Research @ Mangaki Recommandation d'anime et de mangas

Wanna join? Contact us or say hi on Twitter!

Mangaki Data Challenge avec Kyoto U !

Mangaki organise cet été un data challenge en partenariat avec le Kashima lab à l’université de Kyoto !


C’est ici ! Jusqu’au 15 septembre 2017.

Partagez le lien :

Sur Twitter ou sur Facebook !


Les participants devront déterminer si certains utilisateurs ont envie de lire certains mangas ou s’ils n’ont pas envie de les lire, à partir du profil de tous les utilisateurs du site.

Les données de Mangaki ont été évidemment anonymisées pour l’occasion.

Types de ratings dans le Mangaki Data Challenge

Chaque semaine, des notebooks pour aider à entrer dans les données seront postés sur notre GitHub.

Inscrivez-vous sur l’arène University of Big Data, la plateforme de compétitions de l’université de Kyoto.

Data challenge organized with Kashima's Lab, Kyoto University

Mangaki va organiser un data challenge avec le labo de machine learning du professeur Kashima à l’université de Kyoto.

Les données de Mangaki seront anonymisées pour l’occasion. On publiera donc un fichier sous cette forme :


Bref, 350000 lignes du genre : « La personne n° 320 a adoré le manga n° 24 ».

Le principe du concours, qui sera sur la plateforme University of Big Data :

  1. À partir de ce fichier qui contient 80 % des données de Mangaki
  2. Les participants devront programmer un système de recommandation
  3. Celui qui prédira le mieux les 20 % restants gagnera.

On espère ainsi faire découvrir l’IA (et Mangaki) à plus d’étudiants !

Si pour une raison quelconque vous ne souhaitez pas participer à cette aventure, vous pouvez retirer votre participation depuis votre profil. Sinon, merci de nous faire confiance !

Mangaki will be at the Anime & Manga Symposium @ Anime Expo, Los Angeles!

Le 2 juillet, nous étions à Anime Expo, Los Angeles ! Voir les slides.

Car oui, nous avons été acceptés pour faire une présentation à la conférence académique d’Anime Expo.

Everyone regularly ask themselves what movie, series or book they should watch next, according to their taste. Mangaki wants to innovate access to Japanese culture by providing a unique user experience through a recommender system.

When a user shows up, Mangaki asks them to rate a few works. Based on their answers, they receive tailored anime recommendations. Mangaki’s machine learning techniques attempt to “guess” the taste of new users, by geometrically positioning their ratings within those collected from other users. Indeed, we will show that a simple factor analysis on the anonymized data (325,000 ratings from 2,220 users and 15,000 works) can reveal interesting and counterintuitive categories of manga that are liked together.

Mangaki started as a French student project in 2014 (it is so much fun to conduct research when it is unleashed on real data!). It received a prize from Microsoft (2015) and the Japanese Cultural Institute in Paris (2016): we won a trip to Tokyo to meet Japanese companies and investors, who were surprised to learn that we wanted to stay non-profit. The founder now holds a PhD in CS and works as a researcher in RIKEN, Tokyo.

In its mission to promote transparency and education, all the code of the Mangaki platform is open source. Consequently, it becomes possible for anyone to understand better the algorithms behind recommender systems. We regularly hold conferences for students, from high school to master’s degree.

The Mangaki dataset will also be released for academic purposes (humanities: dōzo!), while respecting the users’ privacy. We are currently organizing a data challenge with Kashima’s Machine Learning Lab in Kyoto University, where students will have to improve the accuracy of the recommendations. There is plenty of exciting research yet to be made, notably the automatic detection of NSFW posters using deep learning.

Know more: presentation.pdf
Demo: MP4

Test our algorithms on Movielens!

You can run our 5-fold cross validation on the Movielens dataset.

Download the Movielens dataset prepared by our team:

Clone the GitHub repo and:

git clone
cd mangaki
python3 -m venv venv
. venv/bin/activate
pip install -r requirements.txt
pip install -r requirements/dev.txt
# Put the ratings-ml.csv file in the data folder
cd mangaki
cp settings.template.ini settings.ini
./ compare movies

It will run everything and display:

Final results
als-20: RMSE = 1.122326
svd-20: RMSE = 1.157234

Feel free to modify the mangaki/mangaki/management/commands/ file to compare more algorithms.

In-flight entertainment systems

Je m’intéresse aux in-flight entertainment systems.

As these systems evolve, recommendations could become more sophisticated. “For example, [the system could say,] ‘The last time you were on board, you watched this movie, but you didn’t finish it. Would you like to finish it now?’” Rhoads suggests. “We can say: ‘The last 15 movies that you watched on flights for the past two months have been around these characters or themes. Here are recommendations from this month’s movie selections.’ These are things that are relatively easy to implement and we’re already seeing some in our companion app.”

Il faut les contacter !

Rhoads expects in-flight connectivity to revolutionize recommendations in the future. “Where it’s going to get interesting is the airline’s ability to use connectivity to the aircraft to load content dynamically, based on passenger experience or passenger requests,” he says.

Contact : Fabienne Regitz, IFE product manager for Lufthansa.

The companion app carries passenger-viewing data from one flight to the next to inform recommendations.

Contact : Cedric Rhoads, executive director, Corporate Sales and Product Management at Panasonic Avionics.

“We are able to see how individual content files … are performing each month and then use that information to curate the next cycle’s content set” – Megan Worley, American Airlines

Apparemment Panasonic ont leur propre système appelé eX2 qui est utilisé par Emirates, Singapore Airlines, Cathay Pacific Airways et qui a remporté en 2007 un prix de la World Airline Entertainment Association, à présent appelée Airline Passenger Experience Association. C’est une surcouche de Red Hat, je crois.

“We’re very proud of our customers,” says Paul Margis, CEO of Panasonic Avionics. “Each one has a very unique system.”

Nouveau, 25 octobre 2016

Emirates has selected Thales’ AVANT in-flight entertainment (IFE) solution for its fleet of 150 Boeing 777X aircraft.

Source : Thales’ AVANT IFE Platform to Foster Innovation on Emirates Boeing 777X Fleet

Donc apparemment il y a Panasonic, Thales et Zodiac Aerospace, ce dernier est open source développé par Open Wide racheté par Smile.

Sur une autre note, un mémoire (de master je crois) d’un Portugais à Amadeus (Sophia-Antipolis) : Design and Implementation of a Flight Recommendation Engine. Mais c’est différent, c’est pas pour les in-flight entertainment.


  • Contacter Fabienne Regitz de Lufthansa, ainsi que l’auteur des articles sur pour lui dire ce qu’on fait.
  • Contacter Thalès.