Ranking with Long-Term Constraints

This repository contains the official implementation of the paper [Ranking with Long-Term Constraints].

Enviroment Setup

Python 3.10

• Create virtual Enviroment with virtualenv, conda or ...

• Install required libraries

pip install -r requirements.txt

Environment Configs

usage: simulate.py [-h] --conf CONF --dataset DATASET --ctrl {mc,sc,schinge,oracle,pc,pchinge,base} [--seed SEED]
                   [--lmbda LMBDA] [--gamma GAMMA] [--output_dir OUTPUT_DIR] [--c C [C ...]]
                   [--b B] [--bo BO] [--relevance_type {offline_relevance,online_relevance,sequence_relevance}]
                   [--lr LR] [--init {one,zero}] [--metrics_file_name METRICS_FILE_NAME]
                   [--targets TARGETS [TARGETS ...]] [--dev] [--hinge_min HINGE_MIN]
                   [--shuffle_bootstraps SHUFFLE_BOOTSTRAPS] [--eps EPS] [--beta BETA]

Simulate the online ranking systemunder differenct control policies

options:
  -h, --help            show this help message and exit
  --conf CONF           config file
  --dataset DATASET     dataset
  --ctrl {mc,sc,schinge,oracle,pc,pchinge,base}
                        controller
  --seed SEED           random seed
  --lmbda LMBDA         lambda used in P-Controller
  --gamma GAMMA         gamma used in BP-Controller
  --output_dir OUTPUT_DIR
                        Output of mterics
  --c C [C ...]         C in weighted objective
  --b B                 B offline in PC with bootstrap
  --bo BO               B online in PC with bootstrap
  --relevance_type {offline_relevance,online_relevance,sequence_relevance}
  --lr LR       learning rate
  --init {one,zero} init value
  --metrics_file_name METRICS_FILE_NAME
  --targets TARGETS [TARGETS ...]
  --dev
  --hinge_min HINGE_MIN
  --shuffle_bootstraps SHUFFLE_BOOTSTRAPS
  --eps EPS
  --beta BETA

Datasets <dataset>:

KuaiRec: kuai
Linear television dataset Tv Audience: zf_tv
Last.fm: lastfm_len_top_10
Fully-synthetic: early_and_late

Datasets-Configurations <conf>:

KuaiRec: experiments/multi_group.yml
Linear television dataset Tv Audience: experiments/zf_tv_single.yml
Last.fm: experiments/multi_lastfm_len_top_10.yml.yml
Fully-synthetic: experiments/multi_early_and_late_group.yml

Controllers <ctrl>:

Myopic Controller: mc
Stationary Controller: sc
Oracle: oracle
Predictive Controller: pc
Myopic Controller with out constraints: base

Example:

The below command works for the mc, sc, oracle, and base controllers:

python -u simulate.py 
    --conf <conf>
    --dataset <dataset>  
    --ctrl <ctrl>
    --output_dir <output-directory>
    --c 10. 10. 
    --metrics_file_name <job-name>  
    --shuffle_bootstraps true

The below command works for pc controller:

python -u simulate.py 
    --conf  <conf>
    --dataset <dataset> 
    --ctrl <ctrl> 
    --output_dir  <output-directory> 
    --c 10. 10.  
    --metrics_file_name <job-name> 
    --shuffle_bootstraps false 
    --relevance_type offline_relevance 
    --b <B offline in PC with bootstrap>
    --bo <B online in PC with bootstrap>

Citation

@article{Brantley2023RankingWL,
  title={Ranking with Long-Term Constraints},
  author={Kiant{\'e} Brantley and Zhichong Fang and Sarah Dean and Thorsten Joachims},
  journal={ArXiv},
  year={2023},
}