User decision improvement and trust building in product recommender systems

As online stores are offering an almost unlimited shelf space, users must increasingly rely on product search and recommender systems to find their most preferred products and decide which item is the truly best one to buy. However, much research work has emphasized on developing and improving the underlying algorithms whereas many of the user issues such as preference elicitation and trust formation received little attention. In this thesis, we aim at designing and evaluating various decision technologies, with emphases on how to improve users' decision accuracy with intelligent preference elicitation and revision tools, and how to build their competence-inspired subjective constructs via trustworthy recommender interfaces. Specifically, two primary technologies are proposed: one is called example critiquing agents aimed to stimulate users to conduct tradeoff navigation and freely specify feedback criteria to example products; another termed as preference-based organization interfaces designed to take two roles: explaining to users why and how the recommendations are computed and displayed, and suggesting critique suggestions to guide users to understand existing tradeoff potentials and to make concrete decision navigations from the top candidate for better choices. To evaluate the two technologies' true performance and benefits to real-users, an evaluation framework was first established, that includes important assessment standards such as the objective/subjective accuracy-effort measures and trust-related subjective aspects (e.g., competence perceptions and behavioral intentions). Based on the evaluation framework, a series of nine experiments has been conducted and most of them were participated by real-users. Three user studies focused on the example critiquing (EC) agent, which first identified the significant impact of tradeoff process with the help of EC on users' decision accuracy improvement, and then in depth explored the advantage of multi-item strategy (for critiquing coverage) against single-item display, and higher user-control level reflected by EC in supporting users to freely compose critiquing criteria for both simple and complex tradeoffs. Another three experiments studied the preference-based organization technique. Regarding its explanation role, a carefully conducted user survey and a significant-scale quantitative evaluation both demonstrated that it can be likely to increase users' competence perception and return intention, and reduce their cognitive effort in information searching, relative to the traditional "why" explanation method in ranked list views. In addition, a retrospective simulation revealed its superior algorithm accuracy in predicting critiques and product choices that real-users intended to make, in comparison with other typical critiquing generation approaches. Motivated by the empirically findings in terms of the two technologies' respective strengths, a hybrid system has been developed with the purpose of combining them into a single application. The final three experiments evaluated its two design versions and particularly validated the hybrid system's universal effectiveness among people from different types of cultural backgrounds: oriental culture and western culture. In the end, a set of design guidelines is derived from all of the experimental results. They should be helpful for the development of a preference-based recommender system, making it capable of practically benefiting its users in improving decision accuracy, expending effort they are willing to invest, and even promoting trust in the system with resulting behavioral intentions to purchase chosen products and return to the system for repeated uses.