Dempster's rule of combination serves as a cornerstone in evidence theory for integrating uncertain information from multiple sources. Despite its theoretical appeal, practical implementations remain computationally demanding, especially when relying on recursive algorithms or the Fast Möbius Transform. In this paper, we introduce a novel tensor-based framework for efficient evidence combination under Dempster's rule. By reformulating mass function aggregation as a tensor operation, our approach enables seamless integration with modern machine learning and data processing pipelines that natively support tensor computations. The core mechanism involves applying a projection matrix to the dyadic product of two encoded mass functions, precisely emulating traditional combination behavior. Extensive empirical evaluations demonstrate significant computational advantages, including up to a 1000-fold acceleration in worst-case scenarios compared to classical recursive methods. This work not only provides a computationally efficient and tensor-compatible alternative for real-time information fusion tasks but also lays the foundation for future research at the intersection of evidential reasoning and tensor-based computation. To support reproducibility and facilitate further research, we also provide our implementation of the developed tensor-based methodology as open-source software which is publicly accessible.1 © 2025 The Authors

Financial distress prediction (FDP) is essential for maintaining a stable economy and mitigating systemic financial risks. Statistical and machine learning models have been applied to the FDP problem; however, their performance is limited due to the imbalanced nature of financial data. To address this challenge, we compare the use of resampling techniques and weighted loss functions in enhancing the predictive performance of a gated recurrent unit (GRU) model on a numerical financial dataset. The experiments show that the GRU model with α-balanced Focal Loss consistently outperforms alternative approaches, achieving an AUC of 0.8104. In terms of time-dependency capturing, three years of historical data are found to be the optimal time range for the GRU model. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

Personalized recommendation systems play a vital role in online travel platforms by helping users navigate the vast number of available accommodations. Traditional collaborative filtering methods, especially matrix factorization, are widely used but limited in capturing complex user–hotel interactions due to their linear nature. This study proposes a DNN–based embedding framework for hotel recommendation that learns user and hotel representations directly from sparse rating data and predicts ratings through a regression layer. We evaluate the model on a real-world TripAdvisor dataset containing and compare it with classical (CF_user, CF_item, SVD, ALS) and modern neural baselines (NCF, RNN4Rec). Experimental results demonstrate that even this simplified DNN architecture achieves competitive, and in many cases superior, performance, delivering lower MAE and RMSE and higher R2 scores than all baselines while maintaining manageable computational requirements. These findings highlight the potential of lightweight deep learning models for practical hotel recommendation tasks and pave the way for future extensions incorporating multi-criteria ratings and hybrid features. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.