Knowledge Graphs (KGs) store structured information but typically require specialized query languages, such as Cypher for Neo4j, creating accessibility challenges for users unfamiliar with graph syntax. Large Language Models (LLMs) offer a solution by translating natural language into Cypher queries. However, existing models—including large-scale LLMs (e.g., ChatGPT) and smaller open-source models (e.g., Llama-7B, 8B) often struggle with accurately generating domain-specific queries due to inadequate alignment with KG schemas and limited domain-specific training data. To address these limitations, we propose a training pipeline tailored specifically for domain-aligned Cypher query generation, emphasizing usability for smaller-scale models. Our method integrates template-based synthetic data generation for diverse, high-quality training samples. We combine supervised fine-tuning with preference learning to enhance domain knowledge and Cypher syntax understanding. Additionally, our approach includes a context-aware retrieval mechanism that dynamically incorporates relevant schema elements at inference, improving alignment with domain-specific knowledge. We evaluated our method on the Hetionet biomedical KG using a benchmark dataset of 240 queries across three complexity levels. Our results show that our context-aware prompting achieves a substantial improvement, increasing component matching accuracy by 23.6% for ChatGPT-4o over the vanilla prompt baseline. When applying our full training pipeline to smaller-scale models, CodeLlama-13B* achieves an execution accuracy of 69.2%, nearly matching ChatGPT-4o’s 72.1%. Importantly, our approach significantly narrows the performance gap, enabling smaller models to effectively manage complex, domain-specific tasks previously dominated by larger models. These findings demonstrate that our method is scalable, computationally efficient, and robust for practical Cypher query generation applications.

Video-based activity recognition has a wide array of applications across various industries. A persistent challenge in practice is the scarcity of labeled datasets from target domains needed for training machine learning models capable of accurate activity recognition. Cross-Domain Few-Shot Learning (CDFSL) offers a promising solution by facilitating knowledge transfer from a label-rich source domain to a data-scarce target domain. However, existing CDFSL approaches often overlook the similarities between the source and target domains on the achievable performance of activity recognition in the target domain. Existing metrics for measuring domain similarity, such as Maximum Mean Discrepancy, focus primarily on data distribution and fail to provide actionable guidance for selecting source domains. To address this gap, we explore domain similarities and their effects from various angles, including video attributes like camera angle, background scene, and label granularity. For our case study, we focus on activity recognition within industrial environments as the target application and apply state-of-the-art CDFSL methods across diverse source-target combinations, using open datasets such as Kinetics-100, HMDB51, HA-VID, and Meccano. Our results offer valuable insights into the influence of domain similarities, which can aid in the selection of source domains.

Cameras are widely deployed for visual inspection in manufacturing and construction industries, generating massive volumes of image and video data that demand automated solutions for defect detection. Knowledge distillation has shown strong potential for unsupervised industrial visual anomaly detection, a task of increasing importance in such large-scale, data-intensive environments. However, the conventional single-student-single-teacher framework often yields suboptimal learning of normal patterns, primarily due to the absence of specific constraints and potential feature loss. To address these challenges, we propose S2TKD, a novel dual-student-single-teacher architecture. It incorporates a pre-trained teacher network, an Anomalous Feature Denoising Student (AFDS) network, and a Normal Feature Student (NFS) network. The AFDS network focuses on filtering out anomalies by imposing stronger constraints on anomalous data, while the NFS network extracts normal features to recover subtle patterns suppressed by the AFDS network. To further improve the richness of feature representations within each student network, we integrate a multi-scale feature fusion module Dual Pyramid Network (DPN) between the encoder and the decoder. Furthermore, we propose a Dual Fusion Network (DFN) that accurately identifies anomalous regions by fusing the similarity of the outputs of each student and teacher network to obtain multi-scale similarity maps, which are then adaptively aggregated to generate the final anomaly map. Experimental results on four benchmark datasets demonstrate S2TKD's outperforms compared to the state-of-the-art. On the representative MVTec AD dataset, the I-AUROC, P-AUROC, and PRO reached 99.5%, 99.1%, and 96.9%, respectively, with the PRO showing a 1.9% increase over the current best result. The results highlight S2TKD’s effectiveness and scalability for large-scale industrial visual anomaly detection, making it a promising solution for real-world big data inspection systems.