| dc.description.abstract | Trajectory-user linking (TUL) is a critical problem in spatio-temporal data mining
that involves associating anonymous trajectories with the users who generated them. In
this thesis, we address the TUL challenge by proposing a Transformer-based model aug-
mented with Rotary Positional Embedding (RoPE) encoder and enriched with semantic
context through Points of Interest (POI) features. Using the GeoLife dataset as the ex-
perimental platform, we conduct extensive evaluations on user subsets with trajectory
lengths constrained to a maximum of 1000 points.
Our findings demonstrate that the incorporation of POI features substantially en-
hances the model’s performance, with the full model achieving an Accuracy@1 of 68.33%
and a Macro F1 score of 51.34%. Ablation studies reveal that removing either the Trans-
former structure or the RoPE encoder significantly degrades performance, highlighting
the importance of both sequential modeling and positional encoding. Furthermore, fea-
ture importance analysis shows that the absence of POI information results in a dramatic
collapse of predictive accuracy and generalization, reaffirming the necessity of contextual
features for effective user discrimination.
While the model achieves strong results, the study also acknowledges limitations, in-
cluding reliance on static spatial features and a relatively narrow evaluation scope. Ethical
considerations regarding user privacy are discussed, emphasizing the need for responsi-
ble data handling and privacy-preserving extensions. Overall, this thesis advances the
trajectory-user linking field by demonstrating the synergistic role of structural modeling
and semantic enrichment, while paving the way for future research in robust, scalable,
and ethically aligned mobility analytics. | en_US |
