Accurate survival prognosis is essential for personalized prognostic assessment and treatment planning in lung adenocarcinoma. This study compares non-parametric and semi-parametric statistical methods for survival prognostic modeling using clinical data from the TCGA-LUAD cohort(n=493). The Kaplan-Meier estimator and the multivariate Cox proportional hazards model were applied to evaluate the prognostic roles of tumor stage, age, and gender. The dataset was divided into a training set (70%) for model fitting and a testing set (30%) for independent validation. The results show that the Kaplan-Meier estimator provides an intuitive visualization of survival differences across tumor stages, with Log-rank tests confirming significant differences among subgroups(p<0.001). The Cox model identified tumor stage as the dominant independent prognostic factor. Compared with Stage I patients, Stage IV patients had a 3.58-fold higher hazard of death(HR = 3.58, 95% CI: 1.67–7.69, p<0.005). Although the C-index increased only slightly from 0.686 to 0.689, the Cox model offered added value through multivariate adjustment and the estimation of interpretable hazard ratios. These findings suggest that Kaplan-Meier estimation and Cox regression play complementary roles in lung cancer survival analysis.

A recommendation system is an important foundation for e-commerce platforms to deliver personalized services. The introduction of large language models (LLMs) has expanded recommendation systems from behavioral matching to semantic reasoning, user modeling, and personalized content generation. This paper argues that this transformation is restructuring the privacy risk landscape within recommendation systems: privacy risks have extended beyond data collection and storage security to encompass inferential privacy risks (where platforms infer sensitive user attributes from ordinary shopping behaviors) and behavioral influence risks (where personalized recommendation outputs reshape users’ decision-making environments). Grounded in Nissenbaum’s contextual integrity theory and Crawford and Schultz’s predictive privacy harms framework, this paper proposes a three-stage privacy risk model of "data collection—attribute inference—behavioral influence," and examines the specific manifestations of this risk structure in shopping recommendation scenarios through conceptual analysis and case analysis. The paper further argues that while technical protection mechanisms such as federated learning and differential privacy can reduce data leakage risks, they struggle to address privacy issues at the inference and output stages. Therefore, a comprehensive governance framework needs to be established across three dimensions: technical protection, platform accountability, and user control rights.

Predicting short-horizon volatility is a difficult machine-learning problem in finance because the data are noisy, temporally dependent, and sensitive to leakage. This paper asks whether simple disclosure-based signals improve next-day volatility forecasting under a leakage-safe walk-forward design and against the strong benchmarks. Using a daily panel of the 800 A-share stocks from 2020 to 2025, next-day absolute returns are predicted from price-based, disclosure-based, and volatility-history features. The empirical framework combines the expanding-window walk-forward testing, train-only thresholding, and comparisons across naive different baselines including Ridge regression, and XGBoost. The results show that the volatility-history benchmarks explain most forecastable variation, so disclosure variables deliver only modest average gains once leakage is controlled for. Their predictive value is more visible in rare high-attention states with heterogeneous same-day disclosures, but the evidence remains statistically fragile because such states are sparse. Overall, disclosure signals appear more useful for conditional than for broad unconditional volatility forecasting.

Multi-agent systems(MASs) are more and more significant for solving complex real problems that demanded coordinated decision and interaction in dynamic environments. However, traditional ways, especially in reinforcement learning(RL), are not competent in the situation, including language grounding, multimodal perception, long-horizon planning, and cross-domain generalization. In contrast, large language models (LLMs) show a stronger and promising potential by equipping agents with strong reasoning, planning, and communication capabilities in its distinctive paradigm shift. In this survey,we will comprehensively demonstrate emerging LLM-based multi-agent systems in four key aspects:(1) communication and interaction, which governs how agents swaps information with coordinated reasoning;(2) system architecture and coordination, defining the way to assign and organise complex tasks;(3) evaluation and benchmarking, which is used to measure agents’abilities in different environments;(4) learning, evolution, characterizing how agents improve, adapt and generalize. Through these advanced procedure of refereed dimensions, we reveal how LLMs reshape MASs’ designing space, which can realize the more flexible, extensible and human-aligned coordination strategies. Finally, we discuss open challenges including efficiency, reliability and standardized evaluation, and outline future research directions toward robust and autonomous multi-agent intelligence.

Online recommendation systems need to balance exploration with exploitation while responding to changing user preferences. Linear Upper Confidence Bound (LinUCB) provides an efficient contextual-bandit framework. However, it usually relies on a relatively stable user-interest representation. This paper examines whether combining long-term and short-term genre-preference features improves LinUCB-based movie recommendation, and when short-term signals are beneficial under preference drift. Using MovieLens-1M, experiments are conducted on the 50 most active users. Each round presents one logged movie and nine sampled distractors. LinUCB-Fusion with α = 0.85 achieves the highest average positive-feedback hit rate of 48.81%, compared with 48.42% for LinUCB-Long and 43.06% for Random. The average improvement over LinUCB-Long is small. A clearer pattern appears after splitting users by preference drift: Fusion helps high-drift users but hurts stable users. This suggests that short-term signals are mainly useful when user interests are changing, not for all users equally.

