Elicitation of Users’ Implicit Feedback for Building Recommender Systems in Complex Decision Environments (Li Chen et al.)

In the current online environment, a decision maker often encounters the challenge of how to overcome the overwhelming data for making an efficient and accurate decision, especially in the complex application domains such as e-commerce, financial investment, resource scheduling, trip planning, online learning, and so on. It has been widely recognized in the area of psychology that accurate decisions should be made through compensatory strategies, such as weighted additive sum rule (WADD), by which users need to carefully consider all relevant alternatives and make tradeoffs among their attributes. However, because the cognitive overload and/or emotional cost that this type of strategies induces, people often tend to carry out inaccurate, non-compensatory strategies for the benefit of saving efforts.

It hence raises a critical research problem of how to support users to make accurate decisions, while keeping the consumed efforts within their acceptable level. Personalized decision support, such as the preference-based recommender system, has been demonstrated as an effective approach to achieving this objective. However, there remain two major issues that have not been well resolved in the existing systems. 1) The required user effort is unavoidably still high as they need to provide explicit feedback to the recommendation for the preference refinement. 2) It is still unclear whether and how users make compensatory decision strategy, especially when they examine the item recommendations, which may limit the system's competency in correctly establishing their preference model and generating more satisfactory recommendation.

The novelty of this project is that we will exert to exploit users' implicit behavior, by means of combining their clicking actions and eye movements, for eliciting their preferences. To be more specific, we attempt to develop a novel critiquing-based recommender system based on implicitly elicited feedback, so as to reduce users' involvement efforts while still allowing them to obtain high decision accuracy. We expect that our work will contribute to the field of complex decision support from the following aspects. 1) The analysis of users' implicit behavior may enable us to in depth understand in which condition, and how, users perform compensatory decision strategy. 2) The implicit preference elicitation could be conducted at both product and attribute levels for potentially more accurately building users' multi-attribute preference model. Particularly, the attribute-level elicitation may help verify the preferential dependency relationship between attributes. 3) The uncertainty of elicited user preferences could further be addressed by suggesting the attributes tradeoff relations for users to consider and choose. 4) The proposed methods will be evaluated in proof-of-concept system (with both desktop and mobile versions), through experimental simulations and controlled user studies.

For further information on this research topic, please contact Dr. Li Chen.