Sentiment Analysis (SA) is a well-known and emerging research field in the area of Natural Language Processing (NLP) and text classification. Feature engineering is considered to be one of the major steps in the Machine Learning (ML) pipeline with effective feature extraction playing a vital role in improving the performance of the SA tasks. Choosing an appropriate feature from the text is considered to be the most challenging task in text classification. This study examines the implementation of different traditional feature extraction models such as Bag of Words (BoW), Term Frequency-Inverse Document Frequency (TF-IDF), N-grams, and word embeddings like Word2Vec and Bidirectional Encoder Representations from Transformers (BERT) on an Amazon review dataset. Furthermore, the hyperparameters of the BERT model were fine-tuned, optimizing its performance on the SA task. Additionally, this research explores the effectiveness of combining the word vectors generated using TF-IDF and BERT for SA. The proposed hybrid model implements an effective negation handling approach and combines TF-IDF with BERT to improve performance for Amazon product review classification.The hybrid model was evaluated using several performance metrics, including accuracy, recall, precision, and F1-score. The proposed hybrid model shows promising results achieving an accuracy of 88%. The integration of BERT and TF-IDF not only enhances the model’s ability to understand and interpret text but also demonstrates the potential of combining advanced and traditional NLP techniques.

Ovarian cancer is one of the most challenging cancers to detect early, often leading to poor survival rates. This study explores supervised and unsupervised machine learning and deep learning approaches to improve predictive performance using clinical and biomarker-based data which was scaled through 2 popular techniques: Min-Max scaling and Z-Score normalization. The research begins by carefully preprocessing the dataset including feature selection to ensure high-quality inputs. Various baseline and ensemble classifiers, including K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), and Logistic Regression (LR), are tested, for better model efficiency on both datasets. To further boost performance, ensemble methods like Stacking, Bagging, and Gradient Boosting, are incorporated. Additionally, unsupervised models like K-Means and DBSCAN clustering are implemented to study further subgroups of the Ovarian Cancer dataset optimizing results. The effects of different feature selection techniques and the impact of standardization versus normalization are compared on both datasets. The Min-Max normalization technique outperformed Z-Score and it is observed that, the Stacking classifier achieved the highest accuracy of 100%, followed by SVM, Logistic Regression, and Bagging, each recording an accuracy of 97%. Further, DBSCAN, a clustering technique outperformed K-Means with a Silhouette Score of 0.7245 and it is observed that clustering performed well with Min-Max when compared with Z-Score normalization technique. The findings highlight that a well-optimized combination of feature selection, ensemble learning, and clustering significantly enhances ovarian cancer prediction , providing a valuable foundation for early diagnosis and clinical decision support.

Data analytics is an integral part of strategic decision making in various fields but not limited to business, education and healthcare systems. Existing research works focus on the discovery of itemsets with rare antecedents and consequent or frequent antecedents and consequent. Analysis of association among itemsets with rare antecedents and frequent consequent is equally important to gain valuable insights before making crucial decisions. Mining these itemsets from large datasets is time and resource intensive process. Expedition in the process of mining aids in quick decision making and hence, the entire dataset needs to be stored in the RAM. In this paper, a novel Weighted Binary Count Tree (WBIN-Tree) is proposed and implemented in CUDA to exploit the power of GPU and discover rules with rare antecedent and frequent consequent using parallel approach. WBIN-Tree stores the entire dataset in an abstract, complete and compact form in the RAM using single database scan. WBIN-Tree is compared with existing sequential and parallel algorithms by varying the data size and dimension. The performance evaluation of WBIN-Tree showed promising results, proving to be the most time and space efficient algorithm to store the entire large dataset in the RAM. However, based on the size of the GPU, the performance drops when executed on datasets with large dimensions which could be handled by processing the attributes in batches. Additionally, a case study is included to understand the importance of mining association rules with rare antecedent and frequent consequent by executing the algorithm on breast cancer dataset.