Pavement Missing Condition Data Imputation through Collective Learning-Based Graph Neural Networks

This paper proposes a collective learning-based Graph Convolutional Network model that effectively imputes missing pavement condition data by integrating features from adjacent road sections and capturing dependencies between observed conditions, demonstrating promising results in a Texas Department of Transportation case study.

The Gist

Imagine a massive, sprawling city of roads. Every few years, inspectors drive over these roads to give them a "health score" (like a report card from 1 to 100). This score tells the city if the road is "Very Good" or "Very Poor" and helps them decide where to spend money on repairs.

But here's the problem: The data is messy.

Sometimes the sensors break, sometimes the inspectors miss a section, and sometimes the schedule gets messed up. This leaves big holes in the data. It's like trying to read a book where someone has torn out random pages. If you try to guess the story based only on the pages you have, you might get it wrong.

The authors of this paper, Ke Yu and Lu Gao, wanted to fix these missing pages using a clever new trick called Collective Learning Graph Neural Networks (CLGNN).

Here is how they did it, explained simply:

1. The Old Way: Ignoring the Neighbors

Traditionally, when data is missing, engineers used two main methods:

• The "Delete" Method: If a road section is missing data, they just threw that whole section away. This is like deleting a chapter from a book because a few words are missing. You lose a lot of information.
• The "Simple Guess" Method: They looked at the road's history (how it was doing last year) or its traffic volume to guess the current score. This is like guessing a character's mood in a story just by looking at what they wore yesterday, ignoring who they are talking to right now.

The Flaw: These methods treat every road section as an island. They forget that roads are connected. A pothole on one street often means the street next to it is also suffering.

2. The New Way: The "Gossip Network"

The authors realized that roads are connected like a giant web. If you want to know the health of a specific road, you shouldn't just look at its own history; you should ask its neighbors.

They built a model that acts like a super-smart gossip network:

• The Graph: They turned the road map into a digital "friendship graph." Every road section is a person, and the lines connecting them are friendships.
• The Collective Learning: The model doesn't just look at one person; it looks at the whole group. If "Road A" is missing its health score, the model asks: "What are Road A's neighbors doing? Are they all 'Very Poor'?" If the neighbors are all struggling, it's highly likely Road A is struggling too, even if we don't have the data yet.

3. How the "Magic" Works (The Analogy)

Imagine you are at a party, and you want to guess how a specific guest (let's call him "Road X") is feeling, but you can't see his face.

• Old Method: You look at a photo of Road X from last year. You guess, "He looked happy then, so he's probably happy now."
• CLGNN Method: You look at the three people standing right next to him. They are all crying and holding tissues. You immediately realize, "Oh, Road X is probably crying too, even though I can't see him."

The model uses this "neighborhood vibe" to fill in the missing data. It learns that if a group of connected roads are deteriorating, the missing one in the middle is likely deteriorating too.

4. The Test Drive

To see if this worked, they tested it on real roads in Austin, Texas.

• They took a bunch of real data and secretly "hid" 30% of it (pretending it was missing).
• They asked their new AI model to guess the hidden scores.
• They compared the AI's guesses to the actual answers.

The Result: The new model was the clear winner. It was about 5% more accurate than the best traditional methods and much better than standard deep learning models. It successfully used the "neighborhood gossip" to fill in the blanks.

Why This Matters

In the real world, missing data leads to bad decisions. If a city thinks a road is "Good" because of a bad guess, they might delay repairs, and the road could collapse. If they think it's "Bad" when it's actually "Good," they waste money fixing a road that didn't need it.

By using this Collective Learning approach, cities can:

1. Fill in the gaps without throwing away data.
2. Make smarter decisions about where to spend tax dollars.
3. Keep roads safer by getting a more accurate picture of the whole network, not just the parts they can see.

In short: Instead of treating every road as a lonely island, this paper teaches computers to look at the whole neighborhood to figure out what's really going on. It's a smarter, more connected way to keep our roads in shape.

Technical Summary

Here is a detailed technical summary of the paper "Pavement Missing Condition Data Imputation Through Collective Learning-Based Graph Neural Networks" by Ke Yu and Lu Gao.

1. Problem Statement

Pavement condition data is critical for managing road network maintenance and rehabilitation. However, this data is frequently incomplete due to sensor failures, non-periodic inspection schedules, and other operational issues.

