Ovarian cancer (OC) is the most lethal malignancy of the female reproductive system, and its accurate subtype classification is crucial for individualized treatment. However, high tumor heterogeneity and morphological similarity make this task extremely challenging. Whole Slide Image (WSI) classification faces two main bottlenecks: the ultra-high resolution and tissue complexity result in high annotation cost and label scarcity; and the inherent Multiple Instance Learning (MIL) nature of WSIs means the diagnosis depends on unannotated local regions. To reduce annotation demand and improve classification performance, we propose a novel Semi-Supervised Multiple Instance Learning framework with a Laplacian Rank Constraint (SSL-DEPL).The proposed SSL-DEPL adopts a dual-network, multi-view architecture that integrates a semi-supervised learning approach with an interactive supervision mechanism for network co-training. The method optimizes the embedded feature space through a Laplacian rank constraint to preserve intrinsic data geometry. Finally, a constrained clustering strategy dynamically selects high-quality pseudo-labeled samples to update the labeled sample set. Tested on the Ovarian Cancer dataset from Yunnan Cancer Hospital (YNC-OC), SSL-DEPL achieved a 92% bag-level accuracy using only 0.011% of the total data as labeled information, and significantly outperformed various existing advanced and classical methods across precision and other metrics. SSL-DEPL effectively uses limited annotations to achieve precise OC subtype classification, and its Laplacian constraint coupled with the dual-network interactive supervision mechanism provides a novel direction for low-annotation learning in highly heterogeneous medical images.