Learning to Recommend using a Deep Reinforcement Agent

This is a project developed as part of my thesis dissertation for the M.Sc programme in Web Science and Big Data Analytics at University College London.

It implements a Deep Reinforcement Learning algorithm that learns to give good recommendations to users under a simulated environment built using the OpenAI gym framework. For more information, refer to the main document.

---

Author

• Santiago Gonzalez Toral | M.Sc Web Science and Big Data Analytics (WSBDA) candidate

Supervisor

• Dr. Jun Wang | Reader at University College London. Programme Director of M.Sc WSBDA

Overview

1. Project-Dependencies
2. Setup
3. Datasets
4. Execution
5. License

---

Project-Dependencies

The following packages are needed to execute the project models

• Python 2.7
• Virtualenv
• OpenAI gym (included in this project)
• Tensorflow
• Jupyter notebook
• Graphlab Create API 2.1 (with Academic License)

Setup

To setup the environment, the following commands need to be executed:

$ git clone https://github.com/santteegt/ucl-drl-msc.git

$ cd ucl-drl-msc
$ git submodule update --init --recursive
$ mkdir .venv
$ virtualenv --system-site-packages --python=python2.7 .venv/
$ source .venv/bin/activate

(venv)$ cd gym
(venv)$ pip install -e .
# installation for Mac OS X. For other platforms, refer to https://www.tensorflow.org/versions/r0.9/get_started/os_setup.html#virtualenv-installation
(venv)$ export TF_BINARY_URL=https://storage.googleapis.com/tensorflow/mac/tensorflow-0.9.0-py2-none-any.whl
(venv)$ pip install --upgrade $TF_BINARY_URL
(venv)$ pip install pymongo pandas gensim fastFM matplotlib scikit-learn scipy
(venv)$ pip install --upgrade --no-cache-dir https://get.graphlab.com/GraphLab-Create/2.1/hernan.toral.15@ucl.ac.uk/55E9-2088-3AF8-120F-D171-1AAB-95A3-B077/GraphLab-Create-License.tar.gz

Datasets

Movielens datasets files were pre-processed for running the models. They can be downloaded from the following URL: ( Dataset files ) | ( Alternative 2 )

Then,

• Extract and copy the contents of FMmodel.zip to the data folder
• To use the Movielens-100k dataset, extract the contents of the data-movielens100k.zip file into the data folder
• To use the Movielens-1M dataset, extract the contents of the data-movielens1m.zip file into the data folder

Execution

To make the running process more easier, executable files for each experiment are provided under the bin folder (make sure to activate the virtualenv before running the exec file):

1. DRL-kNN-CB: run-v2.sh runs the model for the Movielens 100k dataset. run-v0.sh runs the model for the Movielens 1M dataset
2. DRL-kNN-CF: run-v1.sh runs the model for the Movielens 100k dataset.
3. DRL-FM: run-v3.sh runs the model for the Movielens 100k dataset.
4. Random agent: run-v1-random.sh run-v2-random.sh runs the DDPG algorithm using a random agent

Human readable rendering of recommendations made while running the model can be seen using the following command:

(venv)$ cd ucl-drl-msc
(venv)$ tail -F run.log

Reward logs obtained by an agent while running the model can be read using the following command:

(venv)$ cd ucl-drl-msc
(venv)$ tail -F ddpg-results/experiment1/output.log

Additionally, to visualize the model parameters during training, you can run the Tensorflow dashboard using the following command:

(venv)$ cd ucl-drl-msc
(venv)$ python src/dashboard.py --exdir ddpg-results/

Finally, to stop the training process, it is necessary to run the following command:

(venv)$ cd ucl-drl-msc/bin
(venv)$ sh stop.sh

License

The Project code is mostly developed under the Apache2 license, except for the module developed using Graphlab Create which is based on an Academic License provided by Turi