Deep Reinforcement Learning of Marked Temporal Point Processes

This is the code produced as part of the paper Deep Reinforcement Learning of Marked Temporal Point Processes

"Deep Reinforcement Learning of Marked Temporal Point Processes" Utkarsh Upadhyay, Abir De, Manuel Gomez-Rodriguez. NeurIPS 2018. arXiv:1805.09360

Packages needed

• tensorflow 1.8.0
• numpy 1.14.3
• pandas 0.22.0
• pip install decorated_options

Needed for Smart Broadcaster experiments, and the RedQueen baseline:

• pip install git+https://github.com/Networks-Learning/RedQueen.git@master#egg=redqueen

Needed for the Karimi baseline:

• pip install git+https://github.com/Networks-Learning/broadcast_ref.git@master#egg=broadcast_ref

We obtained the parameters from the item difficulties via personal correspondence with the authors of MEMORIZE.

Experiment execution

Running experiments:

sbatch/exp_run.py

This script is used for running Smart Broadcasting experiments on a SLURM cluster using the job scripts sbatch/r_2_job.sh and sbatch/top_k_job.sh.

The job scripts assumes that there is a conda environment with the name tf-cpu and that the code resides in ${HOME}/prog/work/broadcast-rl/ folder. These details can be edited easily to match with any host's configuration via the scripts themselves.

train-broadcasting.py

This script is used for running an experiment for one-user in the Smart Broadcasting setup.

It can be executed as:

export USER_IDX=218   # Which user to train the model for.
mkdir -p output-smart-broadcasting/
python train-broadcastring ./data/twitter_data.dill ${USER_IDX} ./output-smart-broadcasting \
  --reward r_2_reward --q 100.0 --algo-feed --save-every 100 \
  --no-merge-sinks

If the ${USER_IDX} variable is looped over the list of indexes in ./data/r_2.csv (or ./data/top_k.csv), then we can reproduce the trained network for all 100 users used for experiments in the paper.

The output will be saved in files ./output-dir/train-save-user_idx-${USER_IDX}.

analyze-broadcasting.py

Reads output produced by train-broadcasting.py and compares it against baselines and saves the results in a CSV file ready for analysis/plotting.

python analyze-broadcasting.py ./data/twitter_data.dill ./output-smart-broadcasting/ broadcasting-analysis.csv  \
        --no-merge-sinks --algo-feed

train-teaching.py

This script is used for running the Spaced Repetition experiment:

mkdir -p output-spaced-repetition;
python train-teaching.py ./data/initial-difficulty.csv 0.049 0.0052 ./output-spaced-repetition \
    --q 0.0001 --q-entropy 0.005

If the training gets stuck on 0 reward, then training an initial version --with-recall-probs as reward instead of binary reward, and/or increasing the batch_size may help.

Reproducing figures

Smart broadcasting

Running:

python plot-smart-broadcasting.py broadcasting-analysis.csv

will reproduce the plots in the paper in the ./output-plots folder.

Spaced repetition

Running:

python analyze-spaced-repetition.py ./data/initial_difficulty.csv 0.049 0.0052 ./output-spaced-repetition

will re-produce the plots in the paper in the ./output-plots folder.