This repository contains the code used for the experiments in "Policy-Aware Unbiased Learning to Rank for Top-k Rankings" published at SIGIR 2020. The paper is available (here)[https://dl.acm.org/doi/abs/10.1145/3397271.3401102], alternatively, a pre-print can be found (here)[https://arxiv.org/abs/2005.09035].

If you use this code to produce results for your scientific publication, or if you share a copy or fork, please refer to our SIGIR 2020 paper:

@inproceedings{Oosterhuis2020Unbiased,
  title={Policy-Aware Unbiased Learning to Rank for Top-k Rankings},
  author={Oosterhuis, Harrie and de Rijke, Maarten},
  booktitle={Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year={2020},
  organization={ACM}
}

The contents of this repository are licensed under the MIT license. If you modify its contents in any way, please link back to this repository.

This code makes use of Python 3 the numpy and sharedmem packages, make sure they are installed.

Then a file is required that explains the location and details of the LTR datasets available on the system, for the Yahoo! Webscope, MSLR-Web30k, and Istella datasets an example file is available. Copy the file:

cp example_datasets_info.txt datasets_info.txt

Open this copy and edit the paths to the folders where the train/test/vali files are placed.

Here are some command-line examples that illustrate how the results in the paper can be replicated. First create a folder to store the resulting models:

mkdir local_output

We will start by creating the pre-trained model, trained on 1% of training data:

python3 pretrain.py local_output/pretrained_model.txt --dataset_info_path datasets_info.txt  --dataset Webscope_C14_Set1 --num_proc 1

To evaluate the resulting model, the following command will print a large number of metrics:

python3 evaluate.py local_output/pretrained_model.txt --dataset_info_path datasets_info.txt --dataset Webscope_C14_Set1 --dataset_partition test 

Then we can train the full-information model (note that --num_proc can be set to the number of processess available for multiprocessing, to speed things up):

python3 fulltrain.py local_output/full_information_model.txt --loss lambdaloss-truncated --dataset_info_path datasets_info.txt  --dataset Webscope_C14_Set1 --num_proc 1

Again we can evaluate the resulting model, which should perform much better than the pre-trained model:

python3 evaluate.py local_output/full_information_model.txt --dataset_info_path datasets_info.txt --dataset Webscope_C14_Set1 --dataset_partition test 

To reproduce Figure 1 and 2 you can use topktrain.py. The following command will simulate 10000 clicks using a pretrained model on a top-5 rankings, and subsequently, apply the Policy-Aware Estimator (with replace-last randomization) to optimize the truncated LambdaLoss:

python3 topktrain.py replacelast_policyaware  output/Webscope_C14_Set1/supervised/pretrained/pretrained_model.txt local_output/model.txt --loss lambdaloss-truncated --dataset_info_path datasets_info.txt  --dataset Webscope_C14_Set1 --num_proc 1 --num_clicks 10000 --cutoff 5

The same setup but optimizing ARP only requires a change of the --loss flag:

python3 topktrain.py replacelast_policyaware  output/Webscope_C14_Set1/supervised/pretrained/pretrained_model.txt local_output/model.txt --loss relevant_rank --dataset_info_path datasets_info.txt  --dataset Webscope_C14_Set1 --num_proc 1 --num_clicks 10000 --cutoff 5

Optimizing ARP with the Policy-Oblivious Estimator (with replace-last randomization) only requires a change of the first argument:

python3 topktrain.py replacelast_oblivious  output/Webscope_C14_Set1/supervised/pretrained/pretrained_model.txt local_output/model.txt --loss relevant_rank --dataset_info_path datasets_info.txt  --dataset Webscope_C14_Set1 --num_proc 1 --num_clicks 10000 --cutoff 5

To reproduce Figure 3 you can use dcgtrain.py, this code always applies the Policy-Aware Estimator (with replace-last randomization) but allows you to optimize different losses for DCG optimization. The following optimizes the truncated-lambdaloss:

python3 dcgtrain.py lambdaloss-truncated output/Webscope_C14_Set1/supervised/pretrained/pretrained_model.txt local_output/model.txt --dataset_info_path datasets_info.txt  --dataset Webscope_C14_Set1 --num_proc 1 --num_clicks 10000 --cutoff 5

Again evaluation can be done with:

python3 evaluate.py local_output/model.txt --dataset_info_path datasets_info.txt --dataset Webscope_C14_Set1 --dataset_partition test 

The following estimators and randomization options are included:

1. deterministic - no randomization and full-ranking i.e. no cutoff at top-k
2. deterministic_cutoff - no randomization and top-k ranking and the policy-oblivious estimator
3. deterministic_cutoff_rerank - no randomization and top-k ranking and the reranking estimator
4. replacelast_oblivious - replace-last randomization and the policy-oblivious estimator
5. replacelast_oblivious_rerank - replace-last randomization and the reranking estimator
6. replacelast_policyaware - replace-last randomization and the policy-aware estimator

Adding _clipped at the end of these options also enables propensity clipping, this is recommended for variance reduction.

The following loss function options are included:

1. monotonic
2. log_monotonic
3. lambdaloss-full
4. lambdaloss@k
5. lambdaloss-truncated
6. relevant_rank

All options are implemented for dcgtrain.py, topktrain.py only supports lambdaloss-truncated and relevant_rank.