This is the source code for the paper titled "PairRank: Online Pairwise Learning to Rank by Divide-and-Conquer" which is published at the Web Conference 2021.

PairRank aims at learning a pairwise learning to rank model online. In each round, candidate documents are partitioned and ranked according to the model's confidence on the estimated pairwise rank order, and exploration is only performed on the uncertain pairs of documents, i.e., divide-and-conquer.
Regret directly defined on the number of mis-ordered pairs is proven, which connects the online solution's theoretical convergence with its expected ranking performance.

To run the code to generate experimental results like those found in our papers, you will need to run a command in the following format:

$ python run_script.py [--algo ALGO] [--data_sets DATASET] [--lr LEARNING_RATE] [--lr_decay LEARNING_RATE_DECAY]
                       [--lambda LAMBDA] [--alpha ALPHA] [--click_models CLICK_MODEL] [--n_impr N_IMPR] [--update UPDATE]
                       [--rank RANK] [--ind] [--refine] [--seed SEED] 

Example:

$ python run_script.py --algo pairrank --data_sets yahoo --click_models per --n_impr 5000 --lambda 0.1 --alpha 0.1 
                       --update gd --rank random --seed 0 --ind --refine

The detailed information about the input parameters can be found in /utils/argparsers/simulationargparser.py

In our paper, we have two widely used benchmark datasets: Yahoo! Learning to Rank dataset and MSLR-WEB10K dataset.