(Various) LogitBoost for multiclass classification with ensemble of vector-valued trees.
About one year after we published AOSOLogitBoost [1], we found that a simple LogitBoost variant also works very well for multiclass classification, sometimes even outperforms AOSOLogitBoost. This new variant, called VTLogitBoost, can be explained in a few lines of words if you are already very familiar with multiclass LogitBoost:
- The tree is still vector valued, as in AOSOLogitBoost
- At each node, all the
Kclasses are updated - The gain for node split is derived by diagonal approximation of the Hessian matrix
We never published VTLogitBoost anywhere, so we simply put the code here (for research purpose). In the folder there are also many other VTLogitBoost variants, including those with importance sampling, total correction, etc, although some of them see a degraded performance, which, we doubt, should attribute to our problematic implementation. However, the VTLogitBoost is guaranteed to improve over/be on par with AOSOLogitBoost.
The code is somewhat messy, but does the job. Also, it is not well documented, but the calling convention is similar to AOSOLogitBoost, see the doc there. Feel free to contact me if you need any further help.
[1] "Peng Sun, Mark D. Reid, Jie Zhou. "AOSO-LogitBoost: Adaptive One-Vs-One LogitBoost for Multi-Class Problems", International Conference on Machine Learning (ICML 2012)"