Skip to content

ContentWise/contentwise-impressions

master
Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.

ContentWise Impressions

This is the repository associated to our article "ContentWise Impressions: An Industrial Dataset with Impressions Included" accepted and presented at CIKM 2020. Full text is available on ACM DL, ArXiv, or ResearchGate.

How to download the dataset?

You can obtain the link to download the dataset by filling this form.

Filling the form is completely optional, and it won't block you from getting the link to download the dataset.

After you receive the dataset link, download the zip file and decompress it on your local environment.

You'll find a README.md file, that includes information about the dataset, authors, license, and more. You'll also find the data folder. Inside this folder you'll find the dataset (interactions, impressions-direct-link, and impressions-non-direct-link) alongside the URM splits that we used in our experiments. Moreover, if you wish to run the scripts inside the repository, you'll need the whole data folder.

Citation

If you use this dataset in a publication, please cite our CIKM paper:

Fernando B. Pérez Maurera, Maurizio Ferrari Dacrema, Lorenzo Saule, Mario Scriminaci, and Paolo Cremonesi. 2020. 
ContentWise Impressions: An Industrial Dataset with Impressions Included. 
In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM '20). 
Association for Computing Machinery, New York, NY, USA, 3093–3100. DOI:https://doi.org/10.1145/3340531.3412774

If you use BibTeX:

@inproceedings{contentwise-impressions,
 author = {P\'{e}rez Maurera, Fernando B. and Ferrari Dacrema, Maurizio and Saule, Lorenzo and Scriminaci, Mario and Cremonesi, Paolo},
 title = {ContentWise Impressions: An Industrial Dataset with Impressions Included},
 year = {2020},
 isbn = {9781450368599},
 publisher = {Association for Computing Machinery},
 address = {New York, NY, USA},
 url = {https://doi.org/10.1145/3340531.3412774},
 doi = {10.1145/3340531.3412774},
 booktitle = {Proceedings of the 29th ACM International Conference on Information & Knowledge Management},
 pages = {3093–3100},
 numpages = {8},
 keywords = {dataset, implicit feedback, impressions, collaborative filtering, open source},
 location = {Virtual Event, Ireland},
 series = {CIKM '20}
}

Full text is available on ArXiv, ResearchGate, or ACM DL. Source code of our experiments and results is available on GitHub.

Experiments Results

You can download the results of our experiments on this link. There you'll find two folders: statistics and result_experiments. The first folder contains the statistical features of the dataset alongside several plots, including others that didn't make it into the paper. The second folder contains all the fine-tuned trained recommender models.

Note: As we exported the models for several recommenders, the results folder takes approximately 2GB on disk.

In this repository we provide several tools to load and use the dataset. We strongly recommend you to go through the Installation and Using the repo sections to know which scripts we provide and how to run them.

Installation

Note: that this repository requires Python 3.7

First we suggest you create an environment for this project using conda. We have tested the installation procedures on Linux 64-bits (Ubuntu 18.04), macOS Catalina 10.15, and Windows 10.

First, install miniconda, instructions on how to install miniconda are found in Their docs

Second, clone this repository, checkout, and install the environment with the following:

git clone https://github.com/ContentWise/contentwise-impressions.git
cd contentwise-impressions
conda env create -f environment.yml
conda activate contentwise-impressions

Now, depending on your platform, there are special installation procedures that you need to perform. If you:

After you have performed your environment specific steps, continue to Compiling Cython.

[Linux] Install more dependencies

At this point, having installed all dependencies, you have to compile all Cython algorithms.

In order to compile you must first have installed: gcc and python3 dev. Under Linux those can be installed with the following commands:

sudo apt install gcc 
sudo apt-get install python3-dev
sudo apt-get install libopenblas-base libopenblas-dev

Now, continue to Compiling Cython.

[macOS] Install more dependencies

You must download Xcode, and the command-line tools in order to have a C compiler installed on your system. More information about Xcode is found on Apple docs.

Now, continue to Compiling Cython.

[Windows] Install more dependencies

If you are using Windows as operating system, the installation procedure is a bit more complex. You may refer to THIS guide.

Continue to Compiling Cython.

Compiling Cython code

Now you can compile all Cython algorithms by running the following command. The script will compile within the current active environment. The code has been developed for Linux and Windows platforms. During the compilation you may see some warnings. These are expected

(contentwise-impressions): python run_compile_all_cython.py

Place the data

Now that you have the environment set, download the dataset and the splits. Please place the data folder inside the repository folder.

