Human-in-the-loop Exploration of Composite Items

Human-in-the-loop data exploration is seeing a renewed interest in our community. With the rise of big data analytics, this area is growing to encompass not only approaches and algorithms to find the next best data items to explore but also interactivity, i.e. accounting for feedback from the data scientist during the exploration. Interactivity is essential to account for evolving needs during the exploration and also customize the discovery process. In this tutorial, we focus on exploration of Composite Items (CIs) that require repeated interaction with human users. CIs address complex information needs and are prevalent in online shopping where products are bundled together to provide discounts, in travel itinerary recommendation where points of interest in a city are combined into a single travel package, and task assignment in crowdsourcing where persoalized micro-tasks are composed and recommended to workers. CI formation is usually expressed as a constrained optimization problem. For instance, in online shopping, package retrieval can retrieve the cheapest smartphones (optimization objective) with compatible accessories (constraints). Similarly, a city tour must be the most popular and conform to a total time and cost budget. A data scientist interested in exploring a variety of CIs has to repeatedly reformulate optimization problems with new constraints and objectives. In this tutorial, we investigate the applicability of interactive data exploration approaches to CI formation. The tutorial will have the following parts: 15 minutes We will first review CI applications and shapes (15mn) that are applicable in different domains. This part will gather different examples and attempt to unify them. 60 minutes We then discuss three big research questions : (i) existing algorithms for CI formation, (ii) human-in-the-loop CIs, and (iii) optimization opportunities. 15 minutes We conclude with ongoing and future research directions. The tutorial targets theoreticians and practitioners interested in the development of data science applications. It should be of particular interest to database researchers, applied machine learners, as well as data scientists in industrial research settings who want to learn about how different domains, such as product recommendation, scientific simulation, or team formation in the social sciences and crowdsourcing, have been developing their "siloed" definitions of CIs. The research direction presented in the tutorial will be helpful to converge these domain specific ideas and creating an overarching generic framework. Tutorial attendees are expected to have basic knowledge in algorithms and data management. Knowledge in constrained optimization is not necessary. The proposed tutorial is timely. It brings together several related efforts and addresses unsolved questions in the emerging area of human-in-the-loop exploration of complex information needs. The tutorial is relevant to the general area of data science and more specifically to Scalable Analytics, Data Mining, Clustering and Knowledge Discovery, Indexing, Query Processing and Optimization, and Crowdsourcing. The technical topics covered are constrained optimization,ranking semantics, clustering, algorithms, and empirical evaluations.