A Graph-Prediction-Based Approach for Debiasing Underreported Data

Hanyang Jiang, Yao XieICASSP 2024 · DOI

TL;DR

This paper introduces GRAUD, a novel graph-based algorithm that efficiently debiases underreported data by jointly estimating true event counts and discovery probabilities. It leverages spatial relationships and avoids strong prior assumptions, offering a practical solution for fields like crime and public health data analysis.

What Problem Does This Paper Address?

Many real-world datasets suffer from underreporting, meaning the observed data is a subset of the true underlying events. This is particularly problematic when trying to estimate both the total number of events ($n$) and the probability of observing each event ($p$). The core challenge is an "identifiability issue": with only the observed count ($y$), there are infinitely many pairs of ($n, p$) that could have produced it. Traditional methods often rely on strong Bayesian priors, which can be difficult to choose appropriately. This paper tackles the "binomial $n$ problem" where both $n$ and $p$ are unknown, aiming for a more data-driven approach that accounts for the inherent structure in the data.

Methods

The authors propose a graph-based debiasing algorithm called GRAUD. The key idea is to leverage the spatial or graphical structure of the data. They make two core assumptions:

  1. Spatial Smoothness: Discovery probabilities ($p_i$) are similar for neighboring regions. This is modeled using a graph Laplacian, penalizing large differences in $p_i$ between connected nodes.
  2. Region-Related Factors: The true count of incidents ($n_i$) is influenced by underlying socioeconomic and demographic factors, which can be modeled using a log-linear model.

These assumptions are incorporated into a constrained convex optimization problem. By taking the logarithm of the observed counts and the model parameters, the problem is transformed into a convex optimization problem that can be solved efficiently using an alternating minimization algorithm. This approach effectively reframes the underreporting problem as a graph signal separation task.

Datasets

  • Simulated Data: Used extensively for evaluating the algorithm's performance under controlled conditions.
  • Real-World Data: 911 call data from Atlanta for the year 2019, comprising approximately 580,000 instances. This data was partitioned into 78 distinct "beats" (police patrol areas), forming the nodes of the graph. Census data from 2019 (population size, income, education level) was used as covariates ($X$) for the log-linear model.

Key Insights and Ideas

  • Graph Structure as a Solution: The paper demonstrates that leveraging the spatial relationships (graph structure) between data points is a powerful way to overcome the identifiability issue in underreported data.
  • Debiasing as Signal Separation: By transforming the problem into a log domain and incorporating graph smoothness, the underreporting problem is elegantly framed as a graph signal separation task, similar to recent work in signal processing.
  • Avoiding Strong Priors: GRAUD's ability to perform well without requiring strong, pre-defined Bayesian priors makes it more broadly applicable and less sensitive to subjective choices.
  • Smoothness is Key: The effectiveness of the method hinges on the assumption that discovery probabilities are spatially smooth, which is often a reasonable assumption in real-world scenarios where factors like law enforcement presence or public awareness tend to vary gradually across regions.

Key Results

  • Accurate Recovery: In simulated experiments, GRAUD demonstrated close alignment between estimated and true values of $n$ and $p$, with relative $l_1$ errors as low as 0.029 for $n$ and 0.035 for $p$.
  • Rapid Convergence: The algorithm converges quickly, as visualized in Figure 2, and this convergence is not highly sensitive to the initial guess.
  • Impact of True Count ($n$): The accuracy of GRAUD's estimates for both $n$ and $p$ improves as the true count $n$ increases, reaching a steady state once $n$ surpasses 10. This suggests that larger underlying event counts lead to less noisy observations and better estimation.
  • Impact of Discovery Probability Smoothness: The accuracy also improves as the spatial variation in discovery probabilities ($p^T L p$) decreases, aligning with the core assumption of spatial smoothness.
  • Real-World Application: Applied to Atlanta's 911 call data, GRAUD identified variations in discovery probability across different beats, with higher probabilities observed in more prosperous downtown and midtown areas, which is intuitively explained by better public security.

Limitations and Future Work

  • Assumption Dependency: The accuracy of GRAUD is contingent on the fulfillment of its core assumptions, particularly spatial smoothness and the log-linear model for true counts. The paper acknowledges that these assumptions might become challenging to maintain in higher dimensions or with more complex data structures.
  • Noise Sensitivity: While the paper provides theoretical guarantees, the practical performance might still be affected by significant noise in the data or deviations from the assumed model.
  • Future Directions: The authors suggest broadening the framework to accommodate more general contexts, refining the algorithm for better noise resistance, and improving its functionality in higher dimensions. They also anticipate relaxing some of the current assumptions.

Summary updated

BibTeX
@inproceedings{jiang2024graph,
  author = {Hanyang Jiang and Yao Xie},
  title = {{A Graph-Prediction-Based Approach for Debiasing Underreported Data}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
  doi = {10.1109/icassp48485.2024.10447272},
}

Loading related content...

Figure