Trajectory-user linking via Complexed-Valued Multi-Layer Perception with Fast Fourier Transform

The Trajectory-User Linking (TUL) task has gained increasing importance as it enables providers to establish connections between diverse online identifiers, ultimately enhancing user profiling and overall experience. Despite the increasing importance of this task, traditional machine learning methods based on statistical analysis lack the capability to capture the inherent interactions within time series and analyze their superior characteristics of a global view and energy compaction. Meanwhile, prevailing deep learning methodologies often tend to be intricate, imposing substantial computational burdens during model training. To address this challenge, the Complex-Valued Multi-Layer Perception with Fast Fourier Transform (CVMLP-FFT) is proposed, which extends traditional MLP to complex space to form complex-valued MLP, and designs two transformation modules on the channel and temporal features. The Channel Transformation Module facilitates communication across distinct channels, functioning on each timestamp by employing shared weights among segments to discern channel dependencies. Simultaneously, the Temporal Transformation Module focuses on learning temporal patterns in the frequency domain. Constructed with shared weights between channels, it also extracts temporal features. These temporal features are subsequently fed into a Softmax module for user identification, thereby recognizing and generating trajectories associated with specific users. Compared with baselines in a real-world user trajectory dataset, the proposed method achieves a notable 5.49% improvement in ACC. The CVMLP-FFT significantly reduces the computational burden, providing a more efficient solution for Trajectory-User Linking classification tasks.