Decision Focused Clustering

Modern digital platforms make large volumes of operational decisions repeatedly and in real time. While rich contextual data are available to support these decisions, treating every decision instance as fully personalized is often computationally expensive and operationally unstable. In practice, platforms frequently rely on aggregation and reuse of decision rules across similar situations, with clustering serving as a key mechanism for reducing decision complexity. However, traditional clustering methods are designed to group instances based on geometric or statistical similarity in feature space, without regard to how the resulting clusters affect downstream decision quality. This thesis proposes a decision-focused clustering framework that explicitly aligns the clustering process with downstream optimization objectives. Building on the Smart Predict- then-Optimize (SPO) framework, we replace conventional distance-based clustering criteria with a decision-aware loss that measures the impact of clustering choices on operational decisions. We study how to partition data such that instances within each cluster can share a common decision with minimal loss in optimality under cluster-level approximation. Each cluster is associated with a representative parameter that yields a shared decision via the underlying optimization problem. We apply the proposed framework to large-scale delivery data and show that decision-focused clustering produces more stable and effective decisions than traditional clustering baselines in high-dimensional and partially observable environments.