Multipath transport protocols including multipath TCP (MPTCP) and multipath QUIC (MPQUIC) are designed to utilize multiple network paths for simultaneous data transfer. These protocols try to improve network performance and offer better resilience in dynamic network environments. Nonethe-less, the actual performance improvement is heavily reliant on the effectiveness of the multipath scheduling algorithms. In specific scenarios such as adaptive video streaming, most existing solutions feature two separate and independent control loops for multipath scheduling and video bitrate adaptation, while multipath scheduling algorithms are usually transparent to the video bitrate adaptation process. Lacking the context of inter-path differences and intra-path fluctuations for both network throughput and latency may potentially result in a suboptimal quality of experience (QoE) for video streaming. Such circumstances may lead to a reduced video bitrate, increased latency, and a greater number of rebuffering events. In this paper, we present a QoE-driven joint decision-making framework based on contextual multi-armed bandit (CMAB) algorithms to efficiently address multipath adaptive video streaming problems. This approach merges application-layer (playback buffer ratio) and network-layer (throughput and latency) metrics to create a context-aware online learning model, which can adaptively select the ideal network path and bitrate for multipath adaptive video streaming. Both network emulation and real-world experiments demonstrate that the proposed algorithm delivers better QoE, including higher average video bitrate and fewer rebuffering events when compared to independent decision-making algorithms.