pith. sign in

arxiv: 2401.06412 · v1 · submitted 2024-01-12 · 💻 cs.HC

Understanding whole-body inter-personal dynamics between two players using neural Granger causality as the explainable AI (XAI)

Ryota Takamido , Chiharu Suzuki , Jun Ota , Hiroki Nakamoto This is my paper

Pith reviewed 2026-05-24 04:15 UTC · model grok-4.3

classification 💻 cs.HC
keywords neural granger causalityinter-personal dynamicsbaseballjoint velocitiesbody coordinationexplainable AIcausal relationshuman movement analysis
0
0 comments X

The pith

Neural Granger causality applied to joint velocities shows a pitcher's throwing arm exerts causal effects on a batter's hands, while weaker reverse effects correlate with batting performance.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper applies neural Granger causality to time-series data of 27 joint velocities from 16 pairs of expert baseball pitchers and batters. It defines causal strength as the magnitude of first-layer weights in neural networks trained to predict future states of each velocity variable. This reveals both intra-player and inter-player causal links, such as the pitcher's throwing arm influencing the batter's hands. A sympathetic reader would care because the approach offers a way to quantify bidirectional influences in real coordination tasks that conventional statistical methods have not previously resolved at the whole-body level. The additional correlation between certain causal strengths and performance outcomes indicates the extracted relations carry behavioral relevance.

Core claim

Causal relationships are defined as the size of the weight parameters of the first layer of a machine-learning model trained to predict the future state of a specific time-series variable. When this NGC procedure is run on input datasets of 27 joint resultant velocity time series collected from 16 expert pitcher-batter pairs, it identifies significant causal relations among intra- and inter-individual body components, including a causal effect from the pitcher's throwing arm to the batter's hands. Although causality running from batter to pitcher is lower overall, the magnitude of these directed effects correlates significantly with batter performance outcomes.

What carries the argument

Neural Granger causality, defined as the magnitude of first-layer weights in a neural network trained to forecast future joint-velocity states.

If this is right

  • Significant causal relations exist both within each player's body and between the two players' bodies.
  • Causality from batter to pitcher remains lower than the reverse direction yet still correlates with measurable performance differences.
  • NGC supplies an effective framework for mapping whole-body inter-personal coordination dynamics.
  • The technique supplies a perspective on complex human behavior that differs from conventional statistical approaches.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same NGC pipeline could be tested on other paired motor tasks such as dance or cooperative object manipulation to see whether similar inter-individual causal patterns appear.
  • If the weight-based causality measure holds up under direct intervention experiments, the method could supply quantitative targets for coaching focused on specific joint linkages.
  • Performance correlations raise the possibility that NGC metrics might serve as an objective index for evaluating coordination skill independent of outcome scores.
  • The asymmetry in causal strength between pitcher and batter directions invites comparison with biomechanical models of timing and anticipation.

Load-bearing premise

The size of first-layer weights in a neural network trained to predict future joint velocities directly measures causal influence rather than mere predictive association, and these magnitudes can be compared across movement directions and across individuals.

What would settle it

An experiment that applies a known physical intervention (such as physically constraining the pitcher's throwing arm) and then checks whether the corresponding NGC weight magnitudes decrease in the predicted direction would falsify the causal interpretation if no such change occurs.

