Congratulations to Dr. Philipp Schneider for obtaining his PhD!

Abstract:

Modern decision systems increasingly unfold on networks of people, firms, and platforms where local events propagate as cascades and aggregate into collective outcomes. While promising, these environments are fragile: data are censored, signals are noisy, and interventions elicit strategic responses. This thesis develops methods to turn event data into dependable decisions by (i) recovering underlying dynamics,

(ii) inferring stable regularities in how actors operate, and (iii) designing interventions that remain effective under constraints.

The first part addresses learning event dynamics from coarse observations. We develop an estimator for self-exciting point processes from time-binned counts; by reconstructing intrabin event histories consistent with the conditional intensity, the estimator reduces finite-sample error and enables credible estimation under censoring.

The thesis then turns to networked platforms and the regularities governing participation. We first address content moderation via a point-process analysis that links delay to expected harm reduction, yielding practical triage rules for investigation under limited capacity. Second, using inverse reinforcement learning, we quantify behavioral homophily, showing it diverges from topical similarity to reveal distinct community roles. Third, we demonstrate that behavior-centric representations identify coordinated actors more robustly than text embeddings, particularly under noise and evasion.

Fourth, coached multi-agent LLM societies are shown to reproduce emergent ties, providing a controlled testbed for counterfactual intervention studies.

The subsequent part addresses the optimization of such interventions under strict capacity and safety constraints. We resolve a failure mode in end-to-end learning where optimization stagnates because gradients vanish at decision boundaries. We introduce Soft-Radial Projection, a layer that maps decisions to the feasible set's interior to guarantee strict feasibility while preserving the gradient flow necessary for robust learning.

The final part examines supplier risk management under buyer distress using a dual approach. Analytically, a Stackelberg game demonstrates that trade creditâ s sales benefits often outweigh non-payment risks.

Empirically, generalized method of moments estimation reveals that effective screening depends on scale: micro suppliers leverage buyer growth signals, while larger suppliers rely on operational efficiency metrics.

Collectively, these chapters provide a principled route from events to decisions in networked societiesâ recovering dynamics, characterizing behavioral regularities, enforcing constraints, and optimizing interventionsâ to enable safe, reliable, constraint-aware decision making.