Sequential recommendation often suffers from sparse ID-based interactions. Although item texts provide complementary semantic signals, existing textual representations either suffer from excessive similarity or weaken preference-related correlations when improving discriminability. To address this problem, we propose LoTSARec, a Low-rank Tensorized Semantic Attention model for sequential recommendation. LoTSARec first introduces a Low-rank Tensorized (LoT) transformation to obtain compact and discriminative textual representations, aiming to reduce redundant textual similarity while preserving preference-related semantic structures. Then, a Semantic-guided Sequential Attention (SA) mechanism uses transformed textual representations to guide attention weight estimation, while retaining ID-based representations to preserve collaborative sequential signals. We further employ contrastive learning objectives to enhance ID-text alignment and target-aware preference modeling. Experiments on public datasets show that LoTSARec improves Hit Rate and NDCG by 6.75% and 5.8% on average over competitive baseline method, while further analyses verify its key components.

Wheat unsound kernel identification is challenged by severe class imbalance, where normal kernels account for most samples while unsound kernels are underrepresented. To improve recognition under such conditions, we propose a Heterogeneous Integration and Decoupled Expert Architecture (HIDEA). The proposed model consists of four components: a feature refinement module that enhances intermediate features through strided convolution and CBAM, a heterogeneous integration module that combines refined stage-wise features with expert-specific representations, an imbalance-aware optimization module based on Focal Loss, and a decoupled classification module for improving discrimination of unsound kernel classes. Experiments conducted on the public GrainSpace dataset demonstrate that HIDEA achieves superior performance over existing state-of-the-art methods in both Top-1 Accuracy and Macro-F1 under imbalanced data distributions.