AAAI Publications, The Thirty-First International Flairs Conference

Font Size: 
Between Multi-Attribute Utility Decision Making and Recommender Systems: Transparent, Instantaneous, Local Recommendations for Sparse Data
James Schaffer, James Michaelis, Adrienne Raglin, Stephen Russell

Last modified: 2018-05-10


One of the most significant contributions to decision technology is multi-attribute utility (MAU) theory. MAU has gained increased traction in determining the value of information in tactical networking, has been a inspiration for some content-based recommender systems, and artifacts of MAU can be found on nearly every e-commerce website. While recommender systems attempt to create a model of the user (often on latent variables) from rating data, MAU attempts to solicit content-relevant attribute weightings explicitly. Both of these methods have trade-offs which might be mitigated if they could be combined. This research presents a method that we call MAUSVR for fusing recommender and MAU decision technology by automatically learning MAU models (from a user's ratings. A comparison with collaborative filtering techniques on the MovieLens dataset suggests that MAUSVR achieves better ranking quality under sparse conditions while also gaining in transparency and locality. Additionally, MAUSVR was able to be built instantaneously (< 100ms) for more than 75% of the evaluated users with an off-the shelf Java implementation of SMOreg. These findings indicate promise for the use of MAUSVR in real-time decision support systems operating in sparse data conditions.


recommender systems; multi-attribute utility; C4ISR; tactical networking; decision-making

Full Text: PDF