Design and Visualization Notes

Why these graphs exist and why they were chosen. This page explains design intent, not interpretation rules.

Why KDE Instead of Histograms

Histogram bins impose hard boundaries that can distort GPX sampling structure. KDE gives a continuous density estimate where each observation contributes smoothly to nearby regions.

• continuous distribution shape (no bin-edge jumps)
• local influence from each point
• better visibility of mixed sampling regimes

Why Gaussian Kernels and Log-Scaled Axes

Time and distance deltas are positive and usually span multiple orders of magnitude. Linear spacing compresses dense near-origin structure and over-emphasizes long-tail outliers.

• KDE is computed in log space for better near-origin resolution
• tails remain visible without clipping
• values are still shown in linear units; only axis spacing is logarithmic

Why Bandwidth Is User-Controlled

There is no single correct smoothing level for raw GPX streams. Different analysis goals require different bandwidths.

• higher bandwidth for broad regime trends
• lower bandwidth for sharp fixed-interval peaks
• manual control keeps the smoothing tradeoff explicit

Why These Plot Types

Each view answers a distinct question.

• Time/Distance KDE: dominant scales and multimodality
• Rug support: where observations exist along the axis
• Time-Distance scatter: joint behavior and isolated outliers
• Sequence-aware scatter: local continuity by GPX index

Why Sequence-Aware Scatter Exists

Global scatter treats points as exchangeable and hides local temporal structure. Sequence-aware plotting preserves order and makes local behavior legible.

• distinguishes true pauses from sparse logging gaps
• shows stationary logging under fixed cadence
• exposes time-local transitions in recording behavior

Coverage caveat: Sequence-aware scatter is rendered from valid joint pairs only (where both time and distance deltas are usable). If the plot appears sparse or visually inconsistent, cross-check Track Details / Deep Inspector for anomalies and excluded points.

What These Plots Do Not Do

• do not prove GPX correctness
• do not score activity quality
• do not classify behavior categories
• do not infer causality

The visualization layer is observation-first: it surfaces structure and leaves interpretation to downstream analysis.