Recommendation systems have permeated society and everyone’s daily life with the expanding use of applications. It plays an important role in providing recommendations to active users based on either their historical interests or interests shown by similar users. This paper aims to introduce several components of a recommendation system, such as Collaborative Filtering (CF) and Matrix Factorization, and combining the dominant approach and main technique: Singular Value Decomposition (SVD). Additionally, with the advances in technology, various recently established SVD-based models are mentioned in this paper, attempting to give an answer to the question that though studies have shown the coexistence of advantages and disadvantages for each mode, the trade-offs between their predictive accuracy, computational cost, and robustness to data sparsity have still not been fully answered. Furthermore, this paper concentrates on the comparison and analysis of diverse matrix factorization variants, meanwhile evaluating their performances. For instance, basic SVD, BiasSVD, and SVD++. Finding indicates that SVD shows more efficiency in small-scale datasets, BiasSVD is suitable for cases with noticeable biases, and SVD++ gives high-quality results when handling both explicit and implicit feedback data. This study provides practical guidelines for practitioners to select the most desirable model based on their existing database.

Pulmonary micronodules often present with weak contrast, blurred boundaries, and easy adhesion to blood vessels or pleura in Computed Tomography (CT), which leads to missed detection and missegmentation in automatic segmentation, affecting the consistency of screening and follow-up evaluation. This paper reviews the research progress of U-Net and its improved structures in the segmentation of pulmonary micronodules. It focuses on summarizing technical routes such as multi-scale feature fusion, attention mechanism, three-dimensional convolution, and convolutional Transformer hybrid structure, and summarizes and compares the performance of common indicators of related methods on datasets such as DCC-IDRI and LUNA16. Overall, multi-scale dense connection, spatial-channel attention, and long-range dependency modeling are helpful to improve the stability of edge characterization and segmentation of small nodules, but they are still constrained by factors such as weak features, background complexity, and cross-device domain offset. In the future, the robustness and generalization ability in multi-center scenarios can be further enhanced from directions such as lightweight 3D modeling, multi-modal fusion, semi-supervised learning, and explainability mechanisms.

Nowadays, the discovery of exoplanets has become a popular field for scientists to study due to the limited resources on Earth. With the help of developed technology and different areas of data, through several years of observation and research, a variable and a huge number of exoplanets have been discovered. The methods to find these exoplanets are really important to learn. However, there are so many methods to use, and each of them has its own shortcomings. This paper will explain the detailed principles, formulas, and information of two common methods, which are Transit and Radial velocity. The drawbacks of the two methods will be illustrated, and then the paper will give the idea of combining the two methods to get more planetary parameters and better outcomes. The result of the work is that although the size of all planets and transit durations can be visually presented through spectroscopy, most of the parameters of the planet still cannot be obtained; also, there is a limitation on the change in inclination. Radial velocity can provide almost all the needed parameters for exoplanets, but this method is only sensitive to massive planets. Easy to tell, both methods have their respective advantages that the other party needs. Therefore, combining these two methods is the best way to complement weaknesses, making it easier and better to study exoplanets.

The integration of different matrix retrieval technologies is likely to shape the future direction of technological development. By combining traditional methods such as TF-IDF with the powerful ability of deep learning to process high-dimensional data, which can build precise retrieval systems. These systems integrate graph neural networks with emerging tools effectively, such as PageRank, improving link analysis by directly combining node attributes like entity types and content details into the sorting process, therefore enhancing context understanding capabilities. For the efficient execution of complex computations similar to transformer models, innovative strategies like scarce attention are crucial techniques, such as sliding window attention and low-rank approximation play a crucial role in handling long text, large code bases, and multiple iterative searches. In addition, technological advancements in dedicated hardware devices like TPU have lessened the challenges brought by intensive matrix computations, making advanced real-time search tasks that were previously unachievable possible. The coordinated development of algorithms and hardware is elevating the semantic parsing capabilities of search systems to a remarkable height. Future search applications will deeply integrate context and personalized cognition, establishing benchmark standards for a comprehensive understanding of complex data, thereby changing educational models and information acquisition methods.

Traditional studies on social issue films lack multi-dimensional quantification and visualization analysis, despite the importance of these films’ distribution, market performance, and correlation rules for social issue dissemination. This study examines four types of social issue films (gender equality, poverty alleviation, education equity, and disability rights) from 2015 to 2024, using 1,000 stratified random samples and tools including statistical analysis, TF-IDF text mining, PageRank algorithm, and visualization methods. It depicts genre and annual output trends, compares box office performance, analyzes rating-box office correlations via relevant charts, and constructs an original association matrix (P matrix) by integrating plot similarity and cast collaboration; the matrix is further optimized with a damping factor (α=0.85) to form matrix A. Results show that the four core genres account for 50% of total samples, with education equity films leading in average box office and ratings; the optimized matrix elevates the weights of education equity-gender equality and poverty alleviation-gender equality genre combinations notably. Compared with traditional PageRank, the optimized algorithm boosts core issue films’ influence proportion by 25.0% and recommendation accuracy by 23.5%. This research quantifies genre supply gaps and linkage potential, offering data support for policy guidance, creation support, and precise communication while enriching the quantitative research paradigm for social issue films.

