The common thread is data that arrives fragmented — duplicated, inconsistently described, split across sources and views — and the question of how much structure can honestly be recovered from it. Each area below states what is being asked, how it is approached, what exists publicly, and where the current limits are.
Matching and deduplicating records across noisy, heterogeneous sources, with pipelines whose every stage can be inspected, compared, and audited.
Methods
- Multi-field similarity combining string, numeric, and embedding-based metrics
- End-to-end pipelines from ingestion to evaluation (erbot toolkit)
- Automated parameter tuning with recorded configurations
Public artifacts
Limitations. No benchmark figures are published in isolation; results are only meaningful with their dataset and evaluation setting.
Status: Active
Blocking and candidate-generation strategies that control the quadratic comparison budget while keeping recall losses measurable rather than invisible.
Methods
- Blocking utilities integrated into the erbot pipeline
- Recall-aware analysis of candidate generation choices
Public artifacts
Limitations. Blocking trade-offs are dataset-dependent; conclusions are reported per setting, not as universal rules.
Status: Active
Multiview Representation and Fusion
How should multiple views of the same records — lexical, numeric, embedding, tensor — be combined without letting one view dominate or contradict the rest?
Representing records through several complementary views and fusing them, including tensor-based canonical correlation approaches for aligning views.
Methods
- TF–IDF with singular value decomposition alongside embedding features
- Tensor-based multiple canonical correlation analysis (tensorMCCA)
- Contributions to R packages for multiview analysis (GCMER)
Public artifacts
Limitations. View-fusion behavior under highly imbalanced view quality remains an open question in this work.
Status: Active
Turning similarity structure into entity clusters, comparing K-means, Louvain community detection, hierarchical clustering, and partitioning around medoids on the same footing.
Methods
- Similarity-graph construction from multi-field comparisons
- Side-by-side clustering strategies within one pipeline (K-means, Louvain, hierarchical, PAM)
Public artifacts
Limitations. Cluster-quality comparisons depend on ground-truth availability, which many real datasets lack.
Status: Active
Systematic threshold selection for match decisions: optimization curves, sensitivity to data shifts, and recording the configuration behind every operating point.
Methods
- Automated parameter tuning with optimization curves in evaluation reports
- Sensitivity analysis of operating points across settings
Public artifacts
Limitations. Calibrated thresholds transfer poorly across domains; per-domain recalibration is assumed, not avoided.
Status: Active
Making evaluation a first-class artifact: ground-truth comparison, reports that bind metrics to the exact parameters that produced them, and pipelines designed to be rerun.
Methods
- Evaluation against ground truth built into the pipeline
- Generated PDF reports carrying parameters, optimization curves, and metrics together
Public artifacts
Limitations. Reproducibility guarantees stop at the data boundary: private or shifting source data limits what any pipeline can promise.
Status: Active
Computational Science
What does it take to move between economics, statistics, and computation without losing the standards of any of them?
Doctoral work in computational science at the University of Massachusetts Boston, building on graduate training in statistics and economics, with earlier published work in information systems research.
Methods
- Statistical modeling and numerical methods
- Machine learning applied to data-integration problems
- Systematic literature review methodology
Public artifacts
Limitations. Interdisciplinary breadth is a bet: it trades depth-per-field for connections between fields, and the copy here claims only the work actually done.
Status: Ongoing