🤖 machine learning

ImputeViz: A Visual Analytics Dashboard for Diagnosing Missing Data and Comparing Imputation Methods

ImputeViz is an interactive visual analytics dashboard that enables researchers to diagnose missing data patterns, configure and compare diverse imputation methods (including a novel geographically informed gKNN), and evaluate their impact through coordinated views and quantitative metrics to support robust, accountable data analysis.

Original authors: Aitik Dandapat, Lalith Punepalle Raveendrareddy, Mithilesh Kumar Singh, Klaus Mueller

Published 2026-07-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aitik Dandapat, Lalith Punepalle Raveendrareddy, Mithilesh Kumar Singh, Klaus Mueller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but half the clues in your notebook have been torn out. In the world of data science, these torn-out clues are called missing data. Sometimes, they are missing because of a random accident (like a coffee spill), but often, they are missing on purpose or because of a pattern—like a town that is too small to report its numbers, or a specific type of person who refuses to answer a survey. If you just guess the missing numbers without thinking, you might end up with a completely wrong picture of the world.

Enter ImputeViz, a new digital dashboard created by researchers at Stony Brook University. Think of it as a "super-powered detective's board" that doesn't just fill in the blanks; it helps you figure out why the blanks are there, how to fill them, and who helped you fill them.

The Problem with "Black Box" Guessing

Usually, when analysts have missing data, they pick a standard tool (like a magic wand) to guess the missing values and hope for the best. They might run one tool, then another, and try to remember if the numbers looked different. It's like trying to compare two different maps of the same city, but one map is zoomed in and the other is zoomed out, making it impossible to tell which route is actually better. The authors argue that this "black box" approach is risky because it hides why a number was chosen and makes it hard to see if the guess makes sense.

The Solution: A Dashboard for "What-If" Scenarios

ImputeViz changes the game by letting you play with different guessing strategies side-by-side. The system includes several famous "guessing engines":

MICE: A method that asks other variables in the data for help, like a group chat solving a puzzle.
Random Forest & XGBoost: Powerful tree-based methods that look for complex patterns.
kNN (k-Nearest Neighbors): A method that says, "If you look like your neighbors, you probably have the same values."
gKNN (Geographically Informed kNN): This is the system's special invention. It's a super-charged version of kNN that doesn't just look at how similar people are; it also looks at how close they are on a map.

The Secret Weapon: gKNN

The researchers introduced gKNN to handle cases where location matters. Imagine you are trying to guess the average temperature in a town you've never visited. A normal guesser might look at a town with the same population size, even if it's on the other side of the country. gKNN is smarter: it blends "social similarity" (like income or education) with "geographic distance." It asks, "Who are the neighbors that are both socially similar and physically close?"

In their tests, when they tried to fill in missing data for prescription-opioid mortality rates across US counties (from 2000 to 2022), gKNN suggested it was the most accurate guesser. In a specific test where they hid data to mimic real-world reporting gaps, gKNN had an average error (MAE) of 2.96, while the next best methods hovered around 3.28 to 3.46. The authors note that this improvement was most obvious when the missing data was clustered in low-county areas, suggesting that knowing where you are helps you guess what is missing.

However, the paper is careful not to say gKNN wins everywhere. When they tested it on a telecom churn dataset (which has no map or geography involved), a different method called Random Forest took the lead, beating MICE with a massive difference in error. This proves the paper's main point: there is no single "best" tool for every job. You have to test them.

The "Donor" Detective Work

One of the coolest features of ImputeViz is provenance. When the system guesses a missing number, it doesn't just give you the number; it shows you who helped. If the system guesses the mortality rate for County A, it highlights the specific "donor" counties that contributed to that guess.

In the opioid study, the researchers found that sometimes a standard method (kNN) would guess a number correctly but rely on a "donor" that was hundreds of miles away. The map view in ImputeViz made this obvious, showing that the guess was based on a long, shaky connection. The gKNN method, by contrast, stuck to nearby neighbors, making the guess feel more trustworthy. The authors suggest that seeing these "donor paths" helps analysts decide if a guess is plausible or if it's just a lucky accident.

How Sure Are They?

The researchers didn't just guess that this works; they measured it. They ran their system on two real-world datasets:

US County Opioid Mortality (2000–2022): They tested four different ways of hiding data (randomly, or by simulating real reporting gaps) and found that gKNN consistently had the lowest error rates in these simulations.
Telecom Churn: A non-geographic dataset where they found tree-based models were better.

They also asked 7 people to try out the dashboard. The users gave it a "SUS" (System Usability Scale) score of 75.4 (out of 100), which is considered a good score. The testers loved being able to switch between methods instantly and see the results update without the scales changing, which made comparing them much easier.

The Bottom Line

ImputeViz suggests that fixing missing data isn't just about running a math formula; it's about understanding the story behind the missing numbers. By letting analysts compare different methods side-by-side, check the "donors" behind every guess, and see how geography plays a role, the system helps them make smarter, more transparent choices. It doesn't claim to have solved the problem of missing data forever, but it offers a much clearer way to navigate the mess.

