Personalized Federated Learning (PFL) enables the training of customized deep models on decentralized, heterogeneous data while preserving privacy. However, existing PFL methods primarily optimize the final layer, overlooking intermediate layers, which degrades backbone training, especially in non-IID settings. In this work, we propose FedAIMS (Federated Adaptive Intermediate Supervision), a novel PFL framework that incorporates intermediate supervision to enhance model training. FedAIMS adopts prototype-based feature alignment to provide effective intermediate supervision and adaptive supervision sampling to reduce computational overhead for resource-limited clients. Experiments on diverse datasets show that FedAIMS outperforms state-of-the-art PFL baselines by up to 36.76% in accuracy.

Graph models based on large language models (LLMs) have recently garnered considerable attention due to its significant success. Although existing methods resort to LLMs to learn unified semantic representations across domains, they disregard the unique structural characteristics of graphs from different domains. To address this problem, in this paper, we boost graph models from structural perspective and propose BooG. The model constructs virtual super nodes to unify structural characteristics of graph data from different domains. Specifically, the super nodes fuse the information of anchor nodes and class labels, where each anchor node captures the information of a node or a graph instance to be classified. Instead of using the raw graph structure, the super nodes, along with virtual edges, establish a standardized aggregation mechanism that fuses rich information from neighborhoods and associated class labels, accommodating graph structural characteristics inherent to different domains. Additionally, we propose a novel pre-training objective based on contrastive learning, which learns more expressive representations for graph data and generalizes effectively to different domains and downstream tasks. Experimental results on various datasets and tasks demonstrate the superior performance of BooG. We provide our code and data here at the website of github.com/cy623/BooG.

In crowdsourcing scenarios, we can obtain each instance’s multiple noisy label set from crowd workers and then infer its unknown true label via label integration. Recent studies show that label integration performs well when the label quality of most workers is high, but seldom considers the crowdsourcing scenario in which the label quality of most workers is low. In this work, we argue that the label quality of most workers is low while the label quality of a few workers is high, label integration can also perform well. Based on this premise, we propose a novel label integration algorithm called random forest-based weighted majority voting (RFWMV). RFWMV uses a random forest to learn multiple labeling rules for each worker and uses the consistency of labeling rules to evaluate the label quality of each worker. Specifically, RFWMV first respectively trains a random forest on the instances labeled by each worker. Then, RFWMV estimates the label quality of each worker based on the outputs of the corresponding random forest’s base classifiers. Finally, RFWMV infers integrated labels of instances by the weighted majority voting based on each worker’s label quality and its corresponding random forest’s output. The extensive experiments show that RFWMV significantly outperforms all the other state-of-the-art label integration algorithms.

Knowledge graph (KG), enriched with abundant auxiliary information, plays an increasingly important role in enhancing recommendation performance while simultaneously improving the system’s explainability. Recently, the development of end-to-end models leveraging graph neural networks (GNNs) has emerged as a leading trend in knowledge-aware recommendation systems. However, the noise data in dataset significantly impact the performance of recommendation systems, often leading to misleading associations between knowledge facts and user behaviors. In this paper, we propose a novel approach that leverages both causal inference and denoising techniques to address the challenge. Specifically, our approach uses two counterfactual generators, built with reinforcement learning, to create counterfactual interactions. The recommender is implemented using two distinct graph neural networks to aggregate information from the Knowledge Graph and user-item interactions, respectively. Additionally, we combine relation-denoising module that filtering out irrelevant relations for both the original and generated interaction graphs. With the help of counterfactual generators and denoising module, the recommender could identify potential spurious correlations and reduce the influence of noise. We conducted extensive experiments on three benchmark datasets, and the results demonstrate the effectiveness of our approach compared to state-of-the-art knowledge-aware methods.