CBiLSTM: A Hybrid Deep Learning Model for Efficient Reputation Assessment of Cloud Services

Abstract

The cloud market is characterized by fierce rivalry among cloud service providers. The availability of various services with identical functionalities on the market complicates the selection decision for service requesters. Although objective trust measurements can be used to evaluate the trustworthiness of services, they are not always available and are static in nature. Subjective approaches are not always viable because they often require repeated service invocations to collect client feedback. To overcome these limitations, we propose, in this paper, a reputation-based trust assessment approach that combines the Net Brand Reputation (NBR) measure with a deep learning-based sentiment analysis model using online user reviews. CBiLSTM is the name of the proposed deep learning model that hybridizes the Convolutional Neural Networks (CNN) and the Bidirectional Long Short-Term Memory (BiLSTM) layers. The CNN layers deal with text inputs' high dimensionality, while the BiLSTM layer explores the context of the extracted features in both forward and backward directions. CBiLSTM was trained on a new dataset named CLOSER-DREAM, containing more than 13,000 reviews relating to several emerging cloud services to classify these reviews and assess the overall reputation of the cloud services providers. The results of the series of experiments that were conducted have shown that CBiLSTM outperforms the classic deep learning models with 98% of precision, 99% of recall, 98% as an F1-score, and 99.7% of accuracy. Also, CBiLSTM offered a reasonable training time of about 519ms with the CLOSER-DREAM dataset. The classification obtained by applying CBiLSTM was proven to be an effective method to calculate the NBR measure used for the reputation assessment of cloud service providers. The proposed technique yielded an NBR score of 98.3% for Google cloud services, which is close to the real/actual NBR of 96.25%.

User credibility evaluation for reputation measurement of online service

“…User feedback collected by existing online systems mainly includes user reviews and user ratings. Research based on user reviews includes review feature analysis (Maddy K. et al, 2014;Fontanarava J. et al, 2017;Zhang L. et al, 2022;Jafarian B. et al, 2023) and sentiment analysis (Dang et al, 2020;Nandwani P. et al, 2021;Al Saleh R. et al, 2022;Wankhade M. et al, 2022), etc. However, the number of user ratings collected by many online systems is much larger than the number of text reviews, so our work focuses on an online service reputation measurement method based on user ratings.…”

“…Common user feedback mainly includes user comments and user ratings. Recently, research based on user reviews has used the concept of an open world, integrating deep neural networks to obtain more fair and accurate measurements of service reputation (Fontanarava J. et al, 2017;Nandwani P. et al, 2021;Al Saleh R. et al, 2022;Zhang L. et al, 2022;Jafarian B. et al, 2023). However, providing feedback requires a significant investment of time and other costs.…”

Beyond Word-Based Model Embeddings: Contextualized Representations for Enhanced Social Media Spam Detection

“…The prevalence of social media platforms and the large number of content and users they host present new challenges in detecting and stopping spam. In their study, Elakkiya et al (2021) introduced a fresh method for identifying spam by utilizing a blend of CNN and BiLSTM models [29], bolstered by a conjoint attention mechanism [30]. The model showcased impressive accuracy across a range of datasets, including those from Twitter and YouTube.…”

“…These parts work together to control the flow of information, which lets the network store data for any amount of time and makes LSTMs perfect for tasks that require sequences of different lengths and long gaps between important events [37]. Consequently, LSTMs have become a cornerstone in deep learning (DL) for effectively categorizing and processing sequential and time-series data, overcoming the limitations associated with discharging and disappearing gradients that plagued earlier RNN models [29]. Tashtouth et al [38] and Al-Eidi et al [39] discussed implementing cybersecurity systems and highlighted the dual nature of technological advancements in cybersecurity, offering insights into both the challenges and opportunities presented by novel attack vectors and defensive mechanisms.…”