RecSim:ACofigurableRecommederSystemsSimulatioPlatform
RecSimisacofigurableplatformforauthorigsimulatioevirometsforrecommedersystems(RSs)thataturallysupportssequetialiteractiowithusers.RecSimallowsthecreatioofewevirometsthatreflectparticularaspectsofuserbehavioraditemstructureatalevelofabstractiowell-suitedtopushigthelimitsofcurretreiforcemetlearig(RL)adRStechiquesisequetialiteractiverecommedatioproblems.Evirometscabeeasilycofiguredthatvaryassumptiosabout:userpreferecesaditemfamiliarity;userlatetstateaditsdyamics;adchoicemodelsadotheruserresposebehavior.WeoutliehowRecSimoffersvaluetoRLadRSresearchersadpractitioers,adhowitcaserveasavehicleforacademic-idustrialcollaboratio.ForadetaileddescriptiooftheRecSimarchitecturepleasereadIeetal.Pleasecitethepaperifyouusethecodefromthisrepositoryiyourwork.
Bibtex@article{ie2019recsim,title={RecSim:ACofigurableSimulatioPlatformforRecommederSystems},author={EugeeIeadChih-weiHsuadMartiMladeovadVihaJaiadSamitNarvekaradJigWagadRuiWuadCraigBoutilier},year={2019},eprit={1909.04847},archivePrefix={arXiv},primaryClass={cs.LG}}DisclaimerThisisotaofficiallysupportedGoogleproduct.
What'sew12/13/2019:Added(abstract)classesforbothmulti-userevirometsadagets.Addedbaditalgorithmsforgeeralizedliearmodels.IstallatioadSampleUsageItisrecommededtoistallRecSimusig(https://pypi.org/project/recsim/):
pipistallrecsimHowever,thelatestversioofDopamieisotiPyPIasofDecember,2019.WewattoistallthelatestversiofromDopamie'srepositorylikethefollowigbeforeweistallRecSim.NotethatDopamierequiresTesorflow1.15.0whichisthefial1.xreleaseicludigGPUsupportforUbutuadWidows.
pipistallgit+https://github.com/google/dopamie.gitHerearesomesamplecommadsyoucouldusefortestigtheistallatio:
gitcloehttps://github.com/google-research/recsimcdrecsim/recsimpythomai.py--logtostderr\--base_dir="/tmp/recsim/iterest_exploratio_full_slate_q"\--aget_ame=full_slate_q\--eviromet_ame=iterest_exploratio\--episode_log_file='episode_logs.tfrecord'\--gi_bidigs=simulator.ruer_lib.Ruer.max_steps_per_episode=100\--gi_bidigs=simulator.ruer_lib.TraiRuer.um_iteratios=10\--gi_bidigs=simulator.ruer_lib.TraiRuer.max_traiig_steps=100\--gi_bidigs=simulator.ruer_lib.EvalRuer.max_eval_episodes=5Youcouldthestartatesorboardadviewtheoutput
tesorboard--logdir=/tmp/recsim/iterest_exploratio_full_slate_q/--port=2222Youcouldalsofidthesimulatedlogsi/tmp/recsim/episode_logs.tfrecord
TutorialsTogetstarted,pleasecheckoutourColabtutorials.IRecSim:Overview,wegiveabriefoverviewaboutRecSim.Wethetalkabouteachcofigurablecompoet:evirometadrecommederaget.
DocumetatioPleaserefertothewhitepaperforthehigh-leveldesig.
评论