Summary: Challenges of Serendipity in Recommender Systems

Full text available: proceedings | direct link.

D. Kotkov, J. Veijalainen, and S. Wang. Challenges of serendipity in recommender systems. In Proceedings of the 12th International Conference on Web Information Systems and Technologies, pages 251-256, 2016.

Challenges of Serendipity in Recommender Systems

The paper does not contain any experimental results. In this paper, we indicate some challenges of suggesting serendipitous items to users. The goal of the paper is to guide and inspire future efforts on serendipity in recommender systems.

Definitions

Recommender systems are software tools that suggest items of use to users, where an item is a general term that is used to denote an object that a recommender system suggests. An item can refer to anything, a movie, a song or a book. An example of a recommender system is YouTube. The service suggests videos to a user based on the videos the user watched.

Serendipity is:

• fortunate happenstance
• pleasant surprise
• the faculty of making fortunate discoveries by accident
• luck that takes the form of finding valuable or pleasant things that are not looked for

Content

The problem is to suggest items that would correspond to the provided definitions. However, the definitions such as a “pleasantly surprising item” or “fortunate item” are very vague. It is not clear what items we should call serendipitous. Currently, there is no agreement on the definition. Different researchers call different items serendipitous. This is the main problem why we cannot suggest serendipitous items.

To design a serendipity-oriented recommendation algorithm, we need to know what we want to suggest, because we cannot suggest if we do know what we want to suggest. The next step is to measure serendipity, because we cannot improve what we cannot measure. These are the main problems with serendipity in recommender system and everything else is affected by these problems.

Why do not we know what items are serendipitous? Due to three reasons: no agreement on definition, lack of data and emotional dimension. The first point has already been discussed.

• Lack of data. Serendipitous encounters are very rare (which makes them valuable) . We therefore do not have enough observations to draw reliable conclusions regarding these items.
• Emotional dimension. It significantly depends on user mood, whether the user finds an item serendipitous (which is included in the definitions). However, we often lack this kind of data.

As there is no agreement of definition, there is no agreement on a way to measure serendipity (it would be weird, if researcher did not have agreement on what items are serendipitous, but completely agreed on a way to measure serendipity). There were many serendipity metrics proposed and there is no agreement on which one to use. Each metric has its own pros and cons.

In a nutshell

Suggesting serendipitous items is challenging, because:

• No definition
• No observations
• No data on emotions
• No way to measure