Advances in computer-human interaction for recommender systems (AdCHIReS)

Recommender Systems produce suggestions of items or content that users have not considered yet, but that might be interesting for them. Such recommendations are produced by analyzing what they previously consumed (bought, watched, or listened), or by the identification of similarities with other users. Users usually express their preferences by providing explicit ratings; however, these might be not very informative, since users might only indicate whether they liked something or not, thus employing only the ratings at the end of the scale (this is also one of the reasons why Netflix employed a thumb-up/thumb-down rating system). However, between these two extremes there lies a set of different actions in the interface that might be interpreted as feedback, but that needs to be collected implicitly. Even if the literature provides different techniques for collecting implicit feedback, they are usually tailored to specific types of applications.

From the user’s point of view, Recommender Systems remain a black box that suggests content, but the users hardly understand why some items are included in the list. The relevance of this issue has increased in recent years, as the introduction of approaches based on latent features (such as Matrix Factorization or Deep Learning) has made it very hard to connect user preferences with the recommended items. Providing the users with an understandable representation of how the system represents them and allowing them to control the recommendation process would lead to benefits in how the recommendations are perceived and in the capability of the system to be persuasive. Such transparency is one of the multiple (and usually conflicting) requirements of Recommender Systems.

Beyond the classical engineering of Recommender Systems, based on data processing, filtering, and sorting, the engineering aspects should also cover aspects related to how users interact with these systems, including how to input data, how to define and evolve the user model, how to present information to users, and how these can manipulate such information. These engineering processes might benefit from practice in specific areas, such as web configurators (which guide the users in the inspection of possible product variants) and safety critical interactive systems (where predictability and consistency over executions are prerequisite to certification). In order to deploy Recommender Systems in broader contexts, there is a need for structured and systematic approaches to engineer such complex computing systems.