<p>Time series classification (TSC), essential for capturing temporal patterns in domains from healthcare to industrial monitoring, faces significant challenges under label scarcity. While deep semi-supervised methods mitigate annotation burdens, they remain vulnerable to epistemic uncertainty arising from insufficient evidence in low-label regimes. This data paucity exacerbates model overconfidence in spurious correlations, ultimately degrading generalization. To address this, we propose Sequence Neural Process with Multi-view Posterior Consistency (SNPMPC), a novel uncertainty-aware semi-supervised TSC framework based on Neural Processes (NPs). Motivated by the capacity of NPs for principled uncertainty quantification, SNPMPC fundamentally reformulates their optimization paradigm for sequence reasoning. It replaces the single global conditional distribution modeling with a sequence of conditional distributions, which dynamically captures evolving dependencies while inherently modeling epistemic uncertainty. Then, we introduce multi-view posterior alignment regularization to modify the distribution divergence regularization in the optimization paradigm of standard NPs that enforces distribution alignment between weak/strong augmented views of unlabeled data and the supervisory manifold of labeled instances, which injects richer signals to elevate latent variable quality. Extensive experiments demonstrate that the proposed approach can simultaneously quantify the epistemic uncertainty and significantly advance state-of-the-art classification accuracy.</p>