Thomas Sandholm

Thomas Sandholm, Sayandev Mukherjee, Bernardo A. Huberman

We present a novel approach to exploring innovation problem and solution domains using LLM fine-tuning with a custom idea database. By semantically traversing the bi-directional problem and solution tree at different temperature levels we achieve high diversity in solution edit distance while still remaining close to the original problem statement semantically. In addition to finding a variety of solutions to a given problem, this method can also be used to refine and clarify the original problem statement. As further validation of the approach, we implemented a proof-of-concept Slack bot to serve as an innovation assistant.

Thomas Sandholm, Sayandev Mukherjee, Bernardo A. Huberman

We propose and experimentally evaluate a novel secure aggregation algorithm targeted at cross-organizational federated learning applications with a fixed set of participating learners. Our solution organizes learners in a chain and encrypts all traffic to reduce the controller of the aggregation to a mere message broker. We show that our algorithm scales better and is less resource demanding than existing solutions, while being easy to implement on constrained platforms. With 36 nodes our method outperforms state-of-the-art secure aggregation by 70x, and 56x with and without failover, respectively.

Thomas Sandholm, Dongman Lee, Bjorn Tegelund et al.

We present a testbed for exploring novel smart refrigerator interactions, and identify three key adoption-limiting interaction shortcomings of state-of-the-art smart fridges: lack of 1) user experience focus, 2) low-intrusion object recognition and 2) automatic item position detection. Our testbed system addresses these limitations by a combination of sensors, software filters, architectural components and a RESTful API to track interaction events in real-time, and retrieve current state and historical data to learn patterns and recommend user actions. We evaluate the accuracy and overhead of our system in a realistic interaction flow. The accuracy was measured to 83-88% and the overhead compared to a representative state-of-the-art barcode scanner improved by 27%. We also showcase two applications built on top of our testbed, one for finding expired items and ingredients of dishes; and one to monitor your health. The pattern that these applications have in common is that they cast the interaction as an item-recommendation problem triggered when the user takes something out. Our testbed could help reveal further user-experience centric interaction patterns and new classes of applications for smart fridges that inherently, by relying on our testbed primitives, mitigate the issues with existing approaches.

Leandro Balby Marinho, Cláudio de Souza Baptista, Thomas Sandholm et al.

With the increasing popularity of location-based social media applications and devices that automatically tag generated content with locations, large repositories of collaborative geo-referenced data are appearing on-line. Efficiently extracting user preferences from these data to determine what information to recommend is challenging because of the sheer volume of data as well as the frequency of updates. Traditional recommender systems focus on the interplay between users and items, but ignore contextual parameters such as location. In this paper we take a geospatial approach to determine locational preferences and similarities between users. We propose to capture the geographic context of user preferences for items using a relational graph, through which we are able to derive many new and state-of-the-art recommendation algorithms, including combinations of them, requiring changes only in the definition of the edge weights. Furthermore, we discuss several solutions for cold-start scenarios. Finally, we conduct experiments using two real-world datasets and provide empirical evidence that many of the proposed algorithms outperform existing location-aware recommender algorithms.