Database Systems Journal

ISSN 2069 - 3230

The journal is published under the sponsorship of
The Bucharest University of Economic Studies
and it is produced by the university's own publishing division,
The Bucharest University of Economic Studies Publishing House


Database Systems Journal, Vol. XVII, 2026


Open PDF Journal


CONTENTS


1. A Hybrid Approach to Real-Time Recommendations using Heterogeneous Data Processing and Distributed Predictive Models (p. 1-9)
Diana-Andreea CAUNIAC, The Bucharest University of Economic Studies, Romania
Simona-Vasilica OPREA, The Bucharest University of Economic Studies, Romania
As e-commerce platforms scale, the gap between what users expect and what static recommendation models can deliver has become hard to ignore. This paper describes the design and implementation of a hybrid recommendation system built on a distributed cloud infrastructure, tested on a dataset of over 3.7 million products, 1.9 million reviews, and 1.1 million user interactions. The system combines collaborative filtering through matrix factorization with sentiment analysis, emotion detection, and topic modeling applied to user reviews, identifying six recurring themes that reflect real purchasing experiences. Geographic proximity and recent behavioral signals are incorporated as contextual features to further refine recommendations. The stack includes Google BigQuery, Vertex AI Feature Store, Pinecone, Apache Kafka, Apache Flink, and BigQuery ML. A feedback loop ties user interactions back to model updates, keeping recommendations relevant as behavior changes. Results suggest that decoupling the processing pipelines reduces latency without sacrificing recommendation quality.
Keywords: recommendation systems, real-time processing, hybrid architecture, collaborative filtering, heterogeneous data, sentiment analysis, topic modeling, distributed computing
2. Data Preparation with Oracle Cloud (p. 10-19)
Cristiana COSTAN, The Bucharest University of Economic Studies, Romania
This article examines the Extract-Transform-Load (ETL) process as an essential method for preparing data from warehouses for analysis. It also covers the difficulties of combining data from numerous external sources, converting it for consistency, and storing it in the cloud for effective querying. Oracle Cloud Infrastructure (OCI) provides a range of services for ETL workflows, including Oracle Data Integrator (ODI), which offers a variety of code-based data processing options, and Data Transforms, also known as ODI Web, a modern no-code solution designed to simplify and automate data transformations and machine learning tasks. A practical example using a League of Legends dataset illustrates how Data Transforms and Autonomous Data Warehouse can be used for effective data preparation and analysis. The goal of this paper is to assist data scientists in simplifying and improving data workflows through the use of modern cloud ETL solutions.
Keywords: Data Warehouse, ETL, Oracle Cloud, Data Integration, Machine Learning
3. Data Visualization in Business Intelligence: A Comparison Between Power BI and Qlik Cloud Analytics (p. 20-28)
Anda-Elena SPATARU, The Bucharest University of Economic Studies, Romania
Florin-Răzvan SOARE, The Bucharest University of Economic Studies, Romania
The visualization component within Business Intelligence (BI) systems plays an essential role in transforming operational data into actionable insights for the decision-making process. BI tools facilitate data interpretation and the formulation of evidence-based conclu-sions, supporting organizations in monitoring performance through clear metrics and adapt-ing rapidly to changes in the competitive environment, including in low-latency information contexts. From this perspective, the article presents general BI concepts and proposes a comparative analysis of two representative platforms, Microsoft Power BI and Qlik Cloud Analytics, highlighting their capabilities and limitations based on relevant evaluation criteria.
