Medical studies have found that Tumor Mutation Burden (TMB) is positively correlated with the efficacy of immunotherapy for Non-Small Cell Lung Cancer (NSCLC), and TMB value can predict the efficacy of targeted therapy and chemotherapy. However, the calculation of TMB value mainly depends on the Whole Exon Sequencing (WES) technology, which is usually expensive and takes too much time. To deal with this problem, this paper explores the connection between TMB and digital pathology images, predicting TMB using common clinical digital slices. We proposes RCA-MSAG, a deep learning model based on the Residual Coordinate Attention (RCA) structure and Multi-Scale Attention Guidance (MSAG) module, enhancing TMB prediction accuracy. The model takes ResNet-50, integrating Coordinate Attention (CA) and MSAG to extract critical information from lung cancer pathology. Using The Cancer Genome Atlas (TCGA) dataset, the model achieves 96.2% accuracy, 96.4% precision, 96.2% recall, and a 96.3% F1 score, outperforming mainstream models. This model shows promise in aiding clinical diagnoses and guiding TMB predictions.

Graph convolutional network (GCN) has rapidly developed in various fields due to its powerful modeling capability. However, most of the researches directly inherit the complex design of GCN, such as feature transformation and nonlinear activation, which lacks thorough ablation analysis on GCN. In addition, implicit feedback is not fully utilized and data sparsity is not well resolved, which are also shortcomings of current recommendation algorithms. To solve the above problems, this paper proposes a lightweight graph convolutional collaborative filtering (F-LightGCCF) recommendation approach incorporating social relationships. Firstly, it abandons the design of feature transformation and nonlinear activation in graph convolutional models and simplifies model training. Additionally, a series of intermediate feedback from users’ implicit negative feedback is generated by taking advantage of social networks, which improves the utilization of implicit negative feedback. Secondly, it can model the long-range dependencies between users and items by using the dual attention mechanism, aggregating the contribution values of neighboring nodes and the importance of the learning vectors in each layer of the graph convolution layer respectively. Lastly, the inner product operation is used to obtain the association score between users and items. Extensive experiment results on two real-world datasets show that F-LightGCCF outperforms existing state-of-the-art recommendation methods. Further ablation studies and analyses validate the efficiency and effectiveness of the F-LightGCCF model.

Abstract Personalized news recommendation is an important technology to help users obtain news information they are interested in and alleviate information overload. In recent years, news recommendation has been increasingly widely studied and has achieved remarkable success in improving the news reading experience of users. In this paper, we provide a comprehensive overview of personalized news recommendation approaches. Firstly, we introduce personalized news recommendation systems according to different needs and analyze the characteristics. And then, a three-part research framework on personalized news recommendation is put forward. Based on the framework, the knowledge and methods involved in each part are analyzed in detail, including news datasets and processing techniques, prediction models, news ranking and display. On this basis, we focus on news recommendation methods based on different types of graph structure learning, including user–news interaction graph, knowledge graph and social relationship graph. Lastly, the challenges of the current news recommendation are analyzed and the prospect of the future research direction is presented.

Diversity point-of-interest recommendation system benefits users to broaden their interests, access and discover new interest points. This paper describes a Diversified and Personalized Point-Of-Interest Recommendation System (DP-POIRS) by leveraging the geo-social relationships between POIs. The system consists of three components. The first component - geo-social distance measuring component is used to construct a correlation matrix to describe the geo-social distance between points of interests. The second component -point-of-interest partition component, divides the interest points into diverse clusters by using the spectral clustering algorithm over the correlation matrix. The third component -personalized sorting component, finds out the user's favorite interest points from each cluster, and then sorts them into a list of recommendation by the use of matrix factorization algorithms.

Spatial database queries are often exploratory. The users often find that their queries return too many answers and many of them may be irrelevant. Based on the coupling relationships between spatial objects, this paper proposes a novel categorization approach which consists of two steps. The first step analyzes the spatial object coupling relationship by considering the location proximity and semantic similarity between spatial objects, and then a set of clusters over the spatial objects can be generated, where each cluster represents one type of user need. When a user issues a spatial query, the second step presents to the user a category tree which is generated by using modified C4.5 decision tree algorithm over the clusters such that the user can easily select the subset of query results matching his/her needs by exploring the labels assigned on intermediate nodes of the tree. The experiments demonstrate that our spatial object clustering method can efficiently capture both the semantic and location correlations between spatial objects. The effectiveness and efficiency of the categorization algorithm is also demonstrated.