Recommendation systems gather and filter information in order to provide users with products they might like. They use the user’s history of interactions with the system in order to recognize one’s preferences. The overall quality of recommendations usually improves with frequent system usage [1].
Web services, such as Netflix [2] and IMDb [3] both use recommendation systems for movies. They track user’s interactions and gather information about movies which one watched, liked or disliked. Later, they compare this information with other data in their database and prepare predictions which movies a user might prefer. The user’s rejection or approval of the recommended movie represents another piece of information on that user, which can improve the overall quality of recommendations.
Recommendation systems often face problems of high similarity between recommended items. These systems recommend similar items due to the lack of data on users or products (cold start problem). The same issue also occurs due to the aim of recommendation systems to make accurate predictions. Very similar recommended products (small diversity) usually do not present any real value to the end user.
The problem of low diversity between recommended products originates from the evaluation metrics of recommendation systems, where usually only the difference between the predicted rating for an item and the user’s real rating for it is measured. The smaller the difference, the better is the recommendation accuracy. For example, with the knowledge of user’s holiday preferences we can recommend 5 different holiday packages at the same location, as the prediction was that the user would like this location. These recommendations might be very accurate as they are based on user’s preferences and history, but they are probably useless for the user when choosing the holiday destination. Instead, we should recommend him 5 packages with different locations and give an option to choose the preferred location and request more details [4].
In this work, we propose a recommendation system which would use a combination of different recommendation algorithms and use the proposed evaluation metric in order to evade the diversity problem. Also, during an experiment we conducted, we evaluated the correlation between diversity in recommended items and user’s satisfaction. For this experiment, we created a conversational recommendation system with controlled diversity, which was based on a recommendation system of the User-adapted Communication and Ambient Intelligence Laboratory of the Faculty of Electrical Engineering at the University of Ljubljana. In the experiment, we showed users 5 sets of recommended items with controlled diversity and asked them for feedback, how they are satisfied with the recommended items. 38 users participated in the experiment and provided us with 190 ratings of diversity and quality for recommended sets of items. We analyzed this data statistically and the results are statistically significant at the risk of 5 percent. The result is that users prefer recommended items with increased diversity. However, we recommend repeating the experiment on a bigger group of participants (larger sample) and on different recommendation systems.
|
