In modern baseball, the relative contributions of velocity, spin, and pitch location to swinging-strike outcomes are still not well understood, and many earlier studies mix up location-driven outcomes with actual pitch quality. This study uses MLB Statcast pitch-level data from 2024–2025, drawn from 221,982 pitches thrown by World Baseball Classic 2026 roster pitchers across 16 countries, to analyze what physical factors predict pitch effectiveness. To separate pitch quality from location effects, this study defines a zone-controlled binary target: in-zone swinging strikes versus in-zone hard-hit contact, while excluding out-of-zone pitches, called strikes, foul balls, and other ambiguous outcomes. This yields a modeling sample of 24,710 pitches with a balanced 45.0%/55.0% class split that needs no resampling. Using a stratified 70/30 split that preserves this class ratio, three classification models are trained and compared: logistic regression, decision tree and random forest, with the two tree models tuned by cross-validated grid search. Exploratory data analysis shows that pitch velocity and vertical plate location separate the two classes most clearly. The random forest classifier achieves the best performance with an accuracy of 67.84%, an ROC AUC of 0.7359, and an F1-score of 0.62, outperforming the decision tree (accuracy 65.72%, AUC 0.7042) and the logistic regression baseline (accuracy 57.37%, AUC 0.5759). Notably, feature importance analysis shows that stuff-related features together account for 46.7% of importance in the random forest—nearly as much as the two location features (53.3%)—showing that pitch quality, not just placement, is an important and often overlooked driver of swinging strikes.

Population aging is increasing demand for unobtrusive in-home monitoring, while cameras raise privacy concerns, and wearables depend on user compliance. This critical survey and design study compares WiFi channel-state information sensing and millimeter-wave FMCW radar for fall detection, activity recognition, and vital-sign monitoring. It identifies cross-environment generalization, multi-occupant attribution, and privacy evaluation as key barriers to deployment. An ISAC-inspired architecture is proposed that combines broad, low-cost WiFi coverage with radar in high-risk areas and uses cross-modal consistency checking to detect modality-specific unreliability. A controlled synthetic simulation illustrates that this mechanism can prevent an unstable modality from degrading fusion under asymmetric domain shift, but does not establish real-world or clinical performance. The paper also distinguishes among visual, behavioral, and health data privacy and identifies priorities for synchronized datasets, cross-home evaluation, and privacy-aware alert governance.

As an urgent problem in image restoration, deblurring aims to recover the original clear image of a blurred picture that has been damaged by motion, defocus, etc. Most of the above are explicit blur-kernel estimates. However, the distribution of these is not uniform. With the progress of deep learning in recent years, many end-to-end methods have been developed based on CNNs, RNNs, GANs and Transformers to reduce the need for hand-crafted priors. The first is to list some types of blur and their corresponding formulas. Then, it will introduce the general ideas, typical examples and applications of the four main architectural styles, as well as their particular advantages and deficiencies in different blurring conditions. List some representative benchmark datasets and image quality assessment methods from the survey, then discuss the main problems of computational efficiency, cross-domain generalization and lightweight deployment, and future research directions.

As the global manufacturing industry accelerates its digital and intelligent transformation, the deep integration of Artificial Intelligence (AI) with the Internet of Things (IoT) is emerging as a central driver behind the evolution of smart manufacturing systems. This paper provides a comprehensive review of the technical architecture, core technologies, and typical application scenarios of smart manufacturing within the context of this integration. From a technological perspective, it examines the transition from convolutional neural networks to vision transformers in industrial quality inspection, and explores how large language models enhance manufacturing interaction and test automation by serving as industrial Copilots, knowledge base question-answering systems, and code generators. The study also highlights the foundational role of IoT in data acquisition, real-time transmission, and intelligent analytics. In terms of applications, it demonstrates productivity improvements and modeling innovations across areas such as predictive maintenance, intelligent quality inspection, production scheduling, and smart warehousing. Despite notable progress, several challenges persist, including data silos, insufficient interpretability of models, difficulties in integrating operational technology with information technology, and a lack of interdisciplinary talent. Future research should prioritize strengthening data infrastructure, optimizing model architectures, advancing system integration and ecosystem development, and facilitating the shift of smart manufacturing from specialized tools to universal platforms, thereby offering stronger support for the intelligent advancement of the manufacturing industry.

