Learning to select for a predefined ranking

Authors: Aleksei Ustimenko, Aleksandr Vorobev, Gleb Gusev, Pavel Serdyukov

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our offline and online experiments with a large-scale product search engine demonstrate the overwhelming advantage of our methods over the baselines in terms of all key quality metrics.
Researcher Affiliation Collaboration 1Yandex, Moscow, Russia 2Skoltech University, Moscow, Russia 3Faculty of Computer Science, Higher School of Economics, Moscow, Russia 4Department of Innovation and High Technology, Moscow Institute of Physics and Technology, Dolgoprudny, Russia.
Pseudocode No The paper refers to 'Algorithm 3 in (Prokhorenkova et al., 2018)' but does not contain its own pseudocode or algorithm blocks.
Open Source Code Yes The whole source code of our learning algorithm and its difference from Cat Boost release 0.10.04 are available10. Later, it was added to Cat Boost as Stochastic Filter loss11. 10https://github.com/Take Over/catboost/tree/0.10.4_release 11https://github.com/catboost/catboost/commit/df18d16
Open Datasets Yes The data with labels used for learning all the models and for evaluation by DCG-RR (see description of labels and metrics below) is available in open source 7. 7https://research.yandex.com/datasets/market
Dataset Splits Yes we randomly split all the queries from the collected dataset D into 5 parts of equal size to run 5-fold cross-validation: at each i of five runs, 80% of the queries (Di train) are used for training, 10% (Di valid) are used for tuning hyperparameters of the algorithms and 10% (Di test) are used for testing.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments.
Software Dependencies Yes We use GBDT implementation in opensourced Cat Boost9 Python package... The whole source code of our learning algorithm and its difference from Cat Boost release 0.10.04 are available10.
Experiment Setup No See Section 4 of the supplementary for Cat Boost parameters we used.