Technical Summary: ImputeViz

Problem Statement

Missing data remains a critical barrier in scientific, social science, and public health research, often introducing bias and forcing analysts to rely on "black-box" preprocessing steps. Current workflows frequently lack support for cross-method visual comparison or for inspecting the provenance of imputed values. Analysts are often compelled to run models sequentially, mentally reconciling shifting scales and parameters, which hinders the ability to diagnose missingness mechanisms (MCAR, MAR, MNAR) and select the most appropriate imputation strategy based on evidence rather than familiarity.

Methodology and System Design

ImputeViz is an integrated visual analytics dashboard designed to support the full imputation workflow: diagnosing missingness, configuring models, comparing outcomes, and inspecting provenance. The system is built on a Python/FastAPI backend and a React/TypeScript/D3.js frontend.

Core Components

Imputation Models: The system integrates widely used algorithms including MICE (IterativeImputer with ridge chains), Random Forest, XGBoost, linear regression, and non-spatial kNN.
gKNN (Geographically informed kNN): A novel contribution of the system, gKNN extends standard kNN by blending socioeconomic and spatial distances. It computes a distance matrix $D_{blend} = \alpha D_{socio} + (1-\alpha) D_{geo}$ , where the blending weight $\alpha$ and neighborhood size $k$ are optimized via Bayesian optimization using training data.
Evaluation Framework: To ensure fair comparison, the system applies a consistent 20% random holdout mask (and structured masking for suppression-like scenarios) across all algorithms. It reports Mean Absolute Error (MAE), Root Mean Square Error (RMSE), $\Delta$ RMSE, and runtime.
Visual Analytics Interface: The dashboard features seven coordinated panels (Figure 2) that facilitate an iterative workflow:
- Diagnosis: Heatmaps and co-missingness summaries reveal missingness structures and correlations with covariates.
- Comparison: A "Method Comparison Summary" table ranks algorithms by performance metrics.
- Validation: Linked scatterplots and histograms with locked axis scales allow analysts to visually inspect distributional shifts and relationship preservation when switching methods.
- Provenance: For geospatial data, a geomap view exposes donor counties and their contributions, enabling analysts to verify if imputed values rely on geographically and socioeconomically plausible neighbors.

Design Goals

The system is guided by four primary goals:

DG1: Make missingness legible through immediate summaries and pre-imputation relationship views.
DG2: Enable fair cross-method comparison via consistent masking, feature handling, and stable-scale visualizations.
DG3: Support visual plausibility checks to detect local artifacts, tail distortions, or implausible clusters that aggregate error metrics might hide.
DG4: Preserve provenance and spatial interpretability, allowing users to interrogate the "who" and "why" behind specific imputed values.

Key Contributions

The paper outlines three primary contributions:

A Visual Analytics Workflow: A method-agnostic environment for missingness diagnosis and cross-method comparison under stable-scale linked views, reducing cognitive overhead during model switching.
Benchmarking Backend: A system that caches per-method runs and enforces consistent holdout-based evaluation, ensuring that comparisons are driven by evidence rather than sequential mental reconciliation.
gKNN Algorithm: A geographically informed kNN variant with donor provenance that can be directly interrogated in the interface, blending socioeconomic and spatial distances to improve robustness in spatially autocorrelated data.

Results

The system was evaluated on two datasets:

US Prescription-Opioid Mortality (2000–2022): A geospatial dataset with structured missingness (suppression in low-count counties).
Telecom Churn Dataset: A non-geographic dataset used to test non-spatial imputation.

Performance Findings:

Spatial Context: In the opioid mortality dataset, gKNN achieved the lowest average MAE across all masking settings (Suppression, Intermediate, and Random), with the most significant gains in suppression-like settings where geographic context is critical. Paired Wilcoxon tests confirmed gKNN's superiority over standard kNN ( $p \approx 0.0396$ ).
Non-Spatial Context: In the telecom churn dataset, Random Forest outperformed MICE and gKNN, demonstrating that no single imputer is universally superior and that flexible non-linear models may dominate in purely tabular settings.
Provenance Utility: Case studies revealed that while standard kNN could achieve competitive error scores, it sometimes relied on long-range donors that were socioeconomically similar but geographically distant. The gKNN provenance view allowed analysts to identify these mismatches and prefer the more localized, contextually appropriate gKNN estimates.

Usability:
In a formative study with 7 participants, the system received a mean System Usability Scale (SUS) score of 75.4. Participants specifically valued the ability to switch methods with immediate visual updates and the stable axis scales that facilitated direct comparison.

Significance and Claims

The paper claims that ImputeViz shifts the paradigm of missing data handling from a predictive task to an analytic reasoning problem. By integrating execution, quantitative benchmarking, and provenance-aware visuals, the system:

Democratizes Method Selection: It prevents "algorithm lock-in" by allowing analysts to select strategies based on evidence specific to their data's missingness structure (e.g., spatial vs. tabular).
Enhances Trust: Through donor-level provenance and stable visual scales, it makes the imputation process auditable, allowing analysts to verify that borrowed values are plausible within their specific domain context.
Supports Robustness: It highlights that spatially informed imputation improves accuracy when missingness is structurally tied to place, while acknowledging that non-spatial models remain superior in purely tabular contexts.

The authors conclude that the system improves transparency and supports evidence-based model choice, though they note limitations regarding computational costs for richer uncertainty analyses and scalability at finer geographic resolutions.