The loss function used in ListMLE or ListNet are smooth and differentiable thus optimization of loss function is easier as compared to pairwise LTR algorithm. But is any increase in performance occur due to reducing the features in ListNet and ListMLE ?
Hi,
What is your loss function and what is the ListNet/ListMLE package/library do you use?
What do you mean by no. of features? is this no. of features in the training dataset or number of features that picked up for learning iterations?
If the whole training set, that has similar test set characteristics, can not be represented in each learning iteration (and this usually happens in most of the machine learning techniques for reducing the application runtime) the ranking model may over-fitting or under-fitting the predictive results in test set data by very low values as in ListNET. Also, bagging by taking average ranking models in Random Forest is a treatment for overfitting for smoothing but taking average models reduces the predictive and training values as well.
Furthermore, pairwise LTR approaches have additional issues compared to listwise LTR approaches.
Thanks
Osman
Hi,
As any learning method, the selection of relevant features is of major importance. About the listwise approaches, as preliminary experiments, you should check the pairwise correlation between features, then eliminate those with no correlation, also remove those with no variances.
There is no neural network technique that has no variance except if you want under-fitting model. Please, read about bias-variance trade-off. This problem caused by sampling methods in each learning iteration. Also, please check this link for over-fitting and under-fitting problem
http://machinelearningmastery.com/overfitting-and-underfitting-with-machine-learning-algorithms/
http://www.cs.bham.ac.uk/~jxb/INC/l9.pdf
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Algorithms