MTS Kion Implicit Contextualised Sequential Dataset for Movie Recommendation

Dataset Description

This is an official repository of the Kion Movies Recommendation Dataset. The data was gathered from the users of MTS Kion video streaming platform from 13.03.2021 to 22.08.2022. It is sample of anonymous data with adding random noise.

The public part of the dataset includes 5,476,251 interactions of 962,179 users with 15,706 items. The dataset includes user_item interactions and their characteristics (such as temporal information, watch duration and watch percentage), as well as user demographics and rich movies meta-information.

Private part of the dataset contains movies that the users watched within one week following the period covered by the public dataset. It is not released to general public, however there is a public the sandbox, where the researchers can measure MAP@10 metric on the private part of the data. Sandbox is accessible by the address https://ods.ai/competitions/competition-recsys-21/leaderboard/public_sandbox.

To make a submission, you need to use the sample_submission.csv file, and replace the sample item ids with the ids of the recommended items according to your recommendation model.

The dataset consists of three parts:

1. Interactions.csv - contains user-item implicit interactions, watch percentages, watch durations
2. Users.cvs - contains users demographics information (sex, age band, income level band, kids flag)
3. Items.cvs - contains items meta-information (title, original title, year, genres, keywords, descriptions, countries, studios, actors, directors)

The users and items files have two versions:

• data_original - original meta-information in Russian language
• data_en - english version of the metadata translated with Facebook FAIR’s WMT19 Ru->En machine translation model.

Comparison with MovieLens-25M and Netflix datasets

Quantitative comparison:

	Netflix	Movielens-25M	Kion
Users	480,189	162,541	962,179
Items	17,770	59,047	15,706
Interactions	100,480,507	25,000,095	5,476,251
Avg. Sequence Length	209.25	153.80	5.69
Sparsity	98.82%	99.73%	99.9%

Qualitative comparison:

Dataset Name	Netflix.	Movielens-25M	Kion
Type	Explicit (Ratings)	Explicit (Rating)	Implicit (Interactions)
Interaction registration time.	After watching	After watching	At watching
Interaction features	Date, Rating	Date, Rating	Date, Duration, Watched Percent
User features	None	None	Age, Income, Kids
Item features	Release Year, Title	Release Year, Title, Genres, Tags	Content Type, Title, Original Title, Release Year, Genres, Countries, For Kids, Age Rating, Studios, Directors, Actors, Description, Keyword

Kion challenge

This dataset was used for the Kion challenge recommendation contest (Official website in Russian Language).

This table contains results of the winners of the competition, measured on the private part of the dataset:

Position	Name	MAP@10	Solution Type
1	Oleg Lashinin	0.1221	Neural and Non-Neural ensemble
2	Aleksandr Petrov	0.1214	Neural and Non-Neural ensemble
3	Stepan Zimin	0.1213	Non-Neural ensemble
4	Daria Tikhonovich	0.1148	Gradient Boosting Trees
5	Olga	0.1135	Gradient Boosting Trees
Popularity baseline		0.0910

Acknowledgements

Igor Belkov, Irina Elisova.

We would would like to acknowledge Kion challenge participants Oleg Lashinin, Stepan Zimin, and Olga for providing descriptions of their Kion Challenge solutions, MTS Holding for providing the Kion dataset, ODS.ai international platform for hosting the competition.

Citations

If you use this dataset in your research, please cite our work:

@article{petrov2022mts,
  title={MTS Kion Implicit Contextualised Sequential Dataset for Movie Recommendation},
  author={Aleksandr Petrov,  Ildar Safilo, Daria Tikhonovich and Dmitry Ignatov},
  year={2022},
  booktitle={Proceedings of the ACM RecSys CARS Workshop 2022, September 23d, 2022 Seattle, WA, USA }
}