfidenceWeighted.jl> (still not registered, but will be soon).
SoftConfidenceWeighted.jl performs online binary linear classification with these advantages: - Large margin training - Confidence weighting - Capability to handle non-separable data - Adaptive margin Since SCW is an online learning algorithm, this also has the characteristics such as rapid execution with low memory usage. Performance comparison with LinearSVC in scikit-learn All code used in this experiment is here <https://github.com/IshitaTakeshi/SoftConfidenceWeighted.jl/tree/master/profile>. I just compared the performance of LinearSVC in scikit-learn and SCW in this package, using the dataset generated by generate_dataset.py in my repository. Result LinearSVC $python3 linearsvc.py 136 function calls in 52.177 seconds Accuracy: 0.9203 SCW $julia profile.jl 5.785278 seconds (787.71 k allocations: 501.247 MB, 0.32% gc time) Accuracy: 0.90405 LinearSVC is faster than SCW on small datasets, but the execution of SCW becomes much faster compare to LinearSVC as the number of samples increases. Environment Julia: Version 0.4.0-dev OS: GNU/Linux CPU: Intel Core i5 Feedback and contributions are appreciated. --Ishita