Synthetic data is widely adopted in embedding models to ensure diversity in training data distributions across dimensions such as difficulty, length, and language. However, existing prompt-based synthesis methods struggle to capture domain-specific data distributions, particularly in data-scarce domains, and often overlook fine-grained relevance diversity. In this paper, we present a Chinese short video dataset with 4-level relevance annotations, filling a critical resource void. Further, we propose a semi-supervised synthetic data pipeline where two collaboratively trained models generate domain-adaptive short video data with controllable relevance labels. Our method enhances relevance-level diversity by synthesizing samples for underrepresented intermediate relevance labels, resulting in a more balanced and semantically rich training data set. Extensive offline experiments show that the embedding model trained on our synthesized data outperforms those using data generated based on prompting or vanilla supervised fine-tuning(SFT). Moreover, we demonstrate that incorporating more diverse fine-grained relevance levels in training data enhances the model's sensitivity to subtle semantic distinctions, highlighting the value of fine-grained relevance supervision in embedding learning. In the search enhanced recommendation pipeline of Douyin's dual-column scenario, through online A/B testing, the proposed model increased click-through rate(CTR) by 1.45%, raised the proportion of Strong Relevance Ratio (SRR) by 4.9%, and improved the Image User Penetration Rate (IUPR) by 0.1054%.

As a significant component of Intelligent Transportation Systems, traffic prediction attempts to forecast traffic trends in advance based on historical traffic characteristics. Although existing deep learning models for traffic prediction possess the capability to model spatial-temporal dependencies of traffic data, these methods tend to suffer from complex feature engineering and face challenges in natural language feature fusion and few-shot learning. In recent years, the rapid development of pre-trained Large Language Models (LLMs) has shown outstanding performance in time series analysis by leveraging their natural language comprehension and generalization ability. Inspired by LLM, this work proposes a novel Multi-level Embedded Spatial-Temporal Large Language Model (MEST-LLM) for traffic prediction. Specifically, we design multi-level embeddings of traffic data, which incorporate traffic historical features, dynamic spatial-temporal features, and prompt text features to model the complex spatial-temporal dependencies of traffic data and help LLM better understand traffic data through prompt text. Through fine-tuning strategies, LLM is capable of maintaining natural language comprehension and enhancing spatial-temporal prediction performance. Experimental results on four real-world traffic datasets demonstrated the excellent prediction ability and robustness of MEST-LLM in both sufficient and scarce data scenarios.

Abstract With the development of informatization, traditional industries have been seriously impacted by cross-border e-commerce (CBEC), and the financial management mode no longer meets the needs of CBEC users. As cross-border electronic trading platforms face enormous financial data challenges, they need financial supervision and effective network risk protection. Artificial intelligence (AI) can effectively improve the financial accounting ability and resource integration of platforms. However, at present, the underdeveloped financial management system of CBEC platforms and problems, such as the running records of platform funds have seriously hindered the development of the financial standardization of platforms because the financial management system affects managers’ statistical analysis of financial data. Therefore, this paper analyzes the financial risks, the management problems and the causes of these problems of CBEC platforms and then uses AI to study the financial operation of CBEC platforms. Subsequently, this paper uses a machine learning (ML) algorithm to analyze the financial data clustering center and the security factor of the financial model. Finally, this paper proposes some corresponding strategies for financial model optimization and construction that can improve the information security and capital management of CBEC finance and promote the long-term development of CBEC platforms. The experimental results show that the classifier value and the safety factor of the financial model of CBEC platforms gradually increase under the ML algorithm. The mean value of the classifier value is approximately 1.14, and the mean value of the safety factor is approximately 1.37. Overall, the initial value of the classifier value of the platform financial model is 0.85, which increases to 1.41 on the seventh day, while the whole process increases by 0.56. The initial value of the safety factor of the platform financial model is 1.10, which increases to 1.64 on the seventh day, while the whole process increases by 0.54. The transaction information security and financial accounting accuracy of the CBEC platform financial model is better than that of the original financial model. The transaction information security is 10.4% higher than that of the original financial model, and the financial accounting accuracy is 10% higher than that of the original financial model. In other words, both AI and ML can promote the financial accounting standards and long-term development of CBEC platforms.

Real-time 3D rendering and interaction is important for virtual reality (VR) experimental education. Unfortunately, standard end-computing methods prohibitively escalate computational costs. Thus, reducing or distributing these requirements needs urgent attention, especially in light of the COVID-19 pandemic In this study, we design a cloud-to-end rendering and storage system for VR experimental education comprising two models: background and interactive. The cloud server renders items in the background and sends the results to an end terminal in a video stream. Interactive models are then lightweight-rendered and blended at the end terminal. An improved 3D warping and hole-filling algorithm is also proposed to improve image quality when the user’s viewpoint changes. We build three scenes to test image quality and network latency. The results show that our system can render 3D experimental education scenes with higher image quality and lower latency than any other cloud rendering systems. Our study is the first to use cloud and lightweight rendering for VR experimental education. The results demonstrate that our system provides good rendering experience without exceeding computation costs.

Combined Heat and Power (CHP) systems are well known for their high efficiency and relatively low emissions. Existing CHP economic dispatch schemes do not use the energy buffer to minimize the average cost in the long term. Motivated by the queueing analysis and buffer management solutions in data communication systems, in this paper, we investigate how to use a battery pack and a water tank to optimize the average cost for the CHP systems by jointly considering the real-time electricity price, renewable energy generation, energy buffer states, etc. We first formulate the queueing models for the CHP systems, and then propose an algorithm based on the Lyapunov optimization technique which does not need any statistical information about the system dynamics. The optimal control actions are obtained by solving a non-convex optimization problem. We then discuss when it can be converted into a convex optimization problem. Since the battery pack queue and water tank queue are correlated by the CHP, the capacity relationship between them is further explored. Through the theoretical performance analysis, we also show the tradeoff between the cost saving and the energy buffer capacity. Finally, the effectiveness of the proposed algorithms is evaluated with practical data.

Dynamic taint analysis based software vulnerability and malware detection is an effective method to detect a wide range of vulnerabilities. Unfortunately, existing systems suffer from requirement of source code, high overhead or shortage of discovery rules, which limit their usage. This paper proposes a low-overhead vulnerability discovery system called DsVD (Dynamic Software Vulnerabilities Discoverer). DsVD works on X86 executables and does not need any hardware change. A new taint state called controlled-taint is introduced to detect more types of vulnerabilities. Our experiments show that DsVD can effectively detect various software vulnerabilities. DsVD incurs very low overhead, only 3.1 times on average for SPECINT2006 benchmarks. With some optimizations such as Irrelevant API Filter and Basic Block Handling, it can reduce runtime overhead by a factor of 4-11 times.