MaaSformer-MMOE: Multi-task Transformer under mixture-of-experts framework for MaaS bundle customization

Abstract

Mobility-as-a-Service (MaaS) platforms are reshaping urban mobility by integrating multiple travel modes into seamless, user-centric systems. However, designing dynamic MaaS bundles that adapt to user-specific preferences, evolving over time in response to changing travel behaviors and shifting needs, remains a significant challenge. The rise of big data and artificial intelligence (AI) has unlocked new opportunities for data-driven personalized MaaS bundle solutions. In this study, we introduce an innovative MaaSformer-MMoE framework to customize user-specific monthly MaaS bundles by predicting each user’s mode-specific usage frequency class (classification tasks) and travel fare (regression task) for the upcoming month based on the user’s previous travel records. Within the multi-gate mixture-of-expert (MMoE) framework, each expert network is a MaaSformer, and each gate determines the weighted contributions of expert outputs relevant to a specific task tower. MaaSformer integrates two key modules: 1) Multi-mode Transformer processes continuous time-series features (e.g., monthly travel time, distance, and fare) employing a multi-feature self-attention mechanism; 2) OD Transformer processes origin-destination (OD)-specific travel features (i.e., journey frequency) using a multi-OD self-attention mechanism. Evaluated on a multimodal (i.e., bus, rail, ferry, and tram) dataset of over 1.5822 million users in Queensland, Australia, from 01/2021 to 01/2023, the proposed MaaSformer-MMoE demonstrates state-of-the-art performance in predicting mode usage frequency class and travel fare compared with 9 baseline models, significantly improving user satisfaction, adoption and retention for MaaS platforms.