Large-Scale Machine Learning
for Big Data Analytics

About

---

Jiayu Zhou is an Assistant Professor in the Department of Computer Science and Engineering at Michigan State University. His research spans large-scale machine learning, data mining, and biomedical informatics. Before joining MSU, he was a staff research scientist at Samsung Research America. He received his Ph.D. in computer science from Arizona State University in 2014.

He is affiliated with the ILLIDAN Lab and serves as Principal Investigator on NSF CRII projects focused on multi-task learning and interactive knowledge integration. His work develops algorithms and open-source software that enable machines to learn from related tasks simultaneously, with applications in disease progression modeling and biomedical discovery.

Large-scale machine learning tackles the challenges of extracting meaningful patterns from datasets that are too vast or complex for conventional approaches. The field demands algorithms that scale gracefully across distributed systems while maintaining statistical rigor. Researchers develop methods that balance computational efficiency with model accuracy, enabling insights that would otherwise remain hidden in the sheer volume of information generated by modern scientific and industrial applications.

Multi-task learning represents a paradigm where models simultaneously learn from several related tasks, sharing representations to improve generalization across each individual problem. By leveraging commonalities among tasks, these approaches reduce the risk of overfitting and often yield more robust predictions than training separate models in isolation. This framework proves especially valuable when labeled data for any single task is scarce but related auxiliary tasks offer abundant training signal.

Biomedical informatics applies computational techniques to complex healthcare data, opening new avenues for understanding disease progression and patient outcomes. Electronic health records, genetic profiles, and medical imaging all contain latent patterns that machine learning can surface. The challenge lies in handling heterogeneous, longitudinal, and often incomplete data while producing models that clinicians can trust and interpret in high-stakes decision-making environments.

Sparse modeling focuses on identifying the most relevant features or parameters from high-dimensional data, discarding noise and redundancy. By enforcing sparsity constraints during optimization, these techniques produce parsimonious models that are easier to interpret and computationally lighter to deploy. Applications range from biomarker discovery in genomics to efficient signal processing, where selecting only the essential variables leads to clearer scientific understanding and practical utility.

Temporal modeling addresses data that unfolds over time, capturing dependencies and trends that static analyses miss. Whether tracking disease progression in patient cohorts or monitoring evolving social networks, these methods must account for irregular sampling intervals and complex longitudinal structures. The resulting models reveal how phenomena develop, deteriorate, or respond to interventions, providing a dynamic lens on processes that shape real-world outcomes.

Knowledge discovery in databases encompasses the end-to-end process of transforming raw data into actionable insights. It bridges data mining algorithms, statistical validation, and domain expertise to ensure that discovered patterns are both statistically significant and practically meaningful. This interdisciplinary pursuit underpins advances in scientific research, business intelligence, and public policy, where rigorous methodology separates genuine signal from spurious correlation.

Factorized models decompose complex relationships into lower-dimensional representations, making them indispensable for recommendation systems and network analysis. By approximating large matrices or tensors with compact latent factors, these techniques uncover hidden affinities among users, items, or entities. The resulting embeddings capture subtle semantic structure while dramatically reducing storage and computation requirements, enabling scalable personalization across massive online platforms.

Interactive machine learning places human judgment at the center of the modeling loop, allowing domain experts to guide and refine algorithms in real time. Rather than treating learning as a fully automated process, these systems invite feedback on intermediate results, feature relevance, and model behavior. This collaboration between human intuition and computational power accelerates convergence toward useful solutions, particularly in exploratory analysis where the right questions evolve alongside the answers.