Life Cycle Assessment (LCA) is a crucial method for evaluating the environmental impact of products throughout their entire lifecycle. However, its practice still faces challenges, including insufficient data credibility, rigid auditing methods, and low motivation for enterprise participation. Research has found that introducing blockchain technology can improve the credibility and auditability of LCA data and incentivize enterprises to participate in LCA calculations. Therefore, this paper proposes a multi-level LCA data governance framework based on blockchain technology. It constructs a corresponding auditing system and a token incentive mechanism that combines cost compensation and quality incentives. This paper provides a generalizable theoretical framework for the design of LCA data supervision and incentive mechanisms.

This paper investigates Ethereum-based smart contracts as a decentralized cybersecurity governance layer for Electronic Medical Records (EMRs). It addresses challenge of fragmented healthcare data silos and strict compliance requirements for electronic protected health information (ePHI), such as HIPAA and GDPR by analyzing how blockchain technology enables explicit, patient-centric access control. Current legacy systems often centralize authorization, creating vulnerablilities. Through topic analysis and a Solidity-based Proof of Concept (PoC), the study demonstrates a hybrid architecture that uses on-chain execution for permission management while storing sensitive clinical data off-chain. The PoC shows how immutable ledger entries enforce least-privilege principles, block unauthorized queries, and provide tamper-evident auditability. The architecture’s trade-offs between enhanced accountability, patient sovereignty, and practical limitations in scalability, governance, and costs are assessed. It concludes that while blockchain offers a robust trust layer, integration with off-chain interoperability standards remains essential for clinical adoption.

Due to information overload in e-commerce, users often suffer from low decision-making efficiency. AI recommendation algorithms have become an effective approach to mine user behavior data and provide personalized product suggestions. Based on a public e-commerce user behavior dataset, this study uses the R language to conduct data cleaning, visualization, conversion funnel analysis, and logistic regression. Results show that user click browsing behavior occupies an absolute dominant proportion in all interactive behaviors, which accounts for 94%, while collection behavior is nearly absent, and add-to-cart behavior only accounts for a small proportion of total behaviors. The conversion funnel indicates serious user loss from click to collection, and the logistic regression shows that click, collection and add-to-cart behaviors have no significant driving effect on purchase conversion under the limitation of existing sample data without purchase records. This paper provides targeted suggestions for e-commerce platforms to optimize recommendation strategies and improve user conversion efficiency.

This review examines single-robot onboard path planning for urban delivery robots operating on known road or sidewalk networks under embedded resource constraints. The analysis emphasizes practical deployment scenarios, where CPU capacity, memory size, power budget, and planning time are limited. Based on representative journal and conference studies in mobile robot navigation, delivery robotics, and embedded robotic systems, the paper reviews how known urban networks can be represented and how planning methods can be selected for resource-constrained platforms. The discussion covers graph-based map representations, global route planning methods, local obstacle handling, incremental replanning, and implementation-oriented strategies like compact data structures, offline preprocessing, bounded online search, and fallback mechanisms. The results demonstrate that dense map representations are generally unsuitable for low-cost onboard deployment, while sparse topological or hybrid graphs provide a better balance among storage requirements, search efficiency, and implementation simplicity in known-network scenarios. Moreover, classical graph search algorithms, like Dijkstra and A*, remain effective for small to medium networks. In contrast, preprocessing-based or incremental approaches are more advantageous when handling repeated queries, local cost updates, or strict timing constraints. It further indicate that lightweight graph-based path planning with bounded runtime and local recovery is more suitable for urban delivery robots than heavy, compute-intensive planning pipelines.

The increasing demand for verifiable computation in privacy-sensitive distributed systems has driven the widespread adoption of Zero-Knowledge Proofs (ZKPs). However, the various kinds of current ZKP frameworks—which include zk-SNARKs, zk-STARKs, Bulletproofs, and folding-based systems—introduce complex trade-offs across proof size, prover cost, and trust assumptions, making system selection challenging in actual practice. This paper presents a systematic, application-oriented survey that connects ZKP design choices with real-world deployment constraints. It provides a comparative analysis of major constructions to evaluate their performance and security properties. Furthermore, these trade-offs are mapped to representative application scenarios, including Layer 1/Layer 2 blockchain scaling, Decentralized Identity (DID), and Verifiable Machine Learning (zkML), explaining how different systems are selected based on application-specific requirements. In addition, the paper discusses emerging paradigms such as hardware acceleration, binary field optimizations, and lookup-based zkVMs, which aim to address the prover bottleneck. Overall, this survey provides a structured understanding of the strengths and limitations of existing ZKP systems and offers insights for the design of scalable and privacy-preserving infrastructures.