Physically-interpretable classification of biological network dynamics for complex collective motions

Keisuke Fujii, Naoya Takeishi, Motokazu Hojo, Yuki Inaba, Yoshinobu Kawahara

TL;DR

This paper introduces Graph Dynamic Mode Decomposition (Graph DMD), a data-driven spectral analysis method to classify complex collective motions in biological networks. By analyzing dynamic network properties, they successfully classified team strategies in basketball, demonstrating that individual player context is crucial alongside physical interactions.

What Problem Does This Paper Address?

Understanding the dynamics of complex biological networks, especially those involving collective motion, is challenging because the relationships between individual components (nodes) and the overall network structure can change rapidly and unpredictably. Traditional network analysis often struggles to capture these transient dynamics. This paper aims to bridge this gap by developing a method to classify these complex collective motions based on physically interpretable dynamical properties, moving beyond static network structures.

Methods

The core of this work is the application of Graph Dynamic Mode Decomposition (Graph DMD). This method extends traditional Dynamic Mode Decomposition (DMD) to handle graph-structured data, specifically sequences of adjacency matrices representing network states over time.

Here's a breakdown of the approach:

  • Graph DMD Framework: Graph DMD decomposes the time-varying adjacency matrices of a network into a set of "spatial modes" (representing network structures) and their corresponding "temporal dynamics." This allows for the extraction of underlying linear dynamics from potentially nonlinear systems.
  • Data-Driven Spectral Analysis: Instead of relying on predefined equations, Graph DMD analyzes observed data to extract these dynamical properties. This is particularly useful for complex biological systems where governing equations are unknown or too complex to model directly.
  • Temporal Sliding Windows: To capture the transient nature of collective motions, Graph DMD is applied to overlapping temporal windows of the data. This allows for the analysis of how network dynamics evolve over short periods.
  • Feature Extraction: The decomposed Graph DMD modes are then used to create feature vectors. The paper highlights that both the spectral properties of these modes and the contextual information of individual nodes are important for classification.
  • Classification: These feature vectors are used to train classification models (e.g., logistic regression) to distinguish between different types of collective behaviors.

Datasets

The primary dataset used in this study is player-tracking data from actual men's Asian basketball games. This data includes the XY positions of players and the ball on the court, recorded at 25 frames per second. The authors analyzed 220 minutes of play, segmenting it into 319 "attack-segments" (periods from entering the attacking half to before a shot). These attack-segments were manually labeled into two main recognition tasks:

  1. Team-Defence: Zone defence vs. Person-to-person defence.
  2. Team-Offence: Offence with screen-play vs. Offence without screen-play.

This dataset is a real-world example of complex collective motion with strategic interactions.

Key Insights and Ideas

  • Beyond Static Networks: The study emphasizes that understanding dynamic collective motions requires analyzing how network structures evolve over time, not just their static properties.
  • The Power of Context: A crucial finding is that classifying collective motions requires not only understanding the physical interactions between agents but also their individual contextual information (e.g., their proximity to the ball or other key players). This "contextual node information" proved critical for accurate classification.
  • Physically and Semantically Interpretable Modes: Graph DMD provides "modes" that can be visualized and interpreted. These modes reveal specific relationships between agents (e.g., between defenders, or attackers and defenders) that are characteristic of different collective behaviors. This allows for both physical understanding (e.g., spatial relationships) and semantic interpretation (e.g., strategic roles).
  • Label-Specific Spectra: The research discovered "label-specific stronger spectra" – particular patterns of interaction (represented by mode strengths) that are uniquely associated with specific collective behaviors (e.g., zone defense).

Key Results

  • Superior Classification Performance: Graph DMD, particularly using its "spectrum" features, outperformed most other tested methods in classifying team-defence and team-offence strategies.
  • Accuracy Gains: Statistical analysis showed significant improvements in accuracy, AUC, and F-measure for Graph DMD spectrum compared to methods like GDMD Laplacian, DMD spectrum, and hand-crafted features. For instance, accuracy in GDMD spectrum was significantly higher than most other methods.
  • Interpretable Patterns: Visualizations of Graph DMD modes revealed distinct patterns for different strategies. For example, zone defense showed stronger interactions among nearest defenders, while offense with a screen-play highlighted interactions among nearest attackers.
  • Validation: The method was validated on a basketball dataset, demonstrating its applicability to real-world complex collective motion scenarios.

Limitations and Future Work

  • Equation-Free Validation: While the method demonstrates strong classification performance, the authors acknowledge that a general quantitative validation method for extracting true dynamical properties (like precise frequencies and decay rates) in an equation-free manner is still needed.
  • Local Interaction Dynamics: The current approach focuses on global dynamics. Future work could explore methods to better capture more specific local interaction dynamics, such as local competitive and cooperative plays, which could enhance practical applications like score prediction.
  • Generalizability: While successful for basketball, further adaptation and customization might be needed for other complex collective motions, especially those with different interaction rules or data structures.
  • Data Requirements: The effectiveness of Graph DMD might depend on the richness of the data and the ability to capture sufficient eigenfunctions.

The authors suggest that their approach has broad potential applications for analyzing complex global dynamics in groups of living organisms or artificial agents, offering a powerful tool for understanding and potentially predicting collective behaviors.

Summary updated

Abstract

Understanding biological network dynamics is a fundamental issue in various scientific and engineering fields. Network theory is capable of revealing the relationship between elements and their propagation; however, for complex collective motions, the network properties often transiently and complexly change. A fundamental question addressed here pertains to the classification of collective motion network based on physically-interpretable dynamical properties. Here we apply a data-driven spectral analysis called graph dynamic mode decomposition, which obtains the dynamical properties for collective motion classification. Using a ballgame as an example, we classified the strategic collective motions in different global behaviours and discovered that, in addition to the physical properties, the contextual node information was critical for classification. Furthermore, we discovered the label-specific stronger spectra in the relationship among the nearest agents, providing physical and semantic interpretations. Our approach contributes to the understanding of principles of biological complex network dynamics from the perspective of nonlinear dynamical systems.
BibTeX
@inproceedings{fujii2020physically,
  author = {Keisuke Fujii and Naoya Takeishi and Motokazu Hojo and Yuki Inaba and Yoshinobu Kawahara},
  title = {{Physically-interpretable classification of biological network dynamics for complex collective motions}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-58064-w},
}
Figure