read the original abstract

Background: Simultaneously focusing on intra- and inter-individual body dynamics and elucidating how these affect each other will help understand human inter-personal coordination behavior. However, this association has not been investigated previously owing to difficulties in analyzing complex causal relations among several body components.To address this issue, this study proposes a new analytical framework that attempts to understand the underlying causal structures behind each joint movement of individual baseball players using neural Granger causality (NGC) as the explainable AI. Methods: In the NGC analysis, causal relationships were defined as the size of the weight parameters of the first layer of a machine-learning model trained to predict the future state of a specific time-series variable. To verify the approach in a practical context, we conducted an experiment with 16 pairs of expert baseball pitchers and batters; input datasets with 27 joint resultant velocity data (joints of 13 pitchers and 14 batters) were generated and used for model training.Results: NGC analysis revealed significant causal relations among intra- and inter-individual body components such as the batter's hands having a causal effect from the pitcher's throwing arm. Remarkably, although the causality from the batter's body to pitcher's body is much lower than the reverse, it is significantly correlated with batter performance outcomes. Conclusions: The above results suggest the effectiveness of NGC analysis for understanding whole-body inter-personal coordination dynamics and that of the AI technique as a new approach for analyzing complex human behavior from a different perspective than conventional techniques.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes neural Granger causality (NGC) as an XAI method to analyze intra- and inter-individual whole-body dynamics in baseball pitcher-batter pairs. NGC is defined operationally as the magnitude of first-layer weights in a neural network trained to predict future joint velocities from 27 time-series inputs. Applied to data from 16 expert pairs, the analysis reports directed causal effects (e.g., pitcher's throwing arm to batter's hands) and a significant correlation between the weaker batter-to-pitcher causality and batter performance outcomes, positioning NGC as a new framework for interpersonal coordination beyond conventional techniques.

Significance. If the operational definition of causality were shown to recover known causal structures and satisfy standard identifiability conditions, the work could introduce a useful XAI lens for complex human movement data. The manuscript does not yet demonstrate this; the reported relations rest on an unvalidated proxy for causality, limiting the strength of the central claims about directed effects and performance correlations.

major comments (3)
  1. [Methods] Methods (NGC definition paragraph): Causal relationships are defined directly as 'the size of the weight parameters of the first layer' of the predictive model. This equates predictive utility with causality without recovery validation on synthetic data with known ground-truth structure, comparison to linear Granger causality, or checks against confounding/temporal precedence. All subsequent claims (directed effects, performance correlation) inherit this load-bearing issue.
  2. [Results] Results (causal relations and performance correlation): The reported 'significant causal relations' and correlation with batter performance lack stated error bars on the weight magnitudes, baseline comparisons, or correction for multiple comparisons across the 27 joints and two directions. Without these, the directionality and outcome correlation cannot be evaluated as robust.
  3. [Methods] Abstract and Methods: No description is given of how the neural network architecture, hyperparameters, or significance threshold for weights were chosen or validated; these free parameters directly determine which relations are declared 'causal.'
minor comments (2)
  1. [Abstract] The abstract states 'causality from the batter's body to pitcher's body is much lower than the reverse' but provides no quantitative comparison or statistical test for this asymmetry.
  2. [Methods] Notation for the 27 joint velocities (13 pitchers + 14 batters) should be clarified with an explicit mapping or table, as inter-individual inputs are central to the inter-personal claims.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and detailed review. The comments highlight important methodological considerations for our operational use of neural Granger causality. We address each major comment below and indicate revisions planned for the next version of the manuscript.

read point-by-point responses

  1. Referee: [Methods] Methods (NGC definition paragraph): Causal relationships are defined directly as 'the size of the weight parameters of the first layer' of the predictive model. This equates predictive utility with causality without recovery validation on synthetic data with known ground-truth structure, comparison to linear Granger causality, or checks against confounding/temporal precedence. All subsequent claims (directed effects, performance correlation) inherit this load-bearing issue.

    Authors: We agree that the operational definition equates weight magnitude with causal strength without synthetic recovery experiments or direct comparison to linear Granger causality. This is a genuine limitation of the presented work, as the manuscript relies on an XAI proxy rather than a fully validated causal estimator. The identified relations (e.g., pitcher throwing arm to batter hands) are consistent with domain knowledge, but we cannot claim identifiability guarantees. In revision we will add an explicit limitations subsection clarifying the proxy nature of the measure, noting the absence of synthetic validation, and stating that future work should include such checks. We will also report a brief comparison to linear Granger causality on the same data if computationally feasible. revision: partial

  2. Referee: [Results] Results (causal relations and performance correlation): The reported 'significant causal relations' and correlation with batter performance lack stated error bars on the weight magnitudes, baseline comparisons, or correction for multiple comparisons across the 27 joints and two directions. Without these, the directionality and outcome correlation cannot be evaluated as robust.

