Understanding how people choose to travel is essential for intelligent transportation planning and related smart services. Recent advances in deep learning, coupled with the increasing market penetration of GPS devices, have paved the way for novel travel mode identification methods based on GPS data mining. While many have shown promising results, most methods have often relied heavily on the few available labeled data, leaving large amounts of unlabeled ones unused. To address this issue, we propose MultiMix, a semi-supervised multi-task learning framework for travel mode identification. Our framework trains a deep autoencoder using batches of labeled, unlabeled, and synthetic data by simultaneously optimizing three corresponding objective functions. We show that MultiMix outperforms several fully- and semi-supervised baselines, achieving a classification accuracy of 66.2% on Geolife using just 1% of labeled data, with accuracy reaching 84.8% when incorporating all available labels. We also verify the necessity of its components through an ablation study designed to provide insights into the proposed approach.