Profiling Driver Behavior Using Graph Centrality

22 concepts11 papers1 proposal

Traditional autonomous driving systems typically treat trajectory prediction and behavior profiling as coordinate-tracking problems. They observe a vehicle’s absolute position, velocity, and acceleration over time, often relying heavily on high-definition (HD) maps to understand environmental constraints.

However, driving is inherently a social, multi-agent interaction. An aggressive driver is not merely defined by a high speedometer reading, but by how they interact with the vehicles around them—tailgating, weaving through traffic, or forcing others to yield.

To capture this social dimension, a prominent research direction models traffic scenes as dynamic geometric graphs. By treating vehicles as nodes and their spatial proximity as edges, researchers can leverage classical social network centrality metrics to profile driver behavior in real time without requiring massive, manually labeled datasets.

The Topology of Traffic

When traffic is represented as a graph, the structural properties of the network directly reflect the physical behaviors of the drivers. As outlined by Rohan Chandra et al. in Game-Theoretic Planning for Autonomous Driving among Risk-Aware Human Drivers, constructing a weighted undirected graph based on spatial distances allows autonomous systems to translate raw coordinates into topological features.

Different driving styles manifest as distinct structural patterns within this graph:

  • Degree Centrality (Neighbor Discovery): Degree centrality measures how many connections a node has. In a traffic graph, a vehicle's degree represents how many other agents are within its immediate interaction zone. Tianrui Guan et al. demonstrate in Forecasting Trajectory and Behavior of Road-Agents Using Spectral Clustering in Graph-LSTMs that tracking the rate of change in an agent's graph degree reveals its "neighbor discovery rate." A driver rapidly gaining and losing neighbors is likely weaving aggressively through a dense platoon.
  • Closeness Centrality (Platoon Position): Closeness centrality quantifies how near a node is to all other nodes in the network. In CMetric: A Driving Behavior Measure Using Centrality Functions, the authors show that closeness centrality effectively tracks a vehicle’s position relative to the center of a traffic cluster. Sudden shifts in closeness indicate longitudinal maneuvers, such as aggressive tailgating or rapid acceleration into a cluster.
  • Betweenness and Eigenvector Centrality (Influence and Friction): Betweenness centrality identifies nodes that act as bridges or bottlenecks between different parts of a network. As explored by Haicheng Liao et al. in MFTraj: Map-Free, Behavior-Driven Trajectory Prediction for Autonomous Driving, vehicles with high betweenness centrality are those that constrain the movement of others. Similarly, eigenvector centrality measures a node's global influence by looking at how well-connected its neighbors are. In BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving, the authors use eigenvector centrality to quantify the "social friction" an agent introduces, identifying key vehicles that dictate the local flow of traffic.

Capturing Dynamics via Derivatives and Spectrums

A static snapshot of a traffic graph can tell you where a vehicle sits in the social hierarchy, but behavior is fundamentally dynamic. To distinguish between a naturally dense traffic jam and an aggressive maneuver, models must analyze how these graph structures distort over time.

One approach is to track the mathematical derivatives of centrality metrics. In Using Graph-Theoretic Machine Learning to Predict Human Driver Behavior, the authors fit quadratic polynomials to centrality values over a sliding temporal window. The first derivative of this polynomial represents the velocity of behavioral change, while the second derivative captures the acceleration or sharpness of a maneuver. A smooth lane change results in a gradual shift in closeness centrality, whereas a reckless cut-in causes a sharp spike in the centrality derivatives, signaling a high-severity maneuver.

[Raw Trajectories] ──> [Dynamic Geometric Graph] ──> [Centrality Metrics over Time]
                                                               │
                                                               ▼
[Behavioral Profile] <── [Maneuver Intensity] <── [1st & 2nd Derivatives]

Alternatively, the entire traffic graph can be evaluated through its spectrum. Rohan Chandra et al. developed a framework called GraphRQI: Classifying Driver Behaviors Using Graph Spectrums, which analyzes the eigenvalues of the graph Laplacian. Because aggressive maneuvers like tailgating or cutting off another vehicle force sudden, localized changes in edge weights, they leave distinct mathematical "fingerprints" in the graph's spectrum. By tracking these spectral distortions, the system can classify drivers on a spectrum from timid to reckless in real time.

Why It Works: Unsupervised, Map-Free Awareness

Shifting from coordinate-based tracking to graph-theoretic profiling provides two massive advantages for autonomous vehicles:

  1. Eliminating Manual Labels: Traditional machine learning models require massive datasets of manually annotated driving behaviors (e.g., labeling thousands of video clips as "aggressive"). Graph centrality metrics bypass this entirely. As noted by Haicheng Liao et al. in Minds on the Move: Decoding Trajectory Prediction in Autonomous Driving with Cognitive Insights, network topology serves as an excellent unsupervised proxy for human-like cognitive insights, profiling risk and intent purely through the geometry of interaction.
  2. Map-Free Generalization: Because metrics like betweenness and eigenvector centrality inherently capture how vehicles constrain one another, the graph structure implicitly encodes the geometry of the road. If a lane narrows, vehicles are forced closer together, changing the graph topology. This allows models like the one proposed in MFTraj to accurately forecast long-term trajectories without relying on expensive, brittle HD map data.

Ultimately, these topological profiles feed directly into downstream planning. For instance, in GamePlan: Game-Theoretic Multi-Agent Planning with Human Drivers at Intersections, Roundabouts, and Merging, the authors use degree centrality to calculate an "impatience score" for human drivers at unsignaled intersections. By incorporating these scores into a game-theoretic auction planner, the autonomous vehicle can safely yield to aggressive drivers and assert its right-of-way with conservative ones, mimicking the intuitive, social reasoning of a human driver.

Key Foundational Work

CMetric: A Driving Behavior Measure Using Centrality Functions (2020) — contributed 3 concept elements to this direction.
Forecasting Trajectory and Behavior of Road-Agents Using Spectral Clustering in Graph-LSTMs (2020) — contributed 2 concept elements to this direction.
GraphRQI: Classifying Driver Behaviors Using Graph Spectrums (2019) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_graph_based_behavior_profiling_a8def08d,
  title = {graph-based behavior profiling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-based-behavior-profiling/rdr_a8def08dc9654c16aa77d5bb7fa690d0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_graph_based_behavior_profiling_a8def08d_v1,
  title = {graph-based behavior profiling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-based-behavior-profiling/rdr_a8def08dc9654c16aa77d5bb7fa690d0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake