Ä¢¹½Ö±²¥

YouTube recommends videos to watch. Spotify suggests songs to listen to. These services need to provide recommendations, because they have so many options that users often simply do not know what the services can offer.

To generate recommendations, these services employ recommender systems. These are systems that suggest items to users, where items can be any objects, such as movies, songs or books. Recommender systems receive the history of interactions of users with items on the service and generate personalized recommendations of items for each user.

Although recommender systems can accurately predict what items users will like, users often complain that recommendations are boring or too conservative. We studied how to make recommendations exciting for users by suggesting serendipitous items.

Serendipitous items are unfamiliar and unexpected to users when recommended, but liked after the users consumed them. For example, let’s say YouTube recommends you a video that is new to you and that you do not expect to like (for example, because you do not normally like these kinds of videos). If you watch and like this video, we will regard it serendipitous to you. There are different variations of serendipity, since you can find a video novel and unexpected in different ways. For example, you might not expect YouTube to recommend you a video, because it does not usually suggest you these kinds of videos. Furthermore, since users have different tastes, videos you find serendipitous might not be serendipitous to others.

We found that different variations of serendipity have different effects on users, but generally broaden user preferences and make recommendations exciting. For example, if YouTube recommended you a serendipitous video and you watched it, this video would be more likely to make you interested in a wider selection of videos compared to a non-serendipitous video you enjoyed watching.

We also found that serendipitous items are rather rare. The typical movie recommender system MovieLens () contains up to 8.5 % of movies that are serendipitous to users in one way or another. This upper boundary estimation suggests that serendipitous items are difficult to recommend, but it is still worth of trying to recommend them in many cases.

In this dissertation, we designed a way to measure serendipity and an algorithm that suggests serendipitous items. Our results might be useful for websites with overwhelming amount of content, such as YouTube, Netflix, Spotify or Amazon. They might use results of our research to improve their recommendations and make their users more satisfied.

M.Sc. Denis Kotkov defends his doctoral dissertation in Computer Science "Serendipity in Recommender Systems" in building Agora, Alfa hall, on June 7, 2018 at 12 o’clock noon. Opponent Dr. Konstantinos Stefanidis (University of Tampere) and Custos Professor Jari Veijalainen (Ä¢¹½Ö±²¥). The doctoral dissertation is held in English.

More information

Denis Kotkov, deigkotk@student.jyu.fi, tel. +358 40 805 3577
Communications Officer Kati Valpe, kati.valpe@jyu.fi, tel. +358 400 247 458

The research conducted in this dissertation was funded by the Academy of Finland, grant #268078 (MineSocMed), the KAUTE Foundation, and the Ä¢¹½Ö±²¥.

The dissertation is published in the series Jyväskylä Studies in Computing number 281, 177 p., Jyväskylä 2018, ISSN: 1456-5390, ISBN: 978-951-39-7437-4, ISBN: 978-951-39-7438-1 (PDF).

Denis Kotkov was born on January 1, 1990 in Saint Petersburg, Russia. He graduated from the Saint Petersburg National Research University of Information Technologies, Mechanics and Optics (ITMO University) (Russia) obtaining a specialist’s degree in Information Systems and Technologies in 2012, and a Master’s degree in Information Technologies and Services in Telecommunications in 2014. He has also worked as a software developer for Devexperts (Russia) for three and a half years (2010-2014), developing software to trade on the stock market. He has been a doctoral student at the Ä¢¹½Ö±²¥ since August 2014.