Evaluation of Autoencoder-Based Trajectory Representations for Reliable Mobility Analytics

Reliable mobility analytics require compact trajectory representations and evaluation procedures that limit information leakage and overly optimistic conclusions. This study evaluates unsupervised autoencoder-based trajectory representations across maritime Automatic Identification System (AIS) and human-mobility Global Positioning System (GPS) data using fixed-length position-and-velocity sequences. Normalized raw features, Principal Component Analysis (PCA), denoising autoencoders (DAE), variational autoencoders (VAE), DAE–VAE feature concatenation, and sequence autoencoders are examined with linear and nonlinear downstream probes. For GeoLife, the previously added handcrafted linear baseline, matched-dimensionality controls, and paired user-level bootstrap are retained, while robustness is extended to all five user-disjoint folds. DAE, VAE, Long Short-Term Memory autoencoder (LSTM-AE), and Transformer autoencoder (Transformer-AE) are trained with two encoder seeds per fold, and Random Forest is evaluated with ten classifier seeds per encoder run. After hierarchical aggregation, DAE–VAE obtains the best average rank under both probes. Its Random Forest Macro F1 is 0.5142±0.1219, only slightly above DAE at 0.5121±0.1187, whereas under Logistic Regression DAE–VAE reaches 0.3852±0.0702 and ranks first in all five folds. Friedman tests indicate overall representation differences for Random Forest (p=0.0076) and Logistic Regression (p=0.0023), while exact pairwise tests are interpreted cautiously because only five independent folds are available. On AIS, Transformer-AE attains the strongest behavioural-probe result, but the descriptor-derived pseudo-labels are treated only as an auxiliary non-semantic probe. Overall, the results support compact learned representations and improved linear separability, while showing that architecture, latent dimensionality, downstream model complexity, and evaluation design materially affect the conclusions. In this work, “reliable” refers to leakage-aware, held-out-user, and uncertainty-conscious evaluation rather than to a separate reliability metric.