    Authors: The original manuscript does not include error bars, baseline comparisons, or multiple-comparison correction. We accept that these omissions weaken the ability to assess robustness. In the revised version we will add bootstrap-derived standard errors on the reported weight magnitudes, include a null baseline (e.g., shuffled time series), and apply FDR correction across the 27 joints and both directions. These additions will be placed in the Results and Methods sections. revision: yes

  3. Referee: [Methods] Abstract and Methods: No description is given of how the neural network architecture, hyperparameters, or significance threshold for weights were chosen or validated; these free parameters directly determine which relations are declared 'causal.'

    Authors: We acknowledge that the manuscript omits details on architecture selection, hyperparameter tuning, and threshold justification. In revision we will expand the Methods section to specify the network architecture (layers, units, activation), the cross-validation procedure used for hyperparameter choice, and the rationale and sensitivity analysis for the weight threshold. These additions will improve reproducibility without altering the core analysis. revision: yes

Circularity Check

1 steps flagged

Causality defined as first-layer weight magnitudes in predictive neural model, rendering reported causal relations tautological by construction

specific steps
  1. self definitional [Methods (Abstract)]
    "In the NGC analysis, causal relationships were defined as the size of the weight parameters of the first layer of a machine-learning model trained to predict the future state of a specific time-series variable."

    Causality is defined directly as the magnitude of weights obtained from the predictive model. Consequently, any 'significant causal relations' identified by NGC analysis are identical to the fitted parameters by construction; the subsequent claims about inter-individual directionality and performance correlations reduce to reporting properties of that fit.

full rationale

The paper's central results rest on an explicit operational definition that equates causality with the fitted first-layer weights of a neural network trained to forecast joint velocities. Reported directed effects (e.g., pitcher's throwing arm to batter's hands) and their correlation with performance are therefore direct extractions from the same predictive fit, without recovery validation, comparison to linear Granger causality, or interventional checks. This matches the self-definitional pattern and forces the key claims by the paper's own definition.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim depends on the untested premise that first-layer neural weights equal causal strength, on standard supervised-learning assumptions, and on the availability of the 27-joint velocity dataset. No new physical entities are introduced.

free parameters (2)
  • neural network architecture and hyperparameters
    Number of layers, hidden units, learning rate, and regularization are chosen to produce the weight matrix interpreted as causality; these choices are not enumerated in the abstract.
  • threshold for declaring a weight 'significant'
    The abstract states 'significant causal relations' without specifying the statistical criterion or multiple-testing correction applied to the 27 x 27 weight matrix.
axioms (2)
  • domain assumption Granger causality can be read from the magnitude of first-layer weights in a neural predictor without further validation against interventions or known ground-truth causal graphs.
    Invoked in the Methods paragraph that defines causal relationships via weight size.
  • domain assumption Joint resultant velocity time series are stationary and sufficiently long for reliable neural training.
    Implicit in the generation of input datasets from motion capture.

pith-pipeline@v0.9.0 · 5811 in / 1593 out tokens · 36605 ms · 2026-05-24T04:15:16.624172+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

  1. [1]

    hyperparameters

    Methods This section provides detailed information regarding the methodology of our approach. Note that since there are many “hyperparameters” included in the process of our approach, such as the selection of the feature variables to the network or dawn-sampling rate, we discussed in detail how to tune those parameters and the effect of changing them on t...

    work page 2000

  2. [2]

    velocity interaction

    E. P. Roetert, M. Kovacs, D. Knudson, J. L. Groppel, Biomechanics of the tennis groundstrokes: Implications for strength training. Strength Cond. J. 31 (2009) 41–49. 13. M. A. Riley, M. J. Richardson, K. Shockley, V . C. Ramenzoni, Interpersonal synergies. Front. Psychol. 2 (2011) 38. 14. M. J. Richardson et al., Comparing the attractor strength of intra-...

    work page 2009