Keywords: Business Intelligence, Microsoft Power BI, Qlik Cloud Analytics, Analytics and Reporting Platforms, BI Architecture, Data Warehouse, Data Mart, Comparative Analysis
4. An Intelligent Recommendation System Built on Emotional Analysis in a Kappa Architecture (p. 29-35)
Diana-Andreea CAUNIAC, The Bucharest University of Economic Studies, Romania
Adela BARA, The Bucharest University of Economic Studies, Romania
Recommendation systems have become a cornerstone of modern e-commerce, directly shaping user experience and conversion rates. Yet most conventional approaches rely solely on behavioral history, clicks, views, purchases, without any awareness of how a user is feeling in the moment of decision. This paper presents an intelligent recommendation system that fills that gap by weaving emotional analysis of text reviews into a real-time data-processing pipeline. The solution is built on a Kappa architecture and leverages Apache Kafka, Google Cloud BigQuery, Bigtable, BigQuery ML, and a RoBERTa-based NLP model for emotion detection. Users are grouped into clusters through K-Means segmentation according to their emotional profiles and recommendations are then derived from well-established correlations between emotional states and product categories. In controlled evaluation, the system achieved a Precision@5 of 0.98, a Recall@5 of 0.94, and an F1-score of 0.96, confirming the strength of the proposed approach.
Keywords: Recommendation systems, sentiment analysis, BigQuery ML, Apache Kafka, Kappa architec-ture, RoBERTa, natural language processing, personalized recommendations
5. SQL vs NoSQL in Polyglot Persistence Architectures (p. 36-53)
Florin-Răzvan SOARE, The Bucharest University of Economic Studies, Romania
Anda-Elena SPATARU, The Bucharest University of Economic Studies, Romania
Miruna SOSEA, The Bucharest University of Economic Studies, Romania
Polyglot persistence is increasingly adopted because large-scale applications combine correctness-critical state with high-throughput, latency-sensitive workloads that cannot be served efficiently by a single datastore. This paper contrasts SQL and NoSQL systems in terms of transactional guarantees, recovery behavior, data modeling, query expressiveness, schema evolution, and operational scaling constraints. Based on this comparison, we derive datastore selection criteria that distinguish system-of-record components from derived serving models. We then discuss integration mechanisms for polyglot architectures, emphasizing explicit data ownership, change propagation via CDC and log-based replication, and saga-style coordination with compensations to manage cross-store failures.
Keywords: Polyglot persistence, SQL, NoSQL, scalability, CDC, sagas, ACID, BASE
6. Hybridizing Evolutionary Algorithms and Reinforcement Learning for Superior Decision-Making in Complex Systems (p. 54-61)
Denis-Claudiu ASAN, The Bucharest University of Economic Studies, Romania
This study examines the effectiveness of hybridizing Evolutionary Algorithms (EA) and Reinforcement Learning (RL) for decision-making in sparse-reward environments. Three models are implemented and compared, namely: (1) a population-based Evolutionary Algorithm with elitism, (2) an actor-critic Reinforcement-Learning method regularized with Behavior Cloning, and (3) a hybrid Lamarckian Memetic framework that combines evolutionary search with gradient-based fine-tuning. All models are trained offline using the antmaze-umaze-v2 dataset from the D4RL benchmark and evaluated both through critic-based offline estimates and online MuJoCo simulations. While all methods achieve comparable offline critic values, the hybrid approach demonstrates superior online normalized performance scores. The results suggest that combining population-based exploration with gradient-based exploitation mitigates local optima in sparse-reward decision-making tasks.
Keywords: Evolutionary Algorithms, Reinforcement Learning, Offline RL, Memetic Algorithms, Sparse Reward, Hybrid Optimization
7. Advantages of Data Warehouses in the Analysis and Study of E-Commerce Markets (p. 62-74)
Mario-Tudor CHIRIAC, The Bucharest University of Economic Studies, Romania
The present paper aims to explore the advantages of using data warehouses in the analysis and study of contemporary e-commerce markets through the development of an applied case study. By integrating Python, Oracle Database, Node.js, and React.js technologies, a comprehensive Business Intelligence solution is implemented. This solution enables the centralization, transformation, and visualization of relevant data for analyzing consumer behavior and identifying market trends. The proposed application provides a user-friendly interface, dynamic report generation capabilities, and advanced filtering functionalities. The study demonstrates that the adoption of data warehouse architectures significantly enhances the efficiency of the decision-making process within organizations, particularly in the context of increasing data volumes. The paper combines a theoretical framework with practical implementation, offering a functional and extensible example for economic market analysis using modern technological tools.
