Séminaire de Paolo Dragone (Criteo) : "Constructive Preference Elicitation"
Speaker(s) : Paolo DRAGONEConstructive Preference Elicitation (Paolo Dragone)
Product configuration is a quite common yet remarkably difficult decision task. It arises in many situations, from choosing components of a personal computer to planning an on-the-road trip.
These problems consist in choosing one among a combinatorial amount of possibilities, taking into account many decision variables and constraints. Configuration systems are the de-facto standard for handling this kind of tasks. They, however, are usually not designed to optimize a well understood utility model, such as MAUT, which makes them less desirable over all. Standard preference elicitation techniques, like regret-based and Bayesian, are better suited for this problem, yet they have some shortcomings that are accentuated when dealing with large combinatorial domains, namely robustness to noise, explicit handling of hard constraints, and scalability. To mitigate these issues, we propose constructive preference elicitation, a technique relying on online machine learning for estimating a utility model from user interactions, and on combinatorial optimization to infer an optimal recommendation for the user. More specifically, I will present a method based on “coactive” learning, in which a user provides feedback in the form of an “improvement” over the recommended object. I will show how this technique has been successfully applied to a product bundling task and validated with real-life participants. I will also talk about some extensions of coactive learning needed to deal with larger constructive problems. Finally, I will briefly touch on few more interesting extensions in this area that we currently have in our pipeline.
-------- Speaker --------
Paolo Dragone is a machine learning engineer at Criteo, working on applied research projects in the areas of recommendation, bidding, and brand safety. He previously obtained a PhD in Computer Science from the University of Trento where he graduated in 2019 under the supervision of Prof. Andrea Passerini. He has published several papers at leading artificial intelligence venues such as AAAI, IJCAI and RecSys.