Large language models (LLMs) augmented with knowledge graphs (KGs) have exhibited great potential for complex reasoning tasks. However, existing approaches often struggle with incomplete subgraph retrieval and inaccurate semantic alignment, which hinder reasoning performance and answer quality. In this paper, we present Faico, a KG-enhanced reasoning framework designed to achieve both semantic faithfulness and structural completeness. Faico decouples model inference from graph traversal by integrating a fine-tuned LLM-based relation type generator for accurate semantic mapping and a KG retriever for reasoning subgraph search. Based on the predicted relation types, we model the reasoning subgraph (RS) as a k-bounded edge type (k-BET) subgraph, where k constrains the recurrence of relation types within paths, and devise a budget-dominance-based algorithm to efficiently identify the maximal k-BET subgraph. Our framework ensures comprehensive coverage of relevant multi-hop relations while reducing computational overhead. Through extensive experiments on multiple KGQA benchmarks, Faico demonstrates improvements in both effectiveness and efficiency over LLM-native and state-of-the-art KG-augmented reasoning baselines, delivering more accurate, complete answers and lower inference latency.

Federated recommender systems (FedRec) have emerged as a promising approach to provide personalized recommendations while protecting user privacy. However, recent studies have demonstrated their vulnerability to poisoning attacks, wherein malicious clients can inject carefully crafted gradients to prompt target items to benign users. Existing attacks typically target the full user group, which compromises stealth and increases the risk of detection. In contrast, real-world adversaries may prefer to target specific user subgroup, such as promoting health supplements to older individual, to maximize attack success while preserving stealth to evade detection. Motivated by this gap, we introduce Spattack, the first poisoning attack designed to manipulate recommendations for specific user subgroups in federated setting. Specifically, Spattack adopts an approximate-and-promote paradigm, which first approximate user embeddings of target/non-target subgroups and then prompts target items to the target subgroups. We further reveal a trade-off in achieving strong attack performance on the target group while keeping the non-target group largely unaffected. To achieve a better trade-off, we propose enhanced approximation and promotion strategies. For the approximation, we push the embeddings of different subgroup away based on contrastive learning and augment the target group's relevant item set via clustering. For the promotion, we align target and relevant item embeddings to strengthen their semantic connections. An adaptive weighting strategy is further proposed to balance promotion effects between target and non-target subgroups. Experiments on three real-world datasets demonstrate that Spattack consistently achieves strong attack performance on the target subgroup with minimal impact on non-target users, even when only 0.1% of users are malicious. Moreover, Spattack maintains competitive recommendation performance and exhibits strong resilience against mainstream defenses.

Although a large amount of valuable knowledge can be obtained from the weighted graph snapshots modeled over time, it may cause privacy issues. Local differential privacy (LDP) provides a strong solution for private graph data publishing in decentralized networks. However, most existing LDP studies over graphs are only applicable to static unweighted graphs. This paper investigates the problem of continuous publication of weighted graph snapshots and proposes a graph publication framework, WGT-LDP, under w -event edge weight LDP, which can protect the privacy of edges and weights over any w consecutive time steps. WGT-LDP consists of four key components: population division-based sampling that overcomes the problem of over-segmentation of the privacy budget, data range estimation that mitigates noise on edge weights, aggregate information collection that obtains important information about the graph structure and edge weights, and graph snapshot generation that reconstructs weighted graph snapshot at each time step. We provide theoretical guarantees on privacy and utility, and perform extensive experiments on three real-world and two synthetic datasets, using four commonly used metrics. Our experiments show that WGT-LDP produces high-quality synthetic weighted graphs and significantly outperforms baseline methods.

Federated learning (FL) has become an emerging paradigm via cooperative training models among distributed clients without leaking data privacy. The performance degradation of F1 on heterogeneous data has driven the development of personalized FL (PFL) solutions, where different models are built for individual clients. However, existing PFL approaches often have limited personalization in terms of modeling capability and training strategy. In this paper, we propose a novel PFL solution, Feed, that employs an enhanced shared-private model architecture and equips with a hybrid federated training strategy. Specifically, to model heterogeneous data for different clients, we design an ensemble-based shared encoder that generates an ensemble of embeddings, and a private decoder that adaptively aggregates these embeddings for personalized prediction. In addition, we propose a server-side hybrid federated aggregation strategy to enable effective training of the heterogeneous shared-private model. To prevent personalization degradation in local model updates, we further optimize the personalized local training on the client-side by smoothing the historical encoders. Extensive experiments on MNIST/FEMNIST, CIFARIO/CIFARIOO, and YELP datasets demonstrate that Feed consistently outperforms state-of-the-art approaches.

Distributed key-value storage and computation are essential components of cloud services. As the demand for high-performance systems has increased significantly, a new architecture has been motivated to separate computing and storage nodes and connect them using RDMA-enabled networks. Existing RDMA-enabled systems use client-side cached indexes to reduce communication overhead and improve performance. However, such approaches could result in high server CPU contention due to heavy dynamic workloads (i.e.,inserts), and cause a large accuracy gap because of the different indexes between client-side and server-side. These drawbacks limit the performance of RDMA-enabled systems. In this paper, to deal with these issues, we introduce AStore to achieve high performance with low memory footprint. AStore employs a new uniformed architecture, utilizing an adaptive learned index as both the server-side learned index and the client-side cached index, to handle dynamic and static workloads. We propose several optimization techniques to optimize dynamic and static workload procedures and design the leaf node lock mechanism to support high concurrent access. Extensive evaluations on YCSB, LGN, and OSM datasets demonstrate that AStore achieves competitive performance on read-only workloads by up to 75.2%, 107.3% and 57.7%, as well as improving performance on write-read workloads by up to 65.7%, 108.7% and 74.3% than XStore.

Recommender systems across many Internet services have become a critical part of online businesses, as consumers would refer to them before making decisions. However, the lack of explicit ratings for items on many services makes it challenging to capture user preferences and item characteristics. Both academia and the industry have drawn attention to rating predications as a fundamental problem in recommendation systems. With the emergence of social networks, social recommender systems have been proposed to utilize the relationship between users and items to alleviate the data sparsity problem for rating predictions. However, they either concentrate on the opinion mining for each user and item, or consider the connections between users only. In this paper, we present an effective framework, Triple-hierarchical Attention Graph-based social rating prediction (TAG), to exploit the social relationships between users, the user-item interest relationships, the correlation relationships between items, and reviews for rating predictions. In order to consider opinions from reviews and these complex relationships, we first employ two triple-hierarchical attention to extract user and item features from reviews. We then design an inductive GNN, which generates effective embedding for users and items. Experiments over Yelp show that TAG outperforms state-of-the-art methods across RMSE, MAE, and NDCG metrics.