Title: Human-Supported Recommender Systems - An Experimental Evaluation
Abstract: During the past decades, recommender systems have become a valuable tool for operators of e-commerce and information systems to reduce consumer choice overload. Traditional recommender techniques succeed in producing useful recommendations in many domains. However, recent research is investigating new recommender system designs for domains with complex constraints where traditional techniques fail to provide suitable recommendations. Those designs often employ complex models and depend upon an extensive data basis. In many cases such systems turn out to be impracticable due to scalability and usability concerns. The broad availability of social network information allows for new recommender system designs, where user networks are utilized to let friends recommend products to each other. By computationally supporting peers in social networks to apply their knowledge and cognitive capabilities, we can make use of human intelligence whose computational reproduction would be unfeasible. Based on these new possibilities, we present the paradigm of human-supported recommender systems, where the computational power of computers is combined with the abilities of humans to remedy shortcomings of previous systems. To evaluate this new paradigm, we compare a general purpose human-supported recommender system with a traditional system in the domain of course recommendations. Our experiment demonstrates that by applying this design, overall user satisfaction can be significantly improved. In addition, we gather experimental evidence that user satisfaction can not only be increased due to a gain in recommendation attractiveness, but also due to behavioral factors such as social influence and trust in the system. These findings demonstrate the potential of human-supported systems to generate recommendations for domains hitherto hard to administer, thus legitimizing further research in this field.