• Consequences of Missing Data: Systematic missing data leads to information loss, reduced statistical power, and biased assessments of road network health.
• Limitations of Existing Methods: Current approaches typically fall into three categories:
  1. Deletion: Discarding entire data points with missing values (reduces dataset size).
  2. Simple Interpolation: Filling gaps using basic mathematical interpolation (ignores complex dependencies).
  3. Statistical Imputation: Using historical data and covariates (e.g., traffic, age) but often failing to fully capture spatial dependencies between adjacent road sections.
• The Gap: Traditional Graph Neural Networks (GNNs) model feature dependencies between adjacent nodes but often neglect the dependencies between node labels (i.e., the condition state of one section influences the likely condition state of its neighbor).

2. Methodology

The authors propose a Collective Learning-based Graph Neural Network (CLGNN) to impute missing pavement condition values by leveraging both node features and label dependencies.

A. Data Representation

• Graph Structure: The road network is modeled as a graph $G = (V, E)$, where nodes ($V$) represent pavement sections and edges ($E$) represent spatial connections between adjacent sections.
• Input Variables:
  • Target: Pavement Condition Score (CS), discretized into five states: Very Good (90–100), Good (70–89), Fair (50–69), Poor (35–49), and Very Poor (0–34).
  • Features ($x$): Historical condition scores (2014–2017), traffic volume (18-kip ESAL), functional class (e.g., Interstate, Farm-to-Market), and pavement type.
• Dataset: Data from the Texas Department of Transportation (TxDOT) Austin District (2014–2018). Climate factors were excluded as the study area is climatically homogeneous.

B. The CLGNN Framework

Unlike standard GNNs that focus on node features, CLGNN incorporates collective learning, which explicitly models the dependencies between the labels (conditions) of connected nodes. The framework operates through an iterative four-step process:

1. Masking: A random binary mask is applied to the data (simulating missing values).
2. Prediction: A Graph Convolutional Network (GCN) layer predicts the label distribution for the masked nodes.
   • GCN Formula: $h^{(k)}_i = \sum_{j \in N(i) \cup \{i\}} \frac{1}{\sqrt{deg(i)deg(j)}} (\Theta h^{(k-1)}_j)$
3. Integration: The predicted labels are combined with the available true labels to form a new input vector.
4. Optimization: The model parameters are updated by minimizing the loss function.

C. Experimental Setup

• Simulation: 30% of the 2018 condition scores were masked as missing.
• Masking Strategy: The selection of masked sections was not purely random; it considered network connectivity to simulate realistic "clustered" missing data scenarios (common in real-world inspections where a specific route segment might be missed).
• Baseline Models: The CLGNN was compared against:
  • Traditional ML: CART, Neural Networks (NN), Random Forest (RF).
  • Standard GNNs: Graph Convolutional Network (GCN), GraphSAGE.

3. Key Contributions

1. Novel Architecture: Adaptation of the Collective Learning framework to pavement management, specifically addressing the limitation of standard GNNs by modeling label dependencies (the correlation between the condition states of adjacent sections) rather than just feature dependencies.
2. Realistic Data Simulation: The study implemented a connectivity-aware masking strategy to simulate real-world missing data patterns (clustering), providing a more rigorous test of model robustness than random masking.
3. Integration of Spatial and Temporal Data: The model successfully fuses historical temporal data (past condition scores) with spatial graph data (neighbor relationships) to improve imputation accuracy.

4. Results

The study evaluated model performance based on classification accuracy for the imputed condition states.

• Performance Comparison (Table 2):
  • CLGNN (Proposed): 0.773 (Highest Accuracy)
  • GCN: 0.725
  • GraphSAGE: 0.721
  • Random Forest (RF): 0.712
  • CART: 0.654
  • Neural Network (NN): 0.556
• Key Finding: The CLGNN model outperformed all other benchmarks, achieving an approximate 5% improvement in accuracy over the next best model (GCN). This demonstrates that explicitly modeling the dependency between the condition labels of neighboring sections significantly enhances imputation performance.

5. Significance and Conclusion

• Practical Impact: The proposed method provides a robust solution for transportation agencies to recover missing pavement data without discarding valuable network segments. This leads to more accurate asset management, better maintenance scheduling, and optimized rehabilitation budgets.
• Theoretical Advancement: The paper validates that "collective learning" (considering label dependencies) is superior to standard graph learning for tasks where the target variable of a node is strongly correlated with the target variables of its neighbors.
• Future Work: The authors suggest extending this approach to impute other specific condition indicators, such as cracking severity and roughness, and testing the model on diverse road networks with varying climatic conditions.

In summary, this research demonstrates that leveraging the spatial interdependence of pavement conditions through a Collective Learning-based GNN is a highly effective strategy for handling missing data in large-scale infrastructure management systems.