Looking for what you are not looking for
Recommendations must be unfamiliar and engaging, sometimes called "serendipity" in the academic literature.
Unfamiliar
- Recommendations must disrupt the flow of similarity, the logic of filtering for nearness or resemblance to past choices. They must be discontinuous with established interests.
- Crude approximation is random offerings.
- Antagonistic filtering begins by identifying users who are significantly different. Then, among those diverse users, a strand of similarity is located. Perhaps a single movie or two that they both strongly enjoy. From there, interests from one and the other can be recommended back and forth. The idea is that the recommendations will be new and different given that they emerge from very different types of people. But, they will not be simply random because some overlap exists between the two.
Engaging
- Opposites (the right kind) attract. Locating these opposites is the task of antagonistic filtering.
- Explainability. Research indicates that explained recommendations may increase user interest and curiosity.
- Context. For example, a user planning a trip to Paris may be predisposed to unfamiliar offerings related to Paris (movies, songs, restaurant, clothes...).