Antonis Mandamadiotis

Abstract

Modern data repositories, such as data lakes and relational databases, have grown too large and heterogeneous for users to explore manually. Data exploration, the process of discovering relevant information and extracting insights from such repositories, is therefore fundamental yet increasingly time-consuming: as data volume and diversity grow, guiding users toward the most informative results becomes progressively harder.

This thesis envisions data exploration as a recommendation problem, in which an intelligent system surfaces the top-K recommendations that best match a user's information need. Rather than requiring users to manually search through overwhelming candidate spaces, the system instead selects a small, high-quality set of recommendations.

We develop this vision across three complementary exploration settings: recommending dataset bundles from data lakes, recommending SQL query components under user feedback, and recommending subgroups with heterogeneous treatment effects. Despite their differing objects of recommendation, each setting requires jointly balancing multiple task-specific objectives under combinatorially large search spaces, where exhaustive solutions are computationally infeasible. To address this, we formalize each recommendation problem and design efficient heuristic algorithms that achieve near-optimal recommendation quality at interactive speeds.