Keywords: Data Warehouses; Python; Oracle Database; Node.js; React.js; Market Analysis; E-commerce; Business Intelligence
8. Scalable and Nonlinear Forecasting with Quantum ARMA Models (p. 75-82)
S. ANU, Department of Statistics, Periyar University, Salem-11, Tamil Nadu, India
S. RITA, Department of Statistics, Periyar University, Salem-11, Tamil Nadu, India
Time series forecasting is very important in planning, climate modelling, and industrial making-decision. It helps make plans and decisions. The Autoregressive Integrated Moving Average model (ARIMA) is a statistical method that is still widely used. It has some problems. It assumes that data changes in a line, which is not always true. This thing needs the data to be more stable which is not easy to do. Working with a lot of data can make it really slow. This study looks at a method called Quantum Autoregressive Moving Average (QARMA). QARMA uses quantum encoding, the Quantum Fourier Transform, and the Harrow-Hassidim-Lloyd algorithm to make forecasts. We used two datasets to test QARMA and compare it to ARIMA. One dataset had information on unemployment rates. The other had weather observations. The results show that QARMA works better than ARIMA with data that changes a lot. QARMA reduces errors in forecasts. Is still fast. The tests also confirmed that QARMA is reliable and works well. The findings suggest that using quantum computing for forecasting can make it better. QARMA can handle data and large amounts of it. This makes QARMA a promising new direction for analytics.
Keywords: Time Series Forecasting, ARIMA, Quantum Computing, Quantum Machine Learning, QARMA.
9. Power BI as an Advanced Alternative to Microsoft Excel (p. 83-91)
Cristiana COSTAN, The Bucharest University of Economic Studies, Romania
This article explores Business Intelligence (BI) systems, focusing on a comparative analysis of Microsoft Excel and Power BI. Power BI is an extension of Excel because it utilizes its add-ons: Power Query, Power Pivot, and Power View. Both Excel and Power BI allow data transformation and visualization, but Power BI is more appropriate for advanced analysis. Its tools are designed for business-oriented analysis through dashboards, while Excel remains well suited for simpler calculations using summarized data. A practical example was developed using Power BI to highlight the power of this program to transform and visualize data. Comparisons with Excel were made to see how Power BI is an extension of the program, not a completely different program. The main purpose of this paper is to inform data analysts, data scientists, and BI enthusiasts about the appropriate contexts for using each tool and to identify when one may be more suitable than the other.
Keywords: Business Intelligence, Excel, Power BI, Data Analysis
10. Security of Business Intelligence Systems in the Face of Cyber Attacks (p. 92-102)
Mario-Tudor CHIRIAC, The Bucharest University of Economic Studies, Romania
This article aims to examine the effects of various types of cyberattacks on Business Intelligence (BI) systems, as well as the protective measures required to prevent such incidents. The simulated attacks include Denial of Service (DoS), Man-in-the-Middle (MITM), and Evil Twin attacks. These represent some of the most frequently employed techniques used by cyber attackers to compromise an organization's infrastructure, either to steal critical data for malicious purposes or to disrupt operations, potentially leading to severe financial losses or even bankruptcy. Therefore, it is essential that Business Intelligence systems be supported by a robust and well-secured infrastructure in order to mitigate risks and prevent catastrophic outcomes.