Sports result prediction is a long-standing topic. Yet most existing studies rely heavily on team-scale data and overlook individual players, leaving abstract qualities like chemistry unexplored. This study emphasizes feature engineering to holistically capture each team's power, style, pace, and chemistry. Since conventional stats are insufficient for modern basketball, novel features are constructed manually based on NBA metrics and in-depth knowledge of the game—including TS%, DNP quantification, weighted standard deviation, winning streaks, and starter-bench depth—via an aggregation step, and game-lag was applied to prevent data leakage. This rich feature engineering successfully integrated players' differentiated performances into machine learning, enabling the model to capture the team's true level in a nuanced way. XGBoost, Random Forest, and CART were selected as the prediction model. The results revealed that these new features significantly boost prediction accuracy. The analysis of feature importance also underscores the significant role these features play.

Urban neighborhoods used to be fairly straightforward—residential here, commercial there. That is less and less true, and planners need ways to keep up. This paper tries to build one such tool, tested on Pudong New Area in Shanghai. We combine a TF-IDF weighting scheme with survey-based perception scores and feed both into a POI-based functional-zone classifier. TF-IDF handles the spatial-rarity side of the problem. If a POI category clusters heavily in a particular grid cell but barely appears elsewhere in the city, the algorithm gives it a high weight. A cargo terminal near an airport, for instance, is rare city-wide and gets flagged; a convenience store, which is everywhere, does not. The perception scores, borrowed from Zhao (2008), get at something else entirely. They reflect how strongly residents mentally connect different facility types with particular urban functions. Metro stations top the scale at 1.0. Residential compounds sit at 0.15—a gap that frequency-density counting, which gives every POI the same weight, has no mechanism to detect. We gridded Pudong into 1,672 cells of 1 km × 1 km. The baseline frequency-density method scored 66.74% overall accuracy. Our model reached 86.33%. Transportation zones and mixed-use areas saw the biggest gains, which makes sense because those are the zone types where brute-force counting is most prone to error. We also mapped single- and mixed-function zones across the district. The practical takeaway is that folding perception data into a TF-IDF framework yields a noticeably sharper read of urban functional structure.

This paper systematically examines bias in Large Language Models (LLMs) through the lens of Data Feminism. The research demonstrates that bias in LLMs is not an issue of technical neutrality, but rather a systematic reproduction of societal power structures embedded in data collection, algorithmic design, and output generation. This analysis reviews relevant empirical studies and reveals both progress and limitations in current research on bias detection and measurement. The paper argues that future research should transcend mere technical adjustments by developing intersectional bias evaluation frameworks, exploring the feasibility of constructing "Feminist LLMs," and translating the critical insights of Data Feminism into actionable ethical standards and policy guidelines. These steps are critical to advancing a more equitable and responsible AI ecosystem.

Heterogeneous computing has emerged as an important architectural response to the slowdown of conventional semiconductor scaling, particularly in the post-Moore era. This review follows the shift from homogeneous multicore processors toward integrated platforms that place CPUs, GPUs, FPGAs, and domain-specific accelerators on the same die. The transition is closely related to two long-standing technology constraints, namely the slowing of Moore's Law and the breakdown of Dennard scaling, while Amdahl's Law still limits the speedup that can be obtained from parallel acceleration. From the system perspective, the review focuses on three recurring obstacles: the memory wall, cache coherence in shared-memory heterogeneous architectures, and on-chip communication supported by Network-on-Chip (NoC) designs. I also reviewed low-power VLSI approaches, especially clock gating and low-swing flip-flop designs, together with design methodologies such as SystemC-based high-level synthesis (HLS), the Catapult HLS flow, and latency-insensitive design. Based on these 24 research contributions, I organize the discussion into a multi-level technical framework for heterogeneous SoC design in AI and IoT applications.