Context-Aware Recommender Systems Evaluation

As the amount of data provided by various software systems increases, there is a need to offer a filtered set of items personalized to user's needs. To enhance user's comfort and thus to satisfy him, we call for recommender system. Recommender systems suggest a set of items that a user might be interested in or might find them useful. Basically, accomplishing recommendation task consists of two steps. At first recommender system has to collect information about user's activities and construct user model, which represents his preferences. Second step is to apply algorithm to compute over user's model and generate a set of items to suggest. However, we may improve the quality of the set of suggested items by including information describing user's environment or state in the user model. These information's value in recommender system may vary depending on the domain of proposed recommender and type of the information (e.g. user's wealth may not be relevant while recommending a movie to watch, however it is relevant while suggesting an item to buy in an e-commerce system). Including context in the process of recommendation matters, because there is a correlation between user's behaviour in certain situations and contexts as Riboni et al. proved in [2]. The question now arises is how to evaluate the quality of suggestions provided by context-aware recommender system? As we mentioned before, context-aware recommenders consider various context information and include them in the user models. On one hand this allows recommenders to generate highly specific suggestions. On the other hand it makes the evaluation of the recommendations slightly more complicated. The simplest thing to do, in the process of context-aware recommender evaluation, is to evaluate every recommendation only if all environment conditions given by recommendation are matched with real contexts. This approach is precise, however may be, and usually is, very costly (e.g. In the middle of the summer we want to evaluate recommendations suggesting user's actions in snowy weather.).