Recommender systems that assist in decision-making are increasingly used in a number of information settings. Most research in this area focuses on algorithms that predict how much a user will prefer a given item based on their own and others' past preferences. Such research tends to treat the question as a machine learning, statistical, or optimization problem; however, this abstracts away important aspects of making recommender systems useful to people, including attending to the decision-making processes and contexts being supported. This project will develop recommender systems that explicitly attend to these concerns using a smart home as the driving domain. The methods developed for reasoning in this context can inform similar decision problems in other domains, such as managing supplies and logistics and accessing information.
This project will develop solutions to a number of important but under-addressed challenges for recommender systems, including (1) recommending groups of related items that work well together, rather than individual items; (2) recommending and re-recommending items primarily from an existing set of items rather than focusing on single-use consumption of new items; and (3) considering important factors about both the physical and social environment that affect decision making. The project will explore these issues in the context of ensemble selection decisions in a smart home. The system will be grounded in case-based reasoning; this will mitigate the cold start problem that plagues ratings-based collaborative filtering algorithms and fits well with the planned attribute-based representations of items, preferences, and context. To do this, the team will use 3D scanners and rendering software to create personalized models of items, and both experts and crowdworkers will then assess the model outputs, accounting for social and physical contexts. This will generate a case library that, along with rulesets for combining items, can be used to generate recommendations for both sets of items and individual items that fit well with already-chosen ones. The system will be based on an existing prototype, adapted to run on a tablet to support remote data collection. The deployment includes a setup process where the team will attach RFID tags to items and photograph them for later analysis and addition of attributes; a three-month period in which users' choices and context will be logged but no recommendations are made, and a six-month period where recommendations are offered. The system will be evaluated in terms of the usefulness of the recommendations, their ability to suggest novel and liked combinations, and their effect on the use and recombination of users? existing items. Through deploying prototype systems in homes, the work will also give insight into designing smart home applications such as activity and medication monitoring that involve object tracking.
