Search

This research paper conducts a thorough examination, comparing BERT and LSTM architectures with machine learning models for sentiment classification task. To establish a foundational reference point, conventional machine learning algorithms including Logistic Regression (LR), Support Vector Machines (SVM), Random Forest (RF), Decision Tree (DT) and BernoulliNB are utilized. The BERT and LSTM models along with machine learning models have been implemented and trained using a dataset comprising hotel reviews. Their performance has been assessed through the utilization of multiple metrics, including accuracy, precision, recall, and F1-score. In the context of text classification and sentiment analysis tasks, the experimental outcomes highlight the superior effectiveness of BERT and LSTM models when compared to traditional machine learning algorithms. The BERT model distinguishes itself through the incorporation of bidirectional training and contextual embedding strategies, ultimately showcasing outstanding performance in its ability to proficiently capture contextual information and subtle nuances prevalent in textual data. Moreover, the LSTM technique has been able to attain noteworthy outcomes through its effective modeling of sequential data and preservation of temporal dependencies. The present research yields significant findings regarding the comparative analysis of machine learning models employing BERT and LSTM architectures concerning text classification and sentiment analysis. The results of this study provide valuable insights into the merits and constraints of each model, thereby aiding researchers and practitioners in making informed decisions when choosing suitable models that are customized for specific natural language processing (NLP) tasks.

In a recent study, Spencer et al. 2021, we have proposed a computational modeling framework for DNA multi strand dynamics implemented using the agent- and rule-based modeling methodology. While this modeling methodology allows for compact representations for systems with large numbers of different species and complexes, such as the case of self-assembly systems, one of its main drawbacks concerns its scalability. Since each agent is individually represented and modeled in the system, the framework becomes slow when dealing with tens- and hundreds of thousands of individual components. In this study we introduce a method to parallelize the computational modeling process by distributing it over several CPU's. We show that such multi-thread models remain equivalent to their sequential counterpart, while the speedup of the computational process can reach even a one-fold increase.

Constructing a large biological model is a difficult, error-prone process. Small errors in writing a part of the model cascade to the system level and their sources are difficult to trace back. In this paper we extend a recent approach based on Event-B, a state-based formal method with refinement as its central ingredient, allowing us to validate for model consistency step-by-step in an automated way. We demonstrate this approach on a model of the heat shock response in eukaryotes and its scalability on a model of the ErbB signaling pathway. All consistency properties of the model were proved automatically with computer support.

This deliverable has the objective to define a mathematical model for an optimised real-time management of railway traffic under Moving Block (MB). The formulated real-time traffic management model contains: i) a core module for the detection and the sub-optimal resolution of track occupation conflicts under MB and ii) a non-vital module for providing early-warning predictions of potentially hazardous MB traffic situations. The proposed real-time traffic management model includes a mathematical translation of requirements and constraints identified for both MB signalling within WP2 (namely deliverables D2.1 and D2.2) and the GNSS localisation and train integrity devices within WP3 (i.e. deliverables D3.1 – D3.3). An extensive literature review on real-time traffic management models and algorithms shows that so far research efforts have mainly focused on fixed-block and distance-to-go railway operations. Significant gaps still exist in the modelling of MB train operations, despite an increasing number of research works on MB signalling technology is observed since year 2003. A modelling gap analysis is here performed which indicates the need of enhancing existing real-time traffic management algorithms to better align them to the MB concept in terms of infrastructure representation and speed-headway functional dependency. To this end, the RECIFE-MILP real-time traffic management algorithm is enhanced. On one hand a finer infrastructure discretisation is implemented to offer a more suitable track representation under moving block which no longer uses fixed block sections. On the other hand, two different speed levels (namely maximum speed and scheduled speed) are introduced enabling a speed-dependent headway computation in either nominal or delayed traffic scenarios, thereby overcoming the limitation of speed-independent headways, typical of fixed-block traffic rescheduling models. A non-vital early-warning prediction model of hazardous MB traffic conditions is al, Transport and Planning

In the current era, a rapid increase in data volume produces redundant information on the internet. This predicts the appropriate items for users a great challenge in information systems. As a result, recommender systems have emerged in this decade to resolve such problems. Various e-commerce platforms such as Amazon and Netflix prefer using some decent systems to recommend their items to users. In literature, multiple methods such as matrix factorization and collaborative filtering exist and have been implemented for a long time, however recent studies show that some other approaches, especially using artificial neural networks, have promising improvements in this area of research. In this research, we propose a new hybrid recommender system that results in better performance. In the proposed system, the users are divided into two main categories, namely average users, and non-average users. Then, various machine learning and deep learning methods are applied within these categories to achieve better results. Some methods such as decision trees, support vector regression, and random forest are applied to the average users. On the other side, matrix factorization, collaborative filtering, and some deep learning methods are implemented for non-average users. This approach achieves better compared to the traditional methods.