Yifan Ji, Zhipeng Xu, Zhenghao Liu, Zulong Chen, Qian Zhang, Zhibo Yang, Junyang Lin, Yu Gu, Ge Yu, Maosong Sun. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Knowledge graphs (KGs) are typically used to enhance the performance of recommender systems by leveraging their characteristic to supplement the sparse user-item interaction data in the latter. This is because KGs feature abundant semantic information about entities (e.g. users or items in recommender systems) and inter-entity relationships. Deep learning has been seen gain wide use in boosting the performance of recommender systems. Thanks to focusing merely on nodes in a KG, or entities, such an approach, however, suffers from well capturing the semantics hidden in the node connectivity when learning node embeddings, leading to undesirable discovery of potential semantic associations between nodes. Against this background, we propose for recommender systems a multi-level cross-view contrastive learning mechanism that learns high-quality feature representations from unlabeled data. By incorporating item-item correlations into item embeddings, we developed an interitem semantic association graph (SAG), which was intended to model more comprehensive and finer-grained inter-item semantic associations. A series of experiments have been carried out and the findings verify that the proposed mechanism outperforms other baseline models on all metrics, confirming the effectiveness of coupling SAGs with contrastive learning in enhancing the performance of recommender systems.

MOTIVATION: Tumor and adjacent normal RNA samples are commonly used to screen differentially expressed genes between normal and tumor samples or among tumor subtypes. Such paired-sample design could avoid numerous confounders in differential expression (DE) analysis, but the cellular contamination of tumor samples can be an important noise and confounding factor, which can both inflate false-positive rate and deflate true-positive rate. The existing DE tools that use next-generation RNA-seq data either do not account for cellular contamination or are computationally extensive with increasingly large sample size. RESULTS: A novel linear model was proposed to avoid the problem that could arise from tumor-normal correlation for paired samples. A statistically robust and computationally very fast DE analysis procedure, contamDE-lm, was developed based on the novel model to account for cellular contamination, boosting DE analysis power through the reduction in individual residual variances using gene-wise information. The desired advantages of contamDE-lm over some state-of-the-art methods (limma and DESeq2) were evaluated through the applications to simulated data, TCGA database and Gene Expression Omnibus (GEO) database. AVAILABILITY AND IMPLEMENTATION: The proposed method contamDE-lm was implemented in an updated R package contamDE (version 2.0), which is freely available at https://github.com/zhanghfd/contamDE. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.