---

Title: Semantic Segmentation of Satellite Imagery for Informal Settlement Growth Monitoring

Abstract:Informal expansion often outpaces cadastral updates used for service planning. We train segmentation models on high-resolution tiles with sparse polygon labels and propagate predictions across annual stacks. Detected footprint growth correlates with later connection requests for water kiosks, supporting proactive infrastructure staging.

---

---

Title: Reinforcement Learning Policies for Dynamic Pricing of Cloud Spot Instances Under Budget Constraints

Abstract:Research groups waste grants when spot prices spike during uninterruptible training jobs. We model bid placement and checkpoint scheduling as a constrained reinforcement problem using historical price traces. Learned policies reduce expected cost versus static bidding while meeting deadline probabilities set by lab administrators.

---

---

Title: Blockchain-Anchored Audit Trails for Multi-Party Clinical Trial Data Provenance

Abstract:Multi-site trials struggle to prove who altered case report forms and when disputes arise. We anchor hashed event logs to a permissioned ledger while keeping PHI off-chain. Investigators retrieve tamper-evident timelines without exposing patient identifiers in public blocks.

---

---

Title: Explainable Gradient Boosting Models for Credit Scoring With Monotonic Business Constraints

Abstract:Regulators expect credit models to respect monotonic relationships between income stability and default risk. We train constrained gradient-boosted trees with SHAP summaries validated by risk officers. Calibration and monotonicity audits pass internal policy checks while maintaining competitive Gini coefficients on out-of-time validation samples.

---

---

Title: Multi-Objective Genetic Algorithms for Container Terminal Yard Crane Scheduling

Abstract:Yard cranes compete for overlapping transfer lanes while vessels face tight departure windows. A multi-objective genetic algorithm minimizes makespan, rehandle count, and energy use on historical berth schedules. Pareto fronts reveal trade-offs where modest rehandle increases buy substantial departure risk reduction.

---

---

Title: Causal Machine Learning Estimators for Measuring Adaptive Learning Tool Effects on Exam Scores

Abstract:Universities deploy adaptive tutors without randomized trials, leaving policy makers unsure of true impact. Doubly robust estimators applied to semester logs separate tool exposure from prior ability and attendance patterns. Gains concentrate among students who receive timely mastery prompts rather than late cramming nudges.

---

---

Title: Quantized Transformer Encoders for On-Device Keyword Spotting in Noisy Factory Environments

Abstract:Hands-free machine commands fail when cloud latency or plant noise breaks recognition pipelines. We distill small transformer encoders and deploy INT8 models on ARM boards near assembly lines. Robustness to impulsive noise remains acceptable for safety-critical stop commands after domain-specific augmentation.

---

---

Title: Knowledge Graph Embeddings for Link Prediction in Heterogeneous Scholarly Citation Networks

Abstract:Recommendation engines for related literature often ignore venue, topic, and author relation types present in bibliographic graphs. We learn embeddings on a heterogeneous graph spanning papers, journals, and keywords. Ranked suggestions improve expert precision at fixed list length compared with cosine similarity on bag-of-words abstracts alone.

---

---

Title: Differentially Private Federated Averaging for Cross-Hospital Sepsis Early-Warning Models

Abstract:Pooling ICU vitals across hospitals accelerates model development but conflicts with strict data residency rules. Federated averaging with calibrated noise preserves discrimination on sepsis onset while bounding membership inference risk. Multi-site models outperform single-hospital training on rare late-stage presentations.

---

---

Title: Graph Attention Networks for Fraud Ring Detection in Mobile Money Transaction Graphs

Abstract:Fraud rings split transfers across many wallets to evade threshold rules. We construct temporal graphs linking agents, devices, and cash-out points and apply attention layers that weight suspicious neighborhood motifs. Early detection rates improve over rule engines while keeping investigator review queues manageable through calibrated risk scores.

---