Keywords: Business Intelligence system security; BI systems; Denial of Service; Man-in-the-Middle; Evil Twin; security measures
11. A Comparative Analysis of Code Generation Accuracy: GPT-5.5 vs. Claude Sonnet 4.6 within Integrated Development Environments (IDEs) (p. 103-111)
Sabin-Marian ARSENE, The Bucharest University of Economic Studies, Romania
The rapid evolution of large language models (LLMs) has profoundly reshaped software development workflows, particularly within Integrated Development Environments (IDEs). This paper presents a comparative analysis of code generation accuracy between GPT-5.5, released by OpenAI in April 2026, and Claude Sonnet 4.6, developed by Anthropic, evaluated across six dimensions: syntactic correctness, semantic accuracy, long-context retention, hallucination rates, IDE integration quality, and multilingual support. Beyond standard benchmarks, this study introduces a practical evaluation framework applied to five representative IDE task categories across three experimental runs, complemented by a cost-efficiency analysis for enterprise deployment. Findings indicate that GPT-5.5 holds advantages in agentic multi-step reasoning and bug detection detail, while Claude Sonnet 4.6 demonstrates superior output consistency, instruction adherence, and lower operational cost. Neither model achieves universal supremacy; optimal selection remains task- and context-dependent.
Keywords: large language models, code generation accuracy, GPT-5.5, Claude Sonnet 4.6, IDE integration, agentic software engineering, automated debugging, benchmark evaluation,software development
12. Predicting Battery RUL: Integrating Classical Regression and Quantum Machine Learning (p. 112-122)
S. ANU, Department of Statistics, Periyar University, Salem-11, Tamil Nadu, India
S. RITA, Department of Statistics, Periyar University, Salem-11, Tamil Nadu, India
Quantum Machine Learning is recently being considered an emerging area in predictive analytics, primarily due to the ability of quantum computing, which is presumed to process complex patterns more effectively and quickly than current traditional models. Most of these assumptions are yet to be validated using concrete data. For the purpose of this study, I assessed the performance of two popular traditional models, Linear Regression and Random Forest, against the performance of the Quantum Circuit Regressor (QCR) model in the context of estimating the Remaining Useful Life (RUL) of batteries, which is an imperative parameter in scenarios involving electric cars and storing energy. The experiments were conducted using two different data sources, and the assessment of these models was done in similar settings. In the case of the first dataset, Random Forest performed outstandingly, with negligible prediction errors and an R² value of almost one, while Linear Regression performed moderately. In contrast, the quantum model struggled to fit the data and resulted in large errors along with negative R² values. A sim-ilar pattern was observed in the second dataset, where all models faced difficulties, but the quantum approach again failed to show any advantage over the classical methods. These results suggest that although quantum models are conceptually attractive, their practical benefits are not yet evident for this type of problem. In such a context, the Random Forest approach was the most accurate method for battery Remaining Useful Life prediction. The classical approach is, in fact, still a very reliable option, also considering the fact that, although it is a quantum approach, it is still hampered by hardware limitations, noise, or unrefined algorithms. The development of such a technology will, possibly, lead to an increasing role, but it is currently rather a complementary tool rather than a substitute.
Keywords: Remaining Useful Life, Random Forest Regressor, Quantum Circuit Regressor, Predictive modelling
13. CV Classification Using Machine Learning Techniques (p. 123-135)
Roxana-Teodora PERSINARU, The Bucharest University of Economic Studies, Romania
Automatic curriculum vitae (CV) classification represents a complex challenge in the field of human resources, driven by the increasing volume of applications and the need to rapidly identify relevant candidates. Manual CV evaluation is a time-consuming and error-prone process, highlighting the necessity of automated solutions based on Machine Learning and Natural Language Processing techniques. This paper proposes a hybrid architecture for CV classification that combines semantic representations based on multilingual embeddings, adaptive linguistic rules, and semantic similarity mechanisms for candidate-job compatibility assessment. The proposed system enables robust classification of semi-structured CVs and automatic candidate-job matching through an explainable multicriteria evaluation mechanism. Experimental results indicate improved performance compared to a baseline approach relying exclusively on heuristic rules, demonstrating the effectiveness of integrating semantic mechanisms and hybrid classification strategies. The proposed hybrid model achieved an accuracy of 79.2% and significantly outperformed the rule-based baseline approach, which obtained an accuracy of 40.0%.
Keywords: CV classification, NLP, semantic similarity, candidate-job matching, information extraction, embeddings