Code and data for "Universal Approximation Functions for Fast Learning to Rank: Replacing Expensive Regression Forests with Simple Feed-Forward Networks"
Switch branches/tags
Nothing to show
Clone or download
Latest commit 966d6a3 Jul 11, 2018


Code and data to train a simple multilayer perceptron (feed-forward neural network) in pytorch from a regression forest learning to rank model (e.g., LambdaMART, RF, GBDT, etc.).

This code is a re-implementation of experiments done for the following SIGIR paper:

Cohen, D., Foley, J., Zamani, H., Allan, J. and Croft, W. B. , "Universal Approximation Functions for Fast Learning to Rank," to appear in the Proceedings of the 41st International ACM SIGIR Conference on Research and Development in Information Retrieval, Ann Arbor, MI, Jul. 8-12 2018 (SIGIR’18)


The paper uses the LTR dataset from Microsoft: MSN30k dataset.

The Gov2/MQ07 data for the paper will be up shortly. The features for MQ07/Gov2 we used are about 1.5GB compressed, and we need to figure out a good way to host that much data.