Human population and urban systems are typical resource-limited dynamic systems, in which growth is constrained by land, food, energy, and ecological capacity. Classical logistic differential equations describe these processes with an S-shaped trajectory, yet most applications still treat the environmental carrying capacity as a fixed constant. This paper uses recent theoretical and empirical studies to revisit resource-limited population growth from the perspective of differential equations. Drawing on the logistic,θ-logistic, and resource feedback models, a unified framework is first summarized, in which both population size and carrying capacity can evolve. Several representative case studies are then analyzed. These include logistic fitting of Wuhan’s population, resource-limited growth models that embed explicit resource dynamics, and logistic-type models of urbanization and ecological environmental quality in Chinese cities. The discussion highlights new advances after 2020, especially models that link carrying capacity to food availability and other socio-ecological variables. The results show that logistic-type differential equations remain powerful tools for capturing long-term trends, but their explanatory power is greatly enhanced when feedback from resource production, technological progress, and policy is incorporated. The paper concludes that future research should move from single-equation curve fitting toward coupled socio-ecological systems, providing a stronger quantitative basis for sustainable population management and urban planning.

Sea level rise is one of the most certain and highly destructive consequences of global warming, posing a serious threat to coastal communities worldwide. The ability to accurately predict its magnitude and rate directly underpins the efficacy of coastal adaptation planning. This paper systematically reviews the main physical mechanisms driving sea level rise and conducts a comparative analysis of semi-empirical models and process models in terms of performance, applicability, and uncertainties. Integrating key research findings from the IPCC Sixth Assessment Report, the article comprehensively presents projections of global and regional sea level rise under different emission scenarios. The results show that under the high-emission scenario SSP5-8.5, the global average sea level could rise by up to 99 centimeters by 2100 and nearly 290 centimeters by 2300. Regional factors, such as land subsidence and ocean dynamics, will exacerbate localized impacts. This paper provides valuable reference for global coastal adaptation efforts.

This paper develops a reinforcement learning framework for optimizing liquidity allocation in multinational enterprises' overseas cash pools. The framework models a multi-regional digital-treasury environment incorporating currency conversion, regulatory frictions, time-zone misalignment, cut-off times, intercompany limits, and intraday payment volatility. An actor-critic agent observes multi-currency balances, forecasted cash flows, FX quotes, and regulatory flags to recommend sweep amounts, buffer adjustments, and internal loans at discrete intraday intervals. The reward function jointly penalizes funding cost, idle balances, constraint breaches, and FX slippage, while safety masks enforce hard limits. Synthetic treasury-plausible data are calibrated to historical distributions of payroll cycles, tax dates, and market stress episodes. Numerical experiments across 9.6×10⁵ simulated days demonstrate that the RL policy reduces net funding cost by 18.7±3.4% and idle liquidity by 26.2±5.1% relative to rule-based approaches, while maintaining regulatory breaches below 0.08±0.02 events per 1,000 transactions. Stress tests confirm graceful performance degradation under FX gaps and sudden outflow shocks. Results suggest enterprise-tailored RL can enhance overseas treasury performance while remaining compatible with operational and compliance requirements.

The large-scale development of online education has given rise to an urgent demand for digital teaching interaction. As the leading role in teaching, teachers' facial expressions contain rich emotional information, which is directly related to students' learning enthusiasm, classroom participation and the effect of knowledge understanding. Aiming at the problem of insufficient capture of key features in dynamic expression classification by LSTM, this paper proposes an expression classification algorithm based on multi-head attention mechanism to optimize LSTM. The research first conducts data mining through violin graph analysis and correlation heat map analysis, and then conducts comparative experiments with various machine learning algorithms such as decision tree, SVM, and KNN. The results show that all performance indicators of the proposed Multihead-Attention-LSTM model are optimal: the accuracy rate, recall rate, precision rate and F1 value all reach 99%, and the AUC value and ExtraTrees are both 100%, which is significantly better than other models. This research effectively enhances the accuracy of dynamic expression classification, providing reliable technical support for the digital implementation of teaching emotional interaction in online education scenarios, and has significant practical significance for improving the effectiveness of online teaching.

Rapid urbanisation and rising academic pressure have made adolescent anxiety more subtle and embedded in daily life, while survey-based studies at administrative scales struggle to capture the spatiotemporal heterogeneity and causal impact of urban noise. To address this gap, a multi-city longitudinal cohort of junior and senior high school students is constructed, combining minute-level noise, physiological and sleep data from wearables with road-network, land-use, nighttime light and vegetation indices from remote sensing and GIS, forming a high-resolution panel at the “student–time slice–spatial unit” level and a spatiotemporal graph with student, grid and commuting-path nodes. Results show that a 5 dBA increase in night-time noise yields a significantly larger causal increment in GAD-7 scores than comparable increases during commuting or school hours, with effects amplified around schools near arterial roads, in low-NDVI communities and among adolescents with high baseline anxiety; greenness and spatial position exhibit clear buffering or amplifying slopes. The study provides multi-source-data-based causal evidence for the “noise–sleep–anxiety” pathway and offers both methodological and empirical foundations for precision design of school quiet zones, urban greening and adolescent mental health interventions.