Conceptual modeling is instrumental in data engineering and knowledge representation. With the increasing adoption of Large Language Models in data-centric tasks, their capacity to generate conceptual models, such as Entity-Relationship Diagrams, continues to require attention. This paper evaluates the performance of Large Language Models in defining Entity-Relationship Diagrams from datasets. To this end, we designed a reproducible experimental framework in which two models of different natures (Mistral Small and Gemini 2.5 Flash) and a deterministic profiling baseline without any language model are evaluated on two datasets, under a full combined scenarios that crosses three types of expert knowledge (a domain description, an attribute-level data dictionary, and business rules) with three levels of informativeness of the attributes’ names, at three sampling temperatures with replicated executions. In summary, 864 generated models, plus six deterministic baseline evaluations, for a total of 870 evaluated models, all scored automatically against reference models curated by domain experts, across entities, relationships, keys, and attributes. The results show that no knowledge type is universally the most effective, that more knowledge is not necessarily better, that the benefit of expert knowledge varies between the two evaluated models and across datasets, and that the informativeness of the attributes’ names, rather than their mere presence, conditions the results; temperature affects the reproducibility of the outputs more than their quality. These findings help to delimit the roles of the data, the language model, and the domain expert in conceptual modeling. The complete experiment is available as a public replication package.
Large language models (LLMs) have advanced question answering, yet remain unreliable for domain-specific multi-hop queries that require structured evidence and traceable reasoning. Text-centric Retrieval-Augmented Generation (RAG) provides external grounding but often fails to preserve the relational dependencies needed to reason over interconnected knowledge-graph facts. DEEPGRAG is a modular GraphRAG framework that combines topic-entity and intent identification, depth-bounded intent-guided graph traversal, two-tier structural and semantic filtering, and edge-aware chain-of-thought reasoning over compact, linearized path representations. A post-generation Factual Auditor verifies relational claims against the question-specific retrieved and pruned graph evidence, explicitly separating Generator-level evidence faithfulness from end-to-end answer correctness. Evaluation is conducted on RESQA, comprising approximately 100k question–answer pairs over a Requirements Engineering scholarly knowledge graph with 18,736 nodes and 10 relation types, and on MetaQA, a standard multi-hop KGQA benchmark, for cross-domain validation. DEEPGRAG achieves Hits@1 of 100%, 100%, and 99.0% on 1-, 2-, and 3-hop RESQA, and 100%, 99.5%, and 99.1% on the corresponding MetaQA splits, respectively. All 12,000 evaluated RESQA relational claims and 98% of 10,000 MetaQA claims are supported by the evidence supplied to the Generator. DEEPGRAG also outperforms Textualized KG-Triplet RAG and LLM-to-Cypher across hop levels and achieves 99% Hits@1 on 3-hop RESQA, compared with 93% for ToG and 95% for PoG, highlighting strong evidence-grounded reasoning on schema-heterogeneous, multi-hop queries.
Obtaining valuable information from large-scale datasets has long been a core research focus in the big data domain. With the rapid development of computer science and the Internet of Things in recent years, various data are continuously generated in the form of data streams. Data in data streams are characterized by infinite size and rapid updates, making it a crucial research topic in data mining to obtain useful information from them. In engineering application scenarios such as industrial product quality inspection and power grid data management, data exhibits the characteristics of being dense and having an infinite volume. Therefore, how to effectively process dense datasets over data stream scenarios has become an urgent problem. To tackle this issue, the present study puts forward an algorithm designated as PFUMS(Prefix-free Utility-Matrix Stream), which employs a sparse matrix structure to simplify the merging process of itemsets while occupying less memory. Additionally, the algorithm adopts a prefix-free itemset extension method and a two-phase local-utility pruning strategy during the join process. The pruning strategy allows for timely pruning of unpromising itemsets without waiting for their complete extension, effectively reducing the search space during the extension phase. Throughout the process, vector pointers and maps are used to maintain necessary data. In addition, during the window sliding process, global variables keep the data of the common part of the two windows from being cleared, so as to reduce the data scanning time in the new window. Experimental results show that it maintains excellent performance in dense datasets.
Critical data, including reports of injured persons and infrastructure damages is shared on social media after disasters. Nevertheless, during the disaster, detecting valuable information from a huge quantity of social media posts is regarded as more challenging. Therefore, a Convolutional Neural Network-based Transfer Learning+Deep CNN_Taylor Archery Algorithm (CNN-based TL+DCNN_TAA) is devised for event detection using images and texts from social media. Firstly, event detection is performed based on the input image using a Deep Convolutional Neural Network (DCNN), which is tuned by a Taylor-based Archery Algorithm (TAA). Here, TAA is formulated by the incorporation of the Taylor series and Archery Algorithm (AA). Simultaneously, input text data is tokenized, and then feature extraction is executed. Next, sentiment classification for event detection is performed by Random Multimodel Deep Learning (RMDL) with Convolutional Neural Network (CNN)-based Transfer Learning (TL). Alternatively, keyword extraction is carried out from the input text. Next, the keyword graph is generated by considering the keywords as nodes, and then keyword graph clustering is executed. Lastly, event detection is performed, and majority voting is utilized for combining all the outputs. Furthermore, the experimental outcomes of CNN-based TL+DCNN_TAA measured a maximum precision, recall, F1-score, and accuracy of 90.754%, 93.766%, 92.235%, and 91.787%.
The datasets used as input for AI models are often orders of magnitude smaller than raw big data files after summarization and feature engineering. Therefore, it is feasible to compute models locally on a modest computer or in the cloud on a powerful virtual server. From a machine learning theory perspective, linear regression is a fundamental predictive model and is used as a building block of neural networks when combined with non-linear activation functions. Historically, database systems have integrated highly optimized algorithms to store, update, and analyze large-scale data, but with somewhat limited mathematical capabilities. As a consequence, AI analytics have progressively migrated outside database systems, mostly to the Python ecosystem, leaving many statistical packages behind. Motivated by this gap, and drawing on core database system principles, we introduce a generic parallel algorithm that accelerates AI model computations by integrating two fundamental database optimizations: block-based I/O and data partitioning for parallel processing. This is programmed in Python via a Single-File Multi-Pointer (SFMP) system architecture and multiprocessing. We theoretically and experimentally show this generic algorithm can be used to accelerate and scale linear regression beyond main memory limits with two complementary, well-established solutions: data summarization with sufficient statistics and stochastic gradient descent. We present an extensive experimental evaluation validating accuracy and showing significant I/O acceleration. We also analyze parallel speedup trends and determine optimal parameter settings with large datasets of diverse dimensionality. Finally, we characterize analytic computations for which our parallel I/O algorithm is a good fit.