After you've done this, you're ready to use the repo.

Using the repo

We have provided several python scripts that uses the dataset in different ways.

In the following sections we describe each script that we provide.

Generating URM splits

Prerequisites: You need to have the environment fully installed.

NOTE: On our tests, this process consumes up to 16GiB of RAM. Please ensure to have these resources or use our splits.

In order to download the data and generate the URM splits that we used in our experiments, you must use the run_generate_splits.py.

  • If it's run without arguments, it will download the interactions, interacted impressions, and non-interacted impressions.
  • If it's run with the -i or --items arguments, it will download the dataset and will generate three URM splits of the interactions Train, Validation, and Test. Using a proportion of 0.7, 0.1, and 0.2, respectively. Users are rows and Items are columns.
  • If it's run with the -s or --series arguments, it will download the dataset and will generate three URM splits of the interactions Train, Validation, and Test. Using a proportion of 0.7, 0.1, and 0.2, respectively. Users are rows and Series are columns.

Examples:

(contentwise-impressions): python run_generate_splits.py -i -s

Tuning Hyper-parameters of recommendation algorithms

Prerequisites: You need to have the environment fully installed, and the data splits, preferably.

NOTE: Depending on your environment and available resources, the process could get killed because of insufficient memory. We used an r4.4xlarge Linux Amazon EC2 Instance to run our experiments. It had 16vCPUs and 128 GiB of RAM. However, we utilized this type of instance to run several experiments on parallel. By our own calculations, running each recommender should take less than 20GiB of RAM using eight cores if the evaluation is done in parallel. Execution times for different recommenders vary.

In order to tune the hyper-parameters of several recommendation algorithms, you must use the run_hyperparameter_tuning.py script. You need to provide the -t or --tune_recommenders arguments in order to make it run. This is to ensure that you're willing to run the hyperparameter tuning.

We ran the experiments using the following recommenders:

  • Random: recommends a list of random items,
  • TopPop: recommends the most popular items,
  • ItemKNN: Item-based collaborative KNN,
  • RP3beta: collaborative graph-based algorithm with re-ranking,
  • PureSVD: SVD decomposition of the user-item matrix,
  • Impressions MatrixFactorization BPR (BPRMF): machine learning based matrix factorization optimizing ranking with BPR, with the possibility to sample negative items at random, inside the impressions, or outside the impressions.

Examples:

(contentwise-impressions): python run_hyperparameter_tuning.py -t

Gathering results

Prerequisites: You need to have the environment fully installed, the data splits, and the result_experiments folder in the repository folder.

The run_results_gathering.py script outputs a table with the hyperparameter tuning results. You need to provide the -s or --show_results arguments in order to make it run. This is to ensure that you're willing to run the script.

Examples:

(contentwise-impressions): python run_results_gathering.py -s

Obtain statistics of the dataset

Prerequisites: You need to have the environment fully installed, and the dataset saved (not necessarily with the splits).

In order to generate the statistics of the dataset, we provide a jupyter notebook, notebook_generate_statistics.ipynb, that lets you to generate several statistics of the dataset on the same place.

In order to run the code.

(contentwise-impressions): jupyter lab --no-browser

Inside the notebook, just run the different sections to obtain different statistics. We provide documentation of what kind of statistics we calculate. All the statistics and plots are generated into the statistics folder. This notebooks generates all of the plots, numbers and figures that we used in the paper.

Run the tests

Prerequisites: You need to have the environment fully installed, and the dataset saved (not necessarily with the splits).

We provide consistency tests of the dataset. It will check several properties of the dataset that are reported in the paper.

We use pytest as test runner. To run the tests is just as easy as to run the following:

(contentwise-impressions): pytest test_dataset_consistency.py --verbose --color=yes

This command doesn't write any report, instead it shows on the console the results of the tests in a PASS/FAIL fashion.

If you run into issues or want to ask us something

Please, don't hesitate to let us know by opening an issue on the Issue Tracker. We highly appreciate your feedback.

Closing remarks

Thanks for using ContentWise Impressions, this repo and supporting our work. We hope that it's useful for your purposes.

Disclaimer

This is not an official ContentWise product.

Contact information

For help or issues using ContentWise Impressions, please submit a GitHub issue.

For personal communication related to ContentWise Impressions, please contact:

About

This repository contains the code used to run generate the data splits, run the hyperparameter tunings, and export the results presented in our article "ContentWise Impressions: An industrial dataset with impressions included"

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published