Semi-synthetic experiments to test several approaches for off-policy evaluation and optimization of slate recommenders.
Contact: Adith Swaminathan (adswamin@microsoft.com)
These python scripts and classes run semi-synthetic experiments on the MSLR and MQ datasets to study off-policy estimators for the slate bandit problem (combinatorial contextual bandits).
For Evaluation experiments: Usage: python Parallel.py Refer Parallel.py::main for examples on how to set up other variants of experiments The make_parallel_eval.sh bash script creates the entire suite of experiments reported in [1] as parallel cluster jobs.
Data: MSLR-30K has 31K queries, each with up to 1251 judged documents on relevance scale of {0, 1, 2, 3, 4} MSLR-10K has 10K queries, each with up to 908 judged documents on relevance scale of {0, 1, 2, 3, 4} Both MSLR datasets have <query, document> features of dimension 136
MQ2007 has 1692 queries, each with between 6 and 147 documents judged on relevance scale of {0, 1, 2} MQ2008 has 784 queries, each with between 5 and 121 documents judged on relevance scale of {0, 1, 2} Both datasets have <query, document> feature vectors of dimension 46
Refer Datasets.py::main for how to read in these datasets Download and uncompress the dataset files in the ../../Data/ folder MSLR: https://www.microsoft.com/en-us/research/project/mslr/ MQ: https://www.microsoft.com/en-us/research/project/letor-learning-rank-information-retrieval/ After reading in the uncompressed datasets once, the Datasets script creates *.npz files in the ../../Data/ folder as pre-processed numpy arrays for faster consumption by other scripts
For Optimization experiments: Usage: python Optimization.py Refer Optimization.py::main for examples on how to set up other variants of experiments The make_parallel_opt.sh bash script creates the entire suite of experiments reported in [1] as parallel cluster jobs.
[1] Off policy evaluation for slate recommendation, https://arxiv.org/abs/1605.04812 ; https://nips.cc/Conferences/2017/Schedule?showEvent=9146