Large language models (LLMs) trained on extensive web and dialogue datasets risk inadvertently memorizing sensitive personal or copyrighted information. Regulatory frameworks, such as the EU’s General Data Protection Regulation (GDPR) “right to be forgotten” and the California Consumer Privacy Act (CCPA) mandate the immediate removal of such data, yet retraining the entire model incurs prohibitive computational and financial costs. Conventional gradient ascent-based unlearning methods apply uniform gradient updates across entire input sequences, significantly degrading overall language capabilities. In this study, we propose an unlearning framework that leverages a sparse autoencoder to selectively identify and remove only the most semantically critical tokens. On the TOFU $1 \%$ forget set, our method achieves a Forget Quality of 0.99 and a Model Utility of 0.61, improving performance by over $50 \%$ and $9 \%$, respectively, compared to standard approaches. This method minimizes unnecessary updates to preserve overall performance and enable more robust and efficient unlearning.

Graph neural networks (GNNs) have recently gained significant attention for their ability to model and analyze complex relationships, leading to numerous applications in a variety of fields. In line with this, explainable models such as GNNExplainer have been developed to provide insights into how GNN models make decisions. However, much of the research has focused on identifying important edges and subgraphs, with relatively less emphasis on feature importance. To address this gap, we propose a novel unsupervised feature scoring method aimed at extracting features that are well-aligned with graph structures. Our approach leverages the distinction between features that explain the overall structural characteristics of the graph and those that do not. Through comparative experiments with other algorithms, we demonstrated that our proposed method is effective in selecting important features in unsupervised settings. Additionally, empirical analysis of gene expression data confirmed that our method could aid in biomarker discovery, highlighting its potential for practical applications in biological studies.

MOTIVATION: To better understand the molecular features of cancers, a comprehensive analysis using multi-omics data has been conducted. In addition, a pathway activity inference method has been developed to facilitate the integrative effects of multiple genes. In this respect, we have recently proposed a novel integrative pathway activity inference approach, iDRW and demonstrated the effectiveness of the method with respect to dichotomizing two survival groups. However, there were several limitations, such as a lack of generality. In this study, we designed a directed gene-gene graph using pathway information by assigning interactions between genes in multiple layers of networks. RESULTS: As a proof-of-concept study, it was evaluated using three genomic profiles of urologic cancer patients. The proposed integrative approach achieved improved outcome prediction performances compared with a single genomic profile alone and other existing pathway activity inference methods. The integrative approach also identified common/cancer-specific candidate driver pathways as predictive prognostic features in urologic cancers. Furthermore, it provides better biological insights into the prioritized pathways and genes in an integrated view using a multi-layered gene-gene network. Our framework is not specifically designed for urologic cancers and can be generally applicable for various datasets. AVAILABILITY AND IMPLEMENTATION: iDRW is implemented as the R software package. The source codes are available at https://github.com/sykim122/iDRW. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Abstract Background Introducing deep learning approach to medical images has rendered a large amount of un-decoded information into usage in clinical research. But mostly, it has been focusing on the performance of the prediction modeling for disease-related entity, but not on the clinical implication of the feature itself. Here we analyzed liver imaging features of abdominal CT images collected from 2019 patients with stage I – III colorectal cancer (CRC) using convolutional neural network (CNN) to elucidate its clinical implication in oncological perspectives. Results CNN generated imaging features from the liver parenchyma. Dimension reduction was done for the features by principal component analysis. We designed multiple prediction models for 5-year metachronous liver metastasis (5YLM) using combinations of clinical variables (age, sex, T stage, N stage) and top principal components (PCs), with logistic regression classification. The model using “1 st PC (PC1) + clinical information” had the highest performance (mean AUC = 0.747) to predict 5YLM, compared to the model with clinical features alone (mean AUC = 0.709). The PC1 was independently associated with 5YLM in multivariate analysis (beta = − 3.831, P < 0.001). For the 5-year mortality rate, PC1 did not contribute to an improvement to the model with clinical features alone. For the PC1, Kaplan-Meier plots showed a significant difference between PC1 low vs. high group. The 5YLM-free survival of low PC1 was 89.6% and the high PC1 was 95.9%. In addition, PC1 had a significant correlation with sex, body mass index, alcohol consumption, and fatty liver status. Conclusion The imaging features combined with clinical information improved the performance compared to the standardized prediction model using only clinical information. The liver imaging features generated by CNN may have the potential to predict liver metastasis. These results suggest that even though there were no liver metastasis during the primary colectomy, the features of liver imaging can impose characteristics that could be predictive for metachronous liver metastasis.

The multiple types of social media data have abundant information, but learning multi-modal social data is challenging due to data heterogeneity and noise in usergenerated data. To address this problem, we propose a multiview network-based clustering approach that is robust to noise and fully reflects the underlying structure of the comprehensive network. To demonstrate the proposed approach, we experimented with clustering challenging tagged images of landmarks. The results show that the proposed method outperforms other previously reported multi-view clustering algorithms and better utilizes the advantages of the network for each view. Furthermore, the tagged-image network constructed by the proposed method and the clustering results are extensively analyzed.

The continuing growth of information overflow has made it hard to obtain valuable information on the web. In this trend, the need for effective Information Retrieval (IR) technique has been increased. Although document data contain much more abundant information, users can retrieve necessary information only from the title and description in conventional web services. In order to meet the demands for fast and accurate retrieval of valuable information, we propose a fast and effective content-based document information retrieval system that retrieves the information from the actual content of a document. The proposed method is based on a topic model of Latent Dirichlet Allocation that is used to extract major keywords for a given document. The main contributions of our system are the increased flexibility, effectiveness, and fast retrieval of information. Our system can easily communicate with existing web service through the standard JSON format. In addition, we increase the speed of information retrieval by using NoSQL based database system with inverted indexing and B-tree based indexing. We validate the performance of our system on real data collected from the SlideShare service. The proposed system shows better retrieval performance over the existing IR system.

Image spams have been annoying users everywhere and it has also been increasingly appearing in mobile phones these days. In accordance with more sophisticated spam filtering system, spams are being more intelligent and have caused severe social problems. However, there has not been effective solution for detecting mobile phone spam images yet. Due to the insufficient spam image data in mobile phones, training the predictive model is quite hard. To resolve this issue, we recently proposed a phone spam image filtering system using e-mail spam images and showed that using e-mail spam data is fairly meaningful in improving the performance of phone spam image detection. In this paper, we further investigate the effectiveness of utilizing the graph structure in e-mail spam data. Furthermore, the classification performance behavior depending on different image descriptors of Pyramid Histogram of Visual Words (PHOW) and RGB histogram is explored extensively.