ADAPTATION OF LAMBDAMART MODEL TO SEMI-SUPERVISED LEARNING

The problem of information searching is very common in the age internet and Big Data. Usually, there are huge collections documents only multiple percent them relevant. In this setup brute-force methods useless. Search engines help to solve optimally. Most based on learning rank methods, i.e. first all algorithm produce scores for they feature after that sorts according score an appropriate order. There a lot algorithms area, but one most fastest robust ranking LambdaMART. This boosting developed supervised learning, where each document collection has estimated by expert. But usually, contain tons their annotation requires resources like time, money, experts, etc. case, semi-supervised powerful approach. Semi-supervised approach machine combines small amount labeled data with large unlabeled during training. Unlabeled data, when used combination quantity can significant improvement accuracy. paper dedicated adaptation LambdaMART learning. author proposes add different weights training procedure achieve higher robustness proposed was implemented using Python programming language LightGBM framework already implementation For testing purposes, datasets were used. One synthetic 2D dataset visual explanation results two real-world MSLR-WEB10K Microsoft Yahoo LTRC.