Publication Type
PhD Dissertation
Version
publishedVersion
Publication Date
7-2026
Abstract
Peer analysis, the comparison of an entity against a relevant group of similar en- tities, underpins benchmarking, due diligence, and risk monitoring across business and regulated domains. Yet it rarely fails at the modeling stage. It fails earlier, when the entities, attributes, and values that analysts intend to compare are not actually comparable. Organizations operate with incomplete information, heterogeneous systems, changing business contexts, and strict limits on data sharing, while the analytical machinery for peer comparison assumes that clean, aligned, and share- able data already exist. This thesis argues that this assumption must be earned, and that earning it is a data-preparation problem with a recurring structure: the roll-up/drill-down duality. Comparability is built by rolling noisy, low-level evidence up into stable business concepts, and it is validated by drilling back down to the evidence whenever a decision demands precision.
The thesis applies this single duality at three layers of the peer-analysis life cycle and contributes a method at each. First, to determine who to compare, NCEXPLORER treats roll-up and drill-down as first-class operators over a knowledge graph’s instance and concept spaces, letting an analyst move from a concrete en- tity to an abstract peer category and back to the supporting evidence; a concept– document relevance score, an unbiased random-walk estimator with a k-hop reach- ability index, and a coverage–specificity–diversity subtopic ranker make the nav- igation tractable and explainable, with gains demonstrated over strong neural and structure-based baselines. Second, to determine what attributes can be compared once a peer set is known, LLMATCH decomposes schema matching into Schema Preparation, Table Selection, and Column Matching, and adds a roll-up/drill-down strategy that abstracts semantically related columns before coarse alignment and refines the result into column-level mappings; the co-developed SCHEMANET benchmark, with multi-table schemas an order of magnitude larger than prior benchmarks, supports realistic n:m alignment, and the study shows that vector similarity, the default retrieval step in many LLM pipelines, is not the most effective choice for selecting candidate tables. Third, to enable peer analysis when raw values cannot be shared at all, the COMB framework separates scalable global generation from targeted local repair: a cheap generator drafts the whole table, a segment tree localizes where the draft is weak, and an acceptance gate admits a per-segment repair only when it improves the segment and leaves every coarser roll-up no worse than before; an iterative LLM repair generator is steered entirely by aggregate statistics and a fixed synthetic texture sample, so it never sees a real record. Across five public datasets the framework improves segment-level fidelity without regressing global quality and matches or beats generators that condition on real rows, at a structural guarantee of zero real-record exposure to the externally hosted LLM.
Taken together, the three works instantiate a single idea: the same roll-up/drill-down duality that lets an analyst navigate a knowledge graph also governs how schema elements are abstracted and how synthetic records are validated against their aggregates. Each layer rolls up to reduce complexity and drills down to restore precision, and each keeps the evidence for a comparison auditable rather than hidden inside an opaque model. The contribution of the thesis is therefore not only three methods for three preparation problems, but the demonstration that a structured curation framework can turn noisy, changing, and restricted business data into a foundation on which peer analysis becomes meaningful.
Keywords
peer analysis, data preparation, knowledge graph, schema matching, large language model, synthetic data generation, privacy-preserving data sharing, data quality, OLAP
Degree Awarded
PhD in Computer Science
Discipline
Databases and Information Systems
Supervisor(s)
LI, Yuchen
First Page
1
Last Page
145
Publisher
Singapore Management University
City or Country
Singapore
Citation
WANG, Sha.
Roll-up and drill-down as a structured curation framework for peer analysis over complex information: Semantic navigation, schema alignment, and synthetic data generation. (2026). 1-145.
Available at: https://ink.library.smu.edu.sg/etd_coll/935
Copyright Owner and License
Author
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.