Provenance for Multi-Source Datasets

EmVIS Lab Meeting, 9 May 2024

Cynthia A. Huang

Department of Econometrics and Business Statistics, MBUS

Introduction

About Me!

• 👩‍🎓 PhD Candidate in Econometrics & Business Statistics, Monash Business School
• 🙏 Supervised by: Rob Hyndman (EBS), Simon Angus (Econ) and Sarah Goodwin!
• 👋 affiliated at Monash with:
  • NUMBATS (Non-Uniform Monash Business Analytics Team)
  • SoDa Labs – alternative data for social science insights
  • MDFI (Monash Data Futures Institute)
• 💱 Previously: Economics at Unimelb

About Me!

• 📊 Interested in adapting data for research purposes:
  • web-scraped retail product & price data for public health research
  • mining wikipedia edit metadata for evidence of state-sponsored misinformation
  • train/test sampling strategies for satellite (image) deep learning
• 👩‍🎓 Provenance and statistical imputation for multi-source datasets
  • 🇩🇪 SoDa Statistiks at LMU in June
  • 🇨🇦 InfoVis group at UBC from Jul-Oct
• 🧗🏻‍♀️ climber, 🎙️ regular host on The Random Sample podcast

Visualising Multi-Source Harmonisation Logic

• Crossmaps Framework
• CS Research (Collaboration) Opportunities
  • Data provenance communication
  • Interactive data merging (multi-table)
• Other downstream projects

 

Domain Problem

 

 

Provenance Model

 

 

Visual Encoding

Interactive Tools

 

• problem set up (borrow Kai’s slide)
  • semantics of data preparation – beyond translating (NLP, glyphs, programming languages)
  • extracting, representing logic of transformations
  • from imperative to declarative specification
• unified framework
  • dual problems appear when you think about [data imputation model]
    • implementation & validation of logic: sensitivity analysis, statistical properties
    • communication & visualisation of logic: code-free implementation interfaces, workflows
• vis/chi challenges / opportunities
  • extraction of data imputation model (graph computation) from existing scripts
    • track-vec; modifying input rather than diff tracing
  • interactive authoring and modification of data imputation model (i.e. harmonisation logic)
• domain applications
  • application 01 –> trade/labour concordance for cross-country comparison
    • xmap workshop paper + sankey layout problem
  • application 02 –> geographic concordance
    • example from merry; what’s the best way to display this information?
  • application 03 –> causal loop diagrams
    • harmonisation can happen between any sets of “related” objects (i.e. nodes OR links)

Crossmaps Framework

Ex-Post Harmonisation 1/2

Ex-post (or retrospective) data harmonization refers to procedures applied to already collected data to improve the comparability and inferential equivalence of measures from different studies (Kołczyńska 2022)

Figure 1: Procedures in ex-post harmonisation

Ex-Post Harmonisation 2/2

Defining or selecting mappings between classifications or taxonomies,

Implementing and validating mappings on given data,

Documenting and analysing the implemented mapping.

Current Approach: Input/Output Comparison

Proposed Alternative: Input & Function Capture

Implications (ASC23 Poster)

https://www.cynthiahqy.com/research/assets/asc-poster.pdf

Contributions

• non-hierarchical provenance task abstraction (c.f. Bors et al. 2019)
• compare with inference/diff-based solutions:
  • smallsets, anteater, COMANTICS etc.
  • translation -> summary/sense-making

• unified approach to:
  • transformation auditing (data quality)
  • transformation documentation (data provenance)
  • data imputation modelling (statistical robustness)

Future Work

Visualisation Concepts

Transformation Logic:

Merged Dataset Quality:

Provenance Communication & Tools

 

Domain Problem

 

 

Provenance Model

 

 

Visual Encoding

Interactive Tools

 

• layouts & ordering (similar to sankey diagram optimisation)
• encoding distribution weights & relation styles
• properties of merged dataset

Data Wrangling Tools

• implementing graph, matrix & table representation in R, with symbolic (fractional) weights
• extracting mapping logic from existing scripts
  • manipulate data input
  • parse AST into computational graph
• authoring and auditing interfaces for non-technical collaborators

Other Projects

• summary properties of the computational graph as a data imputation model
• provenenace model for causal loop diagrams
• multiverse analysis of alternative mapping decisions
• provenance abstraction for statistical sampling vs. database queries

References

Bors, Christian, John Wenskovitch, Michelle Dowling, Simon Attfield, Leilani Battle, Alex Endert, Olga Kulyk, and Robert S. Laramee. 2019. “A Provenance Task Abstraction Framework.” IEEE Computer Graphics and Applications 39 (6): 46–60. https://doi.org/10.1109/MCG.2019.2945720.

Kołczyńska, Marta. 2022. “Combining Multiple Survey Sources: A Reproducible Workflow and Toolbox for Survey Data Harmonization.” Methodological Innovations 15 (1): 62–72. https://doi.org/10.1177/20597991221077923.