KCIN: A Kolmogorov-Arnold Convolutional Network-Based Model for Transportation Mode Recognition With Incomplete Sensor Series

Advances in ubiquitous computing and artificial intelligence have catalyzed the rapid evolution of Human Activity Recognition (HAR). As a branch of HAR, Transportation Mode Detection (TMD) have been widely studied in recent years. However, traditional machine learning-based methods are constrained by the limited expressiveness of handcrafted features. Deep learning-based TMD algorithms have shown significantly outperform by automatically extracting high-level features through extensive supervised learning to enhance the model’s representation ability of complex behaviors. Yet, most existing Deep Neural Network models, inspired by Multilayer Perceptions (MLPs), are limited by the nonlinear expressiveness of MLPs and data integrity, failing to learn correct feature representations when sensor data is incomplete. To address this challenge, we introduce a method named KAN-Conv based Context Interpolation Attention Network (KCIN). KCIN adopts the Kolmogorov-Arnold Convolution Network (KAN-Conv) as its foundational structure and leverages Context Interpolation Attention (CA) to recover representations of missing sensor data. This approach effectively enhances the model’s capability to extract robust motion features from sensor series, particularly in scenarios with data loss. Experiments conducted on three realworld datasets substantiate the efficacy of KCIN across various practical settings. Experiments conducted on three real-world datasets have demonstrated KCIN’s effectiveness, achieving accuracies of ${7 8 . 4 6 \%}, 76.31 \%$, and $81.49 \%$ even with ${9 0 \%}$ missing data.