Awesome Open Source
Awesome Open Source


PyPI - Python Version Version GitHub repo size GitHub

Docs | Paper


Elliot is a comprehensive recommendation framework that analyzes the recommendation problem from the researcher's perspective. It conducts a whole experiment, from dataset loading to results gathering. The core idea is to feed the system with a simple and straightforward configuration file that drives the framework through the experimental setting choices. Elliot untangles the complexity of combining splitting strategies, hyperparameter model optimization, model training, and the generation of reports of the experimental results.

system schema

The framework loads, filters, and splits the data considering a vast set of strategies (splitting methods and filtering approaches, from temporal training-test splitting to nested K-folds Cross-Validation). Elliot optimizes hyperparameters for several recommendation algorithms, selects the best models, compares them with the baselines providing intra-model statistics, computes metrics spanning from accuracy to beyond-accuracy, bias, and fairness, and conducts statistical analysis (Wilcoxon and Paired t-test).

Elliot aims to keep the entire experiment reproducible and put the user in control of the framework.

We did our best to put all the bibliographic information of the methods, techniques, and metrics available in Elliot to give the right credits to their authors. Please, remember to cite them when you use Elliot in your research.

Actually, the same holds also for Elliot :-) If you find Elliot useful for your research or development, remember to cite the following paper:

  author    = {Vito Walter Anelli and
               Alejandro Bellog{\'{\i}}n and
               Antonio Ferrara and
               Daniele Malitesta and
               Felice Antonio Merra and
               Claudio Pomo and
               Francesco Maria Donini and
               Tommaso Di Noia},
  editor    = {Fernando Diaz and
               Chirag Shah and
               Torsten Suel and
               Pablo Castells and
               Rosie Jones and
               Tetsuya Sakai},
  title     = {Elliot: {A} Comprehensive and Rigorous Framework for Reproducible
               Recommender Systems Evaluation},
  booktitle = {{SIGIR} '21: The 44th International {ACM} {SIGIR} Conference on Research
               and Development in Information Retrieval, Virtual Event, Canada, July
               11-15, 2021},
  pages     = {2405--2414},
  publisher = {{ACM}},
  year      = {2021},
  url       = {},
  doi       = {10.1145/3404835.3463245},
  timestamp = {Thu, 15 Jul 2021 15:30:48 +0200},
  biburl    = {},
  bibsource = {dblp computer science bibliography,}


Elliot works with the following operating systems:

  • Linux
  • Windows 10
  • macOS X

Elliot requires Python version 3.6 or later.

Elliot requires tensorflow version 2.3.2 or later. If you want to use Elliot with GPU, please ensure that CUDA or cudatoolkit version is 7.6 or later. This requires NVIDIA driver version >= 10.1 (for Linux and Windows10).

Please refer to this document for further working configurations.

Install from source


git clone && cd elliot
conda create --name elliot_env python=3.8
conda activate elliot_env
pip install --upgrade pip
pip install -e . --verbose


git clone && cd elliot
virtualenv -p /usr/bin/python3.6 venv # your python location and version
source venv/bin/activate
pip install --upgrade pip
pip install -e . --verbose

Quick Start

Elliot's entry point is the function run_experiment, which accepts a configuration file that drives the whole experiment. In the following, a sample configuration file is shown to demonstrate how a sample and explicit structure can generate a rigorous experiment.

from import run_experiment


The following file is a simple configuration for an experimental setup. It contains all the instructions to get the MovieLens-1M catalog from a specific path and perform a train test split in a random sample way with a ratio of 20%.

This experiment provides a hyperparameter optimization with a grid search strategy for an Item-KNN model. Indeed, it is seen that the possible values of neighbors are closed in squared brackets. It indicates that two different models equipped with two different neighbors' values will be trained and compared to select the best configuration. Moreover, this configuration obliges Elliot to save the recommendation lists with at most 10 items per user as suggest by top_k property.

In this basic experiment, only a simple metric is considered in the final evaluation study. The candidate metric is nDCG for a cutoff equal to top_k, unless otherwise noted.

  dataset: movielens_1m
    strategy: dataset
    dataset_path: ../data/movielens_1m/dataset.tsv
      strategy: random_subsampling
      test_ratio: 0.2
        hyper_opt_alg: grid
        save_recs: True
      neighbors: [50, 100]
      similarity: cosine
    simple_metrics: [nDCG]
  top_k: 10

If you want to explore a basic configuration, and an advanced configuration, please refer to:



You can find the full description of the two experiments in the paper.


There are many ways to contribute to Elliot! You can contribute code, make improvements to the documentation, report or investigate bugs and issues

We welcome all contributions from bug fixes to new features and extensions.

Feel free to share with us your custom configuration files. We are creating a vault of reproducible experiments, and we would be glad of mentioning your contribution.

Reference Elliot in your blogs, papers, and articles.

Talk about Elliot on social media with the hashtag #elliotrs.

The Team

Elliot is developed by

It is maintained by SisInfLab Group and Information Retrieval Group.

* Corresponding authors


ELLIOT uses APACHE2 License.


SliM and an alternative KNN-CF implementation refer to RecSys2019_DeepLearning_Evaluation

Related Awesome Lists
Top Programming Languages

Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
Python (890,046
Machine Learning (40,892
Keras (10,610
Recommendation (4,994
Recommendation System (2,975
Slim (2,579
Tensorflow2 (1,200
Vae (1,120
Knowledge Graph (1,085
Collaborative Filtering (734
Matrix Factorization (568
Knn (441
K Nearest Neighbours (91
Content Based Recommendation (72
Deepfm (38
Neural Collaborative Filtering (19
Nfm (14
Funksvd (7
Bprmf (5